WO2001095253A2 - Quality and safety assurance of medical applications - Google Patents

Quality and safety assurance of medical applications Download PDF

Info

Publication number
WO2001095253A2
WO2001095253A2 PCT/DK2001/000396 DK0100396W WO0195253A2 WO 2001095253 A2 WO2001095253 A2 WO 2001095253A2 DK 0100396 W DK0100396 W DK 0100396W WO 0195253 A2 WO0195253 A2 WO 0195253A2
Authority
WO
WIPO (PCT)
Prior art keywords
data
pattern
determination system
pattern determination
treatment
Prior art date
Application number
PCT/DK2001/000396
Other languages
French (fr)
Other versions
WO2001095253A3 (en
Inventor
Peter Michael Nielsen
Jan Larsen
Original Assignee
Medimage Aps
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Medimage Aps filed Critical Medimage Aps
Priority to AU2001263784A priority Critical patent/AU2001263784A1/en
Publication of WO2001095253A2 publication Critical patent/WO2001095253A2/en
Publication of WO2001095253A3 publication Critical patent/WO2001095253A3/en

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks

Definitions

  • the present invention relates to a method for controlling the performance of a pattern determination system such as a pattern recognition system, and to a method for generalising data determined by the pattern determination system. More particular, the invention relates to a method for controlling the performance of a pattern recognition system used, in a particular embodiment, for determining a pattern of antibody composition in a measured sample, which pattern determination serves the basis of diagnosing different diseases.
  • the present invention relates to a method of determining an area of a human body to be treated by radiation, which determination being based on diagnosing by recognition of a pattern of matter represented in a series of images.
  • a need for this invention within the technical field of flow cytometry is based on the fact, that each flow crew has been working for long periods of time to optimise their measurement and/or analysing procedure, and now, being able to reproduce results, they are reluctant to standardise their procedure regarding experiences of other crews. Other crews might get considerably different results; but still they are able to reproduce their results.
  • one object of the present invention is to provide a method for bringing Flow Cytometry technology to a state of multidimensional data comprehension, allowing complex immunological interactions to be recognised, whereby complex immunological interactions may be diagnosed be the crew.
  • a further object of the present invention is to provide a method for quality and assurance control, and a method for normalizing of data either on a local computer or similar means for computation or downloaded from a remote Flow Cytometry set-ups.
  • the invention relates to a method for controlling the performance of a first pattern determination system adapted to determine a pattern of matter in a measured sample, said first pattern determination system is further adapted to exchange data with a second pattern determination system which also is adapted to determine a pattern of matter in a measured sample and is adapted to exchange data with the first pattern determination system, which method comprises the steps of providing to the first pattern determination system a first set of data representing a distribution of matter in the measured sample, determining a second set of data, by use of the first pattern determination system and based on the first set of data, representing a pattern in the distribution of matter, transferring the first and second sets of data to the second pattern determination system, and - determining, by use of the second pattern determination system and based on at least the first set of data, a fourth set of data representing a pattern in
  • the method for controlling the performance of a first pattern determination system may further comprise the steps of
  • the second pattern determination system based on the predetermined criterion detects differences between the patterns determined by the first and the second pattern determination systems.
  • One such situation could be where a component used in connection with the measurement of the antibody composition has changed characteristics for instance due to seasoning.
  • the fifth set of data could represent changes to be made during future measurements using that component.
  • the method further comprises the step of determining by use of the second pattern determination system a set of normalised data representing a normalisation of the third [sic: "third" ?] set of data, the determination being based on the second set of data.
  • Such a generalization is preferably a normalization of data, such that for instance biasing, different base point etc. occurring between data obtained by different crews may be made to look like they originate from one single crew.
  • the normalization is performed with a specially designed neural network.
  • the normalization performed with the specially designed neural network will refer results of each crew to a standard following procedures of data generalizations and/or pattern recognition performed on corresponding sets of data and the analysis hereof downloaded from the remote flow lab. Said data analysis is preferably performed on a predetermined sample of known content. Thus data normalizing is obtained without interfering with delicate procedures developed locally.
  • the data normalizing application may work completely independent both of the neural network and the neural network software, the specially designed neural network being fed with data directly from the flow data analysis.
  • the neural network is being trained based on normalized data.
  • Fig. 1 Illustrates the neural network operation modes.
  • the network is trained to minimize the prediction error or maximizing performance.
  • the trained network is copied to operate on new inputs x, b in the operation mode predicting the diagnosis.
  • Technical and medical basis for the invention - Flow Cytometry
  • Flow Cytometry is a technique gaining widespread use in molecular biology and medicine. Flow cytometry allows measurements of multiple cellullar parameters and the subsequent purification of sub-populations of cells. Immunological, microbiological and molecular biological research has benefited from the various applications of this technique to explore multiple correlated measurements.
  • the use of fluorescent dyes conjugated to ligands, e.g. antibodies allows measurements of the expression of cell surface and intracellular molecules, receptors can be studied and subpopulations of cells can be identified and purified for further functional studies by cell sorting facilities incorporated in advanced flow cytometers.
  • the latest multi parameter flow cytometers allows the identification and characterisation of patterns of interactions between several molecules. In contrast to earlier flow cytometers for routine use, this can be accomplished with several internal controls in each sample studied, enabling a more standardised identification of rare events by flow cytometry.
  • infectious diseases such as AIDS, autoimmune diseases such as multiple sclerosis and rheumatoid arthritis, and cancer.
  • the measurements are based on light as the source of excitation. Intense illumination is required because the cells are small and pass through the detection point rapidly. Furthermore the light source must be capable of producing specific wavelengths to excite fluorescent dyes. The scattered and fluorescent light generated by cells passing through the illumination beam is collected by photo detectors which convert the photon pulses into electronic signals. The specification of light quality is met by laser technique. The laser produces a coherent, plane, polarised, intense narrow beam of light at specific wavelengths.
  • the light scatter and fluorescent light information are led through a complex array of lenses, filters and mirrors before the photons reach the photo multiplier tubes (PMT's).
  • the signal intensity in each of the PMT's is determined by the optics which determines the wavelengths of light that passes on to the PMT and by the spectral sensitivity of the individual PMT's.
  • the flow cytometer is a delicate instrument of great precision which generates highly reproducible results.
  • the multiple immunofluorescense labelling in combination with morphological parameter measurements allows precise identification of single cell properties.
  • the display of this complex information using frequency histograms does not exploit all available information and data due to our inability to comprehend multidimensional data sets.
  • the lack of possibilities for standardisation in the field of flow cytometry is resulting in a reproducibility gap and the problem is increasing.
  • Neural networks have been used in computer vision for feature extraction, line and edge detection, image restoration and recognition, character recognition, speech processing, object location and motion tracking. Systems have also been trained for machinery imperfection detection by noise analysis and processing. Recently neural network applied to medical decisive sytems have proved successful.
  • neural networks will be used to control model based interpretation procedures.
  • This approach allows us to build interactive procedures into early deliverables and to progressively remove these procedures as more automated and intelligent procedures are developed.
  • the neural networks are used to identify appropriate interpretations based on descriptive models.
  • the use of descriptive models is intended to provide a basis for implementing interpretation procedures that are robust and verifiable.
  • the training sets for the model are easily identified. Calibration samples are repeatedly run on the cytometer, each run with slight differs regarding the set up: Parameters are varied incrementally, and the system may even be trained using conclusive data sets available from other flow cytometry set-ups. Thus a trained multi-layered model is built with the ability to abstract calibration imperfections and specified error corrections from flow cytometry generated data. The application performs normalising of data for standardisation simulation.
  • the training sets for a model capable of automated diagnosis comprises conclusive data sets comprising flow data/diagnosis (the diagnosis meaning either one specific diagnosis or one subtype of a known diagnosis or one defined progression of a specific disease)
  • the Flow Cytometer is a precision instrument capable of counting molecular events in e.g. "white blood cells", leucocytes.
  • the patient sample such as blood, spinal fluid
  • the patient sample is prepared for counting in the following procedure:
  • the leucocytes are separated and stained with conjugate reagents comprising e.g. a specially designed antibody part, which will react with the molecules of interest (CD4, CD8 etc..) on the leucocyte surface.
  • conjugate reagents comprising e.g. a specially designed antibody part, which will react with the molecules of interest (CD4, CD8 etc..) on the leucocyte surface.
  • the antibody is linked to a fluorescent dye.
  • the stained leucocyte suspension is then fed to the instrument.
  • the instrument will create a flow of cells, which is elicited in the counting chamber in the form of a row of droplets each containing exactly one leucocyte. This is achieved through the use of a combination of vacuum and a nozzle connected to an ultrasound generator delicately calibrated; when the liquid passes the nozzle the stream is broken into droplets of a controllable size.
  • each molecule of interest e.g. CD4
  • each molecule of interest e.g. CD4
  • the different molecules are tagged with different dyes to make separate counts. (E.g. green colour counts for CD4, red for CD8, etc..)
  • the neural network is installed on a Flow Cytometer computer system or in similar means for electronically data processing but it could also be installed on a similar computer system dedicated for data analysis or dedicated for other means.
  • the neural network data analysis computer system may be connected to the Flow Cytometer computer system through a network or similar means for data transfer e.g. by means of discs for optical or magnetic or electronic data storage.
  • the Flow Cytometer is installed together with the Flow Cytometer computer system which could be a standard Macintosh computer or a computer running a similar operating system such as win 95, win 97, win-NT, dos or UNIX, connected to the Flow Cytometer through SCSI or similar interface for flow data analysis, whereas the neural network data analysis can preferably be performed on a standard P.C. or similar means for computation connected to the Mac through a network or similar means for data transfer.
  • the procedure regarding the training of the neural network application comprises the steps of collecting large amounts of Flow data, feeding said data to the neural network.
  • the training sets comprises flow data/diagnosis (the diagnosis meaning either one specific diagnosis or one subtype of a known diagnosis or one defined progression of a specific disease). This "clinical" neural network is supposed to work completely independent of the application described below. As an option applications for each category of diseases might be developed from the "mother" clinical neural network.
  • the invention relating to a method for quality and assurance control by means of a first computer system
  • the invention comprises a specially designed neural network software being able to detect imperfections or aberrations in Flow Cytometry procedures or in antibody reagents procedures through data generalisation and/or pattern recognition.
  • the software may be installed on a second server system remote from the Flow Cytometer and adjacent computer system or systems performing the flow data analysis.
  • the second server system is then connected to the first computer system by transmission means.
  • the second server system may preferably comprise means for downloading corresponding sets of data and data analysis results from one or more third computer systems installed in remote Flow Cytometry laboratories, for the purpose of comparing the data analysis results downloaded from the third computer system with results obtained from performing data analysis with the second server system software on the corresponding data downloaded from the third computer system.
  • the successions of downloaded data sets may be utilised for further training of a second specially designed neural network software.
  • Said further training may enhance the performance of the second neural network in comparison to the first neural network, and when the enhancement of performance have reached a predetermined significance, the first neural network may be replaced by the second neural network on the second server system.
  • the training may be transmitted to the third computer systems replacing existing neural network software versions with the second neural network. From this point a new training session may be initiated leading to the development of a third neural network, and so on.
  • Neural networks have proven to be very efficient generic models for analysis of multidimensional, non-linear complex data such as Flow Cytometry data.
  • the neural network analysis is applied to samples of multi-dimensional fluorescence expressions and/or analysis of the signature constituted by distribution of various molecules on the surface of the individual cells.
  • the applied neural network methods can be divided into supervised and unsupervised.
  • the overall aim is to assist disease diagnosis and progress as well as providing an interactive exploration framework.
  • the supervised methods originate from paired training samples of: Multi-dimensional flow Cytometry data x.
  • Vector of behavioral data such as physiological measurements and demographic data (sex, age).
  • the neural network is trained to predict the diagnosis/state of disease embodied by y from Flow Cytometry and behavioral data, x and b.
  • the neural network forms a non-linear probabilistic model of y given x and b.
  • the prediction of y is generally non-linear and little information of the complex relationship is available a priori. This urges the need for data-driven neural network methods rather than simple parametric linear methods.
  • the design of the neural network comprises the following steps:
  • Training of the network by adjusting characteristic network weights so as to minimize the prediction error on the available training data.
  • Training can be improved by using an active learning scheme, that is, the training data is gradually actively selected to achieve maximum performance improvement per training sample. Active learning thus reduces the cost involved in expert training data labeling.
  • Unsupervised methods aim at finding structure in the multi-dimensional Flow Cytometry measurements, x and behavioral data b from the training data set without having access to the diagnosis variable y.
  • Unsupervised learning is used as a preprocessing step in the supervised setup or as an independent explorative tool.
  • structure is found by clustering methods such as the k-means algorithm; however, the neural network literature offers many improved techniques for unsupervised learning, such as: independent components analysis, generative topographic mapping, and advanced density estimation optimized by using the generalization concept.
  • combined unsupervised- supervised schemes may improve the overall performance.
  • Intra laboratory control is done by using a novelty detection scheme.
  • the novelty detection is based on monitoring the behavior of the inputs x, b using a model of the joint probability density p(x; b). This model is trained using unsupervised techniques, as described above. If unlikely inputs occur a warning of the predicted diagnosis being uncertain is posted.
  • the detected novelty requires action from an expert to interpret the data or to use the inter laboratory control to search for similar events.
  • the novelty could either be interesting unknown Flow Cytometry data expressions or simply measurement/procedure errors.
  • the inter laboratory quality control is procured by comparing results from individual laboratories with results from a mother neural network running on a remote server system.
  • the comparison is based on predictive performance (generalization error), reproducibility and other quality measures, as mentioned in the supervised method section above.
  • performance is significantly different, the results from the laboratory are signified as dubious.
  • the events are stored, as similar discrepancies may occur at a later stage, thus increasingly providing evidence of possible unknown Flow Cytometry data expressions.
  • Comparable data from different laboratories are fused in a mother network using an ensemble or consensus neural network method, providing general opinion predictions.
  • the neural networks of the individual Flow Cytometry laboratories are gradually improved via retraining by using the massive mother neural network database.
  • the present invention relates to a method for automated detection of abnormalities (e.g. tumour tissue) in an image or sequence of images obtained from a person (or any living creature), and subsequently, automated suggestion of the space volume to be exposed to radiation treatment, by use of a knowledge enhancement (e.g. neural network based) software application.
  • a knowledge enhancement e.g. neural network based
  • the 3D RadioPlan module is applicable to an existing radio neurosurgery device called the Leksell GammaKnife® as described in US patent no. 6,049,587 which hereby is incorporated by reference.
  • the 3D RadioPlan is replacing the existing dose planning procedures, with regard to import of patient image sequences and image manipulation, but not the functions of the radiation treatment control module.
  • a number of (and at least one) patient images are provided. These patient images are obtained for instance by CT and/or MRI (fMRI) and/or PET scanners, where each image is an image of a cross section of a part of a human body. Such sequence is for short termed a "patient image sequence”. If these images are not in pixel format, each image is digitised/pixelised in such a manner that the images are constituted by pixels each having assigned thereto a pixel value representing a physical quantity of the tissue the pixel image.
  • the 3D RadioPlan module is capable of importing images containing information about tissue (cell-) functions such as PET, fMRI, interpret the functionality information and add to each pixel value representing a physical quantity another value representing a physical quality through the characterising of an identified tissue entity as either abnormal (e.g. malignant) or functionally critically important (e.g. speech or motor area of the brain)
  • tissue cell-
  • fMRI fMRI-induced pixel value representing a physical quantity another value representing a physical quality
  • the image sequences are manipulated in a 3D segmenting procedure enabling the multiple slices to be rendered visible in a 3D display allowing the doctor (who retains responsibility for the treatment) to view the anatomical/functional features "in toto".
  • the multiple slices are rendered visible by assigning a transparent colour to each pixel value (quantitative and qualitative) and visually stacking the image sequence onto each other in the order they where obtained and with a distance between each image proportional to the actual physical distance between the different images.
  • the procedure leading to the training of the neural network comprises the steps of -building a knowledge of the architectural variety of normal body parts (in this embodiment human brains) -building an experience in recognising an architectural variety of abnormal human brains enabling the diagnosis of e.g. different brain tumours, AVM (arterio-venous malformations) suitable for radiation treatment
  • the neural network will add a value (malignant!) to each pixel of this specific space volume. If, by importing a PET activation image sequence, part of the same tissue entity is recognised being part of the (e.g. Broca) speech center, another value (critically important area!) is added to each pixel of this specific space volume. In the space volume of pixels having both values mentioned above added, in case of malignancy, the first added value is regarded as the more important and all of the space volume with that value added will be included in the space volume suggested for treatment.
  • tissue entity in the above mentioned example is recognised as being part of an AVM the value added is regarded the less important, and the space volume with the added value of "speech center! is not included in the space volume suggested for treatment. It is understood that information about abnormalities will lead to a decision of inclusion of detected abnormal space volume suggested for treatment, whereas information about critically important brain areas will lead to a decision of exclusion.
  • the patient image material database used for the training procedures is built from de novo scans, already existing in-house database resources or data downloaded from other diagnostic imaging centres.
  • the neural network based knowledge enhancement module (first pattern determination system) will, referring to the large database of normal and pathological patient image sequences perform a pattern recognition procedure which comprises detection and classification of any abnormality. Based on the pattern determined (the detection and classification) and the values added the enhancement module will
  • the optimised dose plan determined comprises a space volume to be included in the radio treatment focus.
  • This determination (suggestion) procedure is also referring to information from the above mentioned database (if for instance the type of brain tumour called astrocytoma is detected, it is necessary to suggest a rather large space volume to be treated, as is it a common feature of this type of tumour to grow in a "star- shaped" manner, and treatment of the centre (easily identified in the image) as well as the "beams" (often not present in the image) is crucial to the success of the treatment initiative.)
  • the determination (suggestion) procedure will also refer to knowledge regarding important areas of the brain. E.g. speech and right hand motor function may be located in the brain by PET activation imaging techniques, or, if PET is not available, predicted from knowledge stored in a database.
  • the optimised dose plan is calculated in such a way that malignant tissue is treated sufficiently, but important areas of the brain are spared to diminish side effects of the treatment.
  • the method for controlling the performance of a pattern determination system described in connection with the Flow Cytometry example is also applied to the radio treatment method. Accordingly, the patient image sequence, which has been inputted to the first pattern determination system (the neural network based knowledge enhancement module) is transmitted to a second pattern determination system, which also (by utilising a neural network) detects and classifies any abnormality. Subsequently, the detection and classification determined by the two systems are examined against each other to judge whether the same pattern has been determined.
  • the dose plan is transmitted to the second pattern determination system which validates said dose plan by comparing said dose plan with a dose plan determined by the second pattern determination system so as to judge whether those dose plans are different.
  • the 3D RadioPlan module is used in connection with a radiation treatment device without means for the complete fixing of coordinates of the organ to be treated (e.g. a linear accelerator).
  • a radiation treatment device without means for the complete fixing of coordinates of the organ to be treated (e.g. a linear accelerator).
  • the part of the human body to be treated changes shape and size during the treatment. This change of shape and size normally requires the space volume to be extended far beyond the volume of the actual diseased tissue in order to be able to guarantee treatment of the diseased tissue causing healthy tissue to be radiated.
  • a neural network is trained to predict the, based on fix-points assigned to the patient instantaneous, shape, size and position of the space volume suggested.
  • a number of fix-points are assigned to the patient's body surface in the form of optically readable marks arranged in such a manner that a 3-dimensional coordinate of each fix-point is obtainable.
  • a reference co-ordinate system is defined to which all movements of the fix-points are referenced. Then the patient image sequence is obtained and each time an image is obtained the position of the fix-points are recorded and the image is transformed into the reference coordinate system, so that the position and size of each pixel in the image is described with reference to the reference co-ordinate system. Each transformation is based on the instantaneous position of the fix-points relative to the reference co-ordinate system.
  • the hereby obtained patient image sequence may suitable be referred to as a normalised patient image sequence.
  • the normalised patient image sequence is ready for further processing and is input to the dose planning module which in turn suggests a normalised space volume to be included in the radio treatment focus (the word normalised is here used to emphasise that the space volume is suggested based on the normalised patient image sequnce).
  • the hereby suggested space volume will be referenced to the reference co-ordinate system.
  • the instantaneous position of the fix-points are recorded and the suggested normalised space volume is transformed according to the fix-point's position in the reference co-ordinate system (the transformation may be viewed upon as the reverse transformation of the transformation providing the normalised patient image sequence).
  • the neural network should be so constructed and trained so as it is able to normalise patient image sequences and able to perform the reverse transformation.
  • Output may be used for controlling the movement of a patient (e.g. executing the above mentioned reverse transformation by moving the patient's bed exactly to keep the coordinates of the space volume to be treated within the focus of the radiation beams), or, if possible, a movement of the focus of the radiation beam during treatment.
  • the method is utilized for the automated diagnosis and/or the data normalizing and/or the quality and assurance control in the field of medical imaging.
  • AD Alzheimer's Disease
  • AD is most often diagnosed either by autopsy or after several years of progression. AD cannot be cured, but new types of medication can alter the progression of AD dramatically mitigateating symptoms for several years, thus improving the quality of life for the patient and ensuring society health care savings.
  • the database may import data for the training sets from other brain imaging modalities (CT/MRI) enabling the knowledge enhancement application to build experience from "tissue functional" imaging data (PET/SPECT/fMRI) as well as “tissue architectural” imaging data (CT/MRI).
  • CT/MRI brain imaging modalities
  • PET/SPECT/fMRI tissue functional imaging data
  • CT/MRI tissue architectural imaging data
  • the patient image material database used for the training procedures is built from de novo scans, already existing in-house database resources or data downloaded from other diagnostic imaging centres.
  • the method for controlling the performance of a pattern determination system described in connection with the Flow Cytometry example is also applied to the medical imaging automated diagnosis method. Accordingly, the patient image sequence, which has been inputted to the first pattern determination system (the neural network based knowledge enhancement module) is transmitted to a second pattern determination system, which also (by utilising a neural network) detects and classifies any abnormality. Subsequently, the detection and classification determined by the two systems are examined against each other to judge whether the same pattern has been determined.
  • diagnosis resulting from analysing the medical image is transmitted to the second pattern determination system which validates said diagnosis by comparing said diagnosis with a diagnosis determined by the second pattern determination system so as to judge whether those diagnoses are different.

Abstract

The present invention relates in a broad aspect to assurance of quality and safety of medical applications determining - or suggesting - diagnosis and/or treatment. In general, such applications are according to the present invention applications which are designed to operate without assistance of a medical adviser in daily use and accordingly these applications must be supervised in some sense in order to be able to guarantee that diagnoses and/or treatments determined by these applications are not determined in a manner that patients are exposed to life threatening treatment due to malfunctioning of these applications. In the general perspective according to the present invention, treatment is suggested based on a diagnosis and a diagnosis is viewed upon as recognition of a pattern. For instance, a diagnosis is the result of recognition of a pattern of antibody substances in a sample taken from the tissue of a patient. Furthermore, the term diagnosis has been given a broad interpretation in the present content so as to include - and/or even be- determination of treatment. The philosophy behind such an approach is that the determination of a treatment is according to the present invention based on determination of a pattern. For instance in case of the prescription of a drug preventing attacks of Multiple Sclerosis, a measured blood sample may be analyzed for patterns of matter in the sample and once a pattern is recognized the treatment assigned to said pattern is the treatment suggested.

Description

TITLE:
QUALITY AND SAFETY ASSURANCE OF MEDICAL APPLICATIONS
TECHNICAL FIELDS RELATED TO THE PRESENT INVENTION
The present invention relates to a method for controlling the performance of a pattern determination system such as a pattern recognition system, and to a method for generalising data determined by the pattern determination system. More particular, the invention relates to a method for controlling the performance of a pattern recognition system used, in a particular embodiment, for determining a pattern of antibody composition in a measured sample, which pattern determination serves the basis of diagnosing different diseases.
In another particular embodiment the present invention relates to a method of determining an area of a human body to be treated by radiation, which determination being based on diagnosing by recognition of a pattern of matter represented in a series of images. BACKGROUND OF THE INVENTION AND INTRODUCTION TO THE INVENTION
In the following the background of the invention is addressed by way of examples. It should be obvious to those skilled in the art that these examples are not considered to be an exhaustive list, but merely expresses some particular important technical fields. It should therefore also be obvious to those skilled in the art, that details described with respect to a specific field of technology is applicable also in other fields of technology.
Flow cytometry:
A need for this invention within the technical field of flow cytometry is based on the fact, that each flow crew has been working for long periods of time to optimise their measurement and/or analysing procedure, and now, being able to reproduce results, they are reluctant to standardise their procedure regarding experiences of other crews. Other crews might get considerably different results; but still they are able to reproduce their results.
Furthermore, the analysing of the measurement is often based on the practical skills and experience of the crew and until now no efficient and safe method for analysing data relating to antibody composition in a measured sample has been present.
Thus, one object of the present invention is to provide a method for bringing Flow Cytometry technology to a state of multidimensional data comprehension, allowing complex immunological interactions to be recognised, whereby complex immunological interactions may be diagnosed be the crew.
A further object of the present invention is to provide a method for quality and assurance control, and a method for normalizing of data either on a local computer or similar means for computation or downloaded from a remote Flow Cytometry set-ups.
This is achieved by the present invention by utilising a specially designed neural network software to perform data generalisation and pattern recognition analysing in multidimensional data, preferably originating from a Flow Cytometer. In a first aspect, the invention relates to a method for controlling the performance of a first pattern determination system adapted to determine a pattern of matter in a measured sample, said first pattern determination system is further adapted to exchange data with a second pattern determination system which also is adapted to determine a pattern of matter in a measured sample and is adapted to exchange data with the first pattern determination system, which method comprises the steps of providing to the first pattern determination system a first set of data representing a distribution of matter in the measured sample, determining a second set of data, by use of the first pattern determination system and based on the first set of data, representing a pattern in the distribution of matter, transferring the first and second sets of data to the second pattern determination system, and - determining, by use of the second pattern determination system and based on at least the first set of data, a fourth set of data representing a pattern in the distribution of matter, and comparing the second set of data and the fourth set of data and determine whether the two sets of data are representing the same pattern of matter in order to judge whether the pattern of matter determined by the first pattern determination system is different from the pattern of matter determined by the second pattern system or vice versa.
In the first aspect of the present invention the method for controlling the performance of a first pattern determination system may further comprise the steps of
determining a fifth set of data indicating changes to be applied to the first pattern determination system and/or indicating changes to be applied to a method used for providing a distribution of antibody substances, and transferring the fifth set of data to the first pattern determination system.
This possibility is especially useful when the second pattern determination system based on the predetermined criterion detects differences between the patterns determined by the first and the second pattern determination systems. One such situation could be where a component used in connection with the measurement of the antibody composition has changed characteristics for instance due to seasoning. In this case the fifth set of data could represent changes to be made during future measurements using that component.
In very important embodiments, the method further comprises the step of determining by use of the second pattern determination system a set of normalised data representing a normalisation of the third [sic: "third" ?] set of data, the determination being based on the second set of data.
Such a generalization is preferably a normalization of data, such that for instance biasing, different base point etc. occurring between data obtained by different crews may be made to look like they originate from one single crew. In a particular preferred embodiment the normalization is performed with a specially designed neural network.
The normalization performed with the specially designed neural network will refer results of each crew to a standard following procedures of data generalizations and/or pattern recognition performed on corresponding sets of data and the analysis hereof downloaded from the remote flow lab. Said data analysis is preferably performed on a predetermined sample of known content. Thus data normalizing is obtained without interfering with delicate procedures developed locally.
The data normalizing application may work completely independent both of the neural network and the neural network software, the specially designed neural network being fed with data directly from the flow data analysis. Preferably the neural network is being trained based on normalized data.
DETAILED DESCRIPTION OF THE INVENTION
Preferred embodiments of the invention will now be described in details with reference to the drawing in which
Fig. 1 : Illustrates the neural network operation modes. The network is trained to minimize the prediction error or maximizing performance. The trained network is copied to operate on new inputs x, b in the operation mode predicting the diagnosis. Technical and medical basis for the invention - Flow Cytometry
Flow Cytometry is a technique gaining widespread use in molecular biology and medicine. Flow cytometry allows measurements of multiple cellullar parameters and the subsequent purification of sub-populations of cells. Immunological, microbiological and molecular biological research has benefited from the various applications of this technique to explore multiple correlated measurements. The use of fluorescent dyes conjugated to ligands, e.g. antibodies allows measurements of the expression of cell surface and intracellular molecules, receptors can be studied and subpopulations of cells can be identified and purified for further functional studies by cell sorting facilities incorporated in advanced flow cytometers.
Measurements of two lights scatter and two or three immunofluorescent parameters on each cell is now in routine use in flow cytometry laboratories. Several new flow cytometers do, however, allow additional parameters to be measured, enabling the simultaneous aquisition of six or seven parameters on each cell. The problem of analysing and displaying six- or seven-dimensional data introduces major obstacles to further progress in the field.
The latest multi parameter flow cytometers allows the identification and characterisation of patterns of interactions between several molecules. In contrast to earlier flow cytometers for routine use, this can be accomplished with several internal controls in each sample studied, enabling a more standardised identification of rare events by flow cytometry. Examples of complexly balanced, unstable systems where there is a need for better methods for the study of the molecular pathology, are infectious diseases such as AIDS, autoimmune diseases such as multiple sclerosis and rheumatoid arthritis, and cancer.
Measurements
The measurements are based on light as the source of excitation. Intense illumination is required because the cells are small and pass through the detection point rapidly. Furthermore the light source must be capable of producing specific wavelengths to excite fluorescent dyes. The scattered and fluorescent light generated by cells passing through the illumination beam is collected by photo detectors which convert the photon pulses into electronic signals. The specification of light quality is met by laser technique. The laser produces a coherent, plane, polarised, intense narrow beam of light at specific wavelengths.
The light scatter and fluorescent light information are led through a complex array of lenses, filters and mirrors before the photons reach the photo multiplier tubes (PMT's). The signal intensity in each of the PMT's is determined by the optics which determines the wavelengths of light that passes on to the PMT and by the spectral sensitivity of the individual PMT's.
Characterisation of technological limitations for further progress
The flow cytometer is a delicate instrument of great precision which generates highly reproducible results. The multiple immunofluorescense labelling in combination with morphological parameter measurements allows precise identification of single cell properties. The display of this complex information using frequency histograms does not exploit all available information and data due to our inability to comprehend multidimensional data sets. The lack of possibilities for standardisation in the field of flow cytometry is resulting in a reproducibility gap and the problem is increasing.
Neural Network Technology
Neural networks have been used in computer vision for feature extraction, line and edge detection, image restoration and recognition, character recognition, speech processing, object location and motion tracking. Systems have also been trained for machinery imperfection detection by noise analysis and processing. Recently neural network applied to medical decisive sytems have proved successful.
Structuring of a particular solution
In the application considered here we propose an approach in which neural networks will be used to control model based interpretation procedures. This approach allows us to build interactive procedures into early deliverables and to progressively remove these procedures as more automated and intelligent procedures are developed. The neural networks are used to identify appropriate interpretations based on descriptive models. The use of descriptive models is intended to provide a basis for implementing interpretation procedures that are robust and verifiable.
It is essential to provide forward pathways for updating evidence and backward pathways to control the interpretation process.
The training sets for the model are easily identified. Calibration samples are repeatedly run on the cytometer, each run with slight differs regarding the set up: Parameters are varied incrementally, and the system may even be trained using conclusive data sets available from other flow cytometry set-ups. Thus a trained multi-layered model is built with the ability to abstract calibration imperfections and specified error corrections from flow cytometry generated data. The application performs normalising of data for standardisation simulation.
The training sets for a model capable of automated diagnosis comprises conclusive data sets comprising flow data/diagnosis (the diagnosis meaning either one specific diagnosis or one subtype of a known diagnosis or one defined progression of a specific disease)
Data analysis set-up
When investigating a particular sample in the flow cytometer the vast amount of data generated is stored in the computer memory. Instead of displaying the results in two- or three-dimensional graphical expressions, data is fed to the neural network module where multi-parameter analysis results without loss of detail are described through pattern recognition and subtle interactions are identified and diagnosed. Simultaneously the Quality and Assurance Control method is applied (on the fly) for documentation of procedure and data analysis.
Summary of daily clinical practice
The Flow Cytometer is a precision instrument capable of counting molecular events in e.g. "white blood cells", leucocytes. The patient sample ( such as blood, spinal fluid) is prepared for counting in the following procedure:
The leucocytes are separated and stained with conjugate reagents comprising e.g. a specially designed antibody part, which will react with the molecules of interest (CD4, CD8 etc..) on the leucocyte surface. The antibody is linked to a fluorescent dye. The stained leucocyte suspension is then fed to the instrument.
The instrument will create a flow of cells, which is elicited in the counting chamber in the form of a row of droplets each containing exactly one leucocyte. This is achieved through the use of a combination of vacuum and a nozzle connected to an ultrasound generator delicately calibrated; when the liquid passes the nozzle the stream is broken into droplets of a controllable size.
In the counting chamber the cells are exposed to laser lights of different wavelenghts exiting the various dye molecules attached to the cells. These dyes will respond with various fluorescent emissions of different colours. The pattern of light is recorded with PMT's (photo multiplier tubes) and the counts are collected in a computer application. Thus, in theory, each molecule of interest (e.g. CD4) on the surface of a single cell will produce a count to be displayed on the computer screen. Of course, the different molecules are tagged with different dyes to make separate counts. (E.g. green colour counts for CD4, red for CD8, etc..)
A PREFERRED EMBODIMENT OF THE INVENTION - CYTOMETRY
In a preferred embodiment of the present invention the neural network is installed on a Flow Cytometer computer system or in similar means for electronically data processing but it could also be installed on a similar computer system dedicated for data analysis or dedicated for other means. The neural network data analysis computer system may be connected to the Flow Cytometer computer system through a network or similar means for data transfer e.g. by means of discs for optical or magnetic or electronic data storage. In a preferred embodiment the Flow Cytometer is installed together with the Flow Cytometer computer system which could be a standard Macintosh computer or a computer running a similar operating system such as win 95, win 97, win-NT, dos or UNIX, connected to the Flow Cytometer through SCSI or similar interface for flow data analysis, whereas the neural network data analysis can preferably be performed on a standard P.C. or similar means for computation connected to the Mac through a network or similar means for data transfer. The procedure regarding the training of the neural network application comprises the steps of collecting large amounts of Flow data, feeding said data to the neural network. The training sets comprises flow data/diagnosis (the diagnosis meaning either one specific diagnosis or one subtype of a known diagnosis or one defined progression of a specific disease). This "clinical" neural network is supposed to work completely independent of the application described below. As an option applications for each category of diseases might be developed from the "mother" clinical neural network.
According to the invention relating to a method for quality and assurance control by means of a first computer system, the invention comprises a specially designed neural network software being able to detect imperfections or aberrations in Flow Cytometry procedures or in antibody reagents procedures through data generalisation and/or pattern recognition. The software may be installed on a second server system remote from the Flow Cytometer and adjacent computer system or systems performing the flow data analysis. The second server system is then connected to the first computer system by transmission means. Furthermore the second server system may preferably comprise means for downloading corresponding sets of data and data analysis results from one or more third computer systems installed in remote Flow Cytometry laboratories, for the purpose of comparing the data analysis results downloaded from the third computer system with results obtained from performing data analysis with the second server system software on the corresponding data downloaded from the third computer system.
The successions of downloaded data sets may be utilised for further training of a second specially designed neural network software. Said further training may enhance the performance of the second neural network in comparison to the first neural network, and when the enhancement of performance have reached a predetermined significance, the first neural network may be replaced by the second neural network on the second server system. The training may be transmitted to the third computer systems replacing existing neural network software versions with the second neural network. From this point a new training session may be initiated leading to the development of a third neural network, and so on.
NEURAL NETWORK
Neural networks have proven to be very efficient generic models for analysis of multidimensional, non-linear complex data such as Flow Cytometry data.The neural network analysis is applied to samples of multi-dimensional fluorescence expressions and/or analysis of the signature constituted by distribution of various molecules on the surface of the individual cells.
The applied neural network methods can be divided into supervised and unsupervised. The overall aim is to assist disease diagnosis and progress as well as providing an interactive exploration framework.
The supervised methods originate from paired training samples of: Multi-dimensional flow Cytometry data x.
Known diagnosis, progression of specific diseases and other characteristics of the specific disease assembled in the vector y.
Vector of behavioral data, b, such as physiological measurements and demographic data (sex, age).
From a training data set of paired samples the neural network is trained to predict the diagnosis/state of disease embodied by y from Flow Cytometry and behavioral data, x and b. The neural network forms a non-linear probabilistic model of y given x and b. The prediction of y is generally non-linear and little information of the complex relationship is available a priori. This urges the need for data-driven neural network methods rather than simple parametric linear methods.
The design of the neural network comprises the following steps:
Selection of the neural network family.
Preprocessing of data using advanced unsupervised neural network techniques as described below or classical methods like principal component analysis.
Training of the network by adjusting characteristic network weights so as to minimize the prediction error on the available training data. Training can be improved by using an active learning scheme, that is, the training data is gradually actively selected to achieve maximum performance improvement per training sample. Active learning thus reduces the cost involved in expert training data labeling.
Optimization of the network structure in order to yield optimal predictive performance.
Evaluating quality of the network by computing performance on test data not used for training. This provides an unbiased, out-of-sample predictive performance assessment (generalization performance), e.g. defined as probability of correctly classifying the disease. More elaborate performance measures such as confusion tables, receiver operation characteristics and confidence intervals are also available.
Visualization/Interpretation of the optimized neural network model via an importance analysis of employed Flow Cytometry measurements and behavioral data. Furthermore reproducibility within and between subjects/patients is analyzed using statistical re-sampling.
When the network is designed it is going to run in an operation mode predicting the diagnosis y from measurements of x and b. The training and operation phases are shown in Fig. 1.
Unsupervised methods aim at finding structure in the multi-dimensional Flow Cytometry measurements, x and behavioral data b from the training data set without having access to the diagnosis variable y. Unsupervised learning is used as a preprocessing step in the supervised setup or as an independent explorative tool. Traditionally structure is found by clustering methods such as the k-means algorithm; however, the neural network literature offers many improved techniques for unsupervised learning, such as: independent components analysis, generative topographic mapping, and advanced density estimation optimized by using the generalization concept. In addition, combined unsupervised- supervised schemes may improve the overall performance.
Quality and Assurance Control of Flow Cytometry laboratories is performed at two levels: an intra laboratory quality control and an inter laboratory control. Intra laboratory control is done by using a novelty detection scheme. When the network is used in operation mode only the predicted diagnosis is available. Thus the novelty detection is based on monitoring the behavior of the inputs x, b using a model of the joint probability density p(x; b). This model is trained using unsupervised techniques, as described above. If unlikely inputs occur a warning of the predicted diagnosis being uncertain is posted. The detected novelty requires action from an expert to interpret the data or to use the inter laboratory control to search for similar events. The novelty could either be interesting unknown Flow Cytometry data expressions or simply measurement/procedure errors.
The inter laboratory quality control is procured by comparing results from individual laboratories with results from a mother neural network running on a remote server system. The comparison is based on predictive performance (generalization error), reproducibility and other quality measures, as mentioned in the supervised method section above. When performance is significantly different, the results from the laboratory are signified as dubious. However still the events are stored, as similar discrepancies may occur at a later stage, thus increasingly providing evidence of possible unknown Flow Cytometry data expressions. Comparable data from different laboratories are fused in a mother network using an ensemble or consensus neural network method, providing general opinion predictions. The neural networks of the individual Flow Cytometry laboratories are gradually improved via retraining by using the massive mother neural network database.
A necessary requirement for fusion of Flow Cytometry data from diverse laboratories with individually tuned operation procedures is normalization of data. This is done by comparing performance on Flow Cytometry samples of known content for the various tasks/diseases handled by the mother network. The normalization data are used to train a neural network with the purpose of modeling systematic, reproducible differences among laboratories. Thus in essence, this lead to the construction of non-linear calibration curves. Technical and medical basis for the invention - radiation treatment
In this example, the present invention relates to a method for automated detection of abnormalities (e.g. tumour tissue) in an image or sequence of images obtained from a person (or any living creature), and subsequently, automated suggestion of the space volume to be exposed to radiation treatment, by use of a knowledge enhancement (e.g. neural network based) software application.
A PREFERRED EMBODIMENT OF THE INVENTION - RADIATION TREATMENT
In a first preferred embodiment of the present invention, the 3D RadioPlan module, is applicable to an existing radio neurosurgery device called the Leksell GammaKnife® as described in US patent no. 6,049,587 which hereby is incorporated by reference. The 3D RadioPlan is replacing the existing dose planning procedures, with regard to import of patient image sequences and image manipulation, but not the functions of the radiation treatment control module.
According to the invention a number of (and at least one) patient images are provided. These patient images are obtained for instance by CT and/or MRI (fMRI) and/or PET scanners, where each image is an image of a cross section of a part of a human body. Such sequence is for short termed a "patient image sequence". If these images are not in pixel format, each image is digitised/pixelised in such a manner that the images are constituted by pixels each having assigned thereto a pixel value representing a physical quantity of the tissue the pixel image.
These patient image sequences are imported to the 3D RadioPlan module in which a further processing is performed.
In a preferred embodiment the 3D RadioPlan module is capable of importing images containing information about tissue (cell-) functions such as PET, fMRI, interpret the functionality information and add to each pixel value representing a physical quantity another value representing a physical quality through the characterising of an identified tissue entity as either abnormal (e.g. malignant) or functionally critically important (e.g. speech or motor area of the brain) The image sequences are manipulated in a 3D segmenting procedure enabling the multiple slices to be rendered visible in a 3D display allowing the doctor (who retains responsibility for the treatment) to view the anatomical/functional features "in toto".
In the preferred embodiment the multiple slices are rendered visible by assigning a transparent colour to each pixel value (quantitative and qualitative) and visually stacking the image sequence onto each other in the order they where obtained and with a distance between each image proportional to the actual physical distance between the different images.
The procedure leading to the training of the neural network comprises the steps of -building a knowledge of the architectural variety of normal body parts (in this embodiment human brains) -building an experience in recognising an architectural variety of abnormal human brains enabling the diagnosis of e.g. different brain tumours, AVM (arterio-venous malformations) suitable for radiation treatment
-building a knowledge of the functional variety of normal brains (emphazising space volume distribution of specialised brain cortex areas such as speech and motor centers) -building an experience in recognising a functional variety of abnormal human brains (e.g. a distribution of malignancy)
-defining rules for hierachisation of information upon which the space volume to be treated is suggested.
(E.g if a tissue entity of the brain is recognised as being part of a malignant tumour the neural network will add a value (malignant!) to each pixel of this specific space volume. If, by importing a PET activation image sequence, part of the same tissue entity is recognised being part of the (e.g. Broca) speech center, another value (critically important area!) is added to each pixel of this specific space volume. In the space volume of pixels having both values mentioned above added, in case of malignancy, the first added value is regarded as the more important and all of the space volume with that value added will be included in the space volume suggested for treatment.
If, on the contrary, the tissue entity in the above mentioned example is recognised as being part of an AVM the value added is regarded the less important, and the space volume with the added value of "speech center!" is not included in the space volume suggested for treatment. It is understood that information about abnormalities will lead to a decision of inclusion of detected abnormal space volume suggested for treatment, whereas information about critically important brain areas will lead to a decision of exclusion.
The patient image material database used for the training procedures is built from de novo scans, already existing in-house database resources or data downloaded from other diagnostic imaging centres.
The neural network based knowledge enhancement module (first pattern determination system) will, referring to the large database of normal and pathological patient image sequences perform a pattern recognition procedure which comprises detection and classification of any abnormality. Based on the pattern determined (the detection and classification) and the values added the enhancement module will
determine an optimised dose plan - in practical use the determined optimised dose plan is regarded as a suggested dose plan as the doctor in charge of the treatment is asked to confirm the dose plan. The optimised dose plan determined (suggested) comprises a space volume to be included in the radio treatment focus. This determination (suggestion) procedure is also referring to information from the above mentioned database (if for instance the type of brain tumour called astrocytoma is detected, it is necessary to suggest a rather large space volume to be treated, as is it a common feature of this type of tumour to grow in a "star- shaped" manner, and treatment of the centre (easily identified in the image) as well as the "beams" (often not present in the image) is crucial to the success of the treatment initiative.) The determination (suggestion) procedure will also refer to knowledge regarding important areas of the brain. E.g. speech and right hand motor function may be located in the brain by PET activation imaging techniques, or, if PET is not available, predicted from knowledge stored in a database. The optimised dose plan is calculated in such a way that malignant tissue is treated sufficiently, but important areas of the brain are spared to diminish side effects of the treatment.
ask the doctor in charge to confirm the suggested dose plan or correct the dose plan at his/her convenience, using the manual image manipulation features (either modelling the space volume by "mouse dragging" or any other interface) export the dose plan, if accepted by the doctor in charge, to the radio treatment control module
Also in this example of the present invention it may be crucial to validate the method outlined above (the determination of the pattern). In this case the method for controlling the performance of a pattern determination system described in connection with the Flow Cytometry example is also applied to the radio treatment method. Accordingly, the patient image sequence, which has been inputted to the first pattern determination system (the neural network based knowledge enhancement module) is transmitted to a second pattern determination system, which also (by utilising a neural network) detects and classifies any abnormality. Subsequently, the detection and classification determined by the two systems are examined against each other to judge whether the same pattern has been determined.
Also the dose plan is transmitted to the second pattern determination system which validates said dose plan by comparing said dose plan with a dose plan determined by the second pattern determination system so as to judge whether those dose plans are different.
In another preferred embodiment of the present invention the 3D RadioPlan module is used in connection with a radiation treatment device without means for the complete fixing of coordinates of the organ to be treated (e.g. a linear accelerator). As the person to be treated breaths and/or moves during the radio treatment the part of the human body to be treated changes shape and size during the treatment. This change of shape and size normally requires the space volume to be extended far beyond the volume of the actual diseased tissue in order to be able to guarantee treatment of the diseased tissue causing healthy tissue to be radiated. In order to minimise radiation of healthy tissue a neural network is trained to predict the, based on fix-points assigned to the patient instantaneous, shape, size and position of the space volume suggested.
In this situation a number of fix-points are assigned to the patient's body surface in the form of optically readable marks arranged in such a manner that a 3-dimensional coordinate of each fix-point is obtainable. Also, a reference co-ordinate system is defined to which all movements of the fix-points are referenced. Then the patient image sequence is obtained and each time an image is obtained the position of the fix-points are recorded and the image is transformed into the reference coordinate system, so that the position and size of each pixel in the image is described with reference to the reference co-ordinate system. Each transformation is based on the instantaneous position of the fix-points relative to the reference co-ordinate system. The hereby obtained patient image sequence may suitable be referred to as a normalised patient image sequence.
Now the normalised patient image sequence is ready for further processing and is input to the dose planning module which in turn suggests a normalised space volume to be included in the radio treatment focus (the word normalised is here used to emphasise that the space volume is suggested based on the normalised patient image sequnce). The hereby suggested space volume will be referenced to the reference co-ordinate system.
When the treatment is performed, the instantaneous position of the fix-points are recorded and the suggested normalised space volume is transformed according to the fix-point's position in the reference co-ordinate system (the transformation may be viewed upon as the reverse transformation of the transformation providing the normalised patient image sequence).
In this mode the neural network should be so constructed and trained so as it is able to normalise patient image sequences and able to perform the reverse transformation. These features are provided by training the neural network comprising the steps of -building a knowledge of the architectural variety of normal body parts and the fluctuation of coordinates (in 3D) of same relating to the fluctuation of coordinates (in 3D) of the above mentioned patient surface fix points
-building a knowledge of the architectural variety of normal body parts and the fluctuation of coordinates (in 3D) of same during breathing
-building an experience in predicting the fluctuations of coordinates (in 3D) of abnormal body parts (e.g. tumour) relating to the fluctuation of coordinates (in 3D) of the above mentioned patient surface fix points
-building an experience in predicting the fluctuations of coordinates (in 3D) of abnormal body parts (e.g. tumour) during breathing. Output may be used for controlling the movement of a patient (e.g. executing the above mentioned reverse transformation by moving the patient's bed exactly to keep the coordinates of the space volume to be treated within the focus of the radiation beams), or, if possible, a movement of the focus of the radiation beam during treatment.
A PREFERRED EMBODIMENT - AUTOMATED DIAGNOSIS OF AD
In another preferred embodiment of the present invention, the method is utilized for the automated diagnosis and/or the data normalizing and/or the quality and assurance control in the field of medical imaging.
An example of this utilization is the automated diagnosis of AD
Automated diagnosis of Alzheimer's Disease (AD) utilizing a neural network based software application capable of data pattern recognition and data generalization in the field of medical imaging data processing.
AD is most often diagnosed either by autopsy or after several years of progression. AD cannot be cured, but new types of medication can alter the progression of AD dramatically eleviating symptoms for several years, thus improving the quality of life for the patient and ensuring society health care savings.
Early diagnosis is possible through the PET (Positron Emission Tomography) technology, whereas a brain PET scanning will display a characteristic pattern of blood flow distribution in the brain. However, the display of this characteristic pattern is dependent on specific procedures followed during scanning and data processing. These techniques are refined and enhanced by the training of a neural network software application capable of data pattern recognition and data generalisation. The training sets are identified: The data from one PET scanning authorised by world leading scientists in the field of Alzheimer PET brain imaging is paired with clinically/by autopsy documented dementia diagnosis and the information is fed into the neural network as a training set. It is expected that a robust knowledge enhancement application will be built for the purpose of the discrimination of dementia diagnosis. The detection of early AD will be possible even if the traditional PET image is inconclusive. The database may import data for the training sets from other brain imaging modalities (CT/MRI) enabling the knowledge enhancement application to build experience from "tissue functional" imaging data (PET/SPECT/fMRI) as well as "tissue architectural" imaging data (CT/MRI). Similar to the embodiment mentioned above (3D RadioPlan) the procedure leading to the training of the neural network comprises the steps of
-building a knowledge of the architectural variety of normal human brains
-building an experience in recognising an architectural variety of abnormal human brains enabling the diagnosis of different types of dementia
-building a knowledge of the functional variety of normal brains -building an experience in recognising a functional variety of abnormal human brains enabling the diagnosis of different types of dementia
The patient image material database used for the training procedures is built from de novo scans, already existing in-house database resources or data downloaded from other diagnostic imaging centres.
Also in this example of the present invention it may be crucial to validate the method outlined above (the determination of the pattern). In this case the method for controlling the performance of a pattern determination system described in connection with the Flow Cytometry example is also applied to the medical imaging automated diagnosis method. Accordingly, the patient image sequence, which has been inputted to the first pattern determination system (the neural network based knowledge enhancement module) is transmitted to a second pattern determination system, which also (by utilising a neural network) detects and classifies any abnormality. Subsequently, the detection and classification determined by the two systems are examined against each other to judge whether the same pattern has been determined.
Also the diagnosis resulting from analysing the medical image is transmitted to the second pattern determination system which validates said diagnosis by comparing said diagnosis with a diagnosis determined by the second pattern determination system so as to judge whether those diagnoses are different.

Claims

1. A method for controlling the performance of a first pattern determination system adapted to determine a pattern of matter in a measured sample, said first pattern determination system is further adapted to exchange data with a second pattern determination system which also is adapted to determine a pattem of matter in a measured sample and is adapted to exchange data with the first pattern determination system, which method comprises the steps of providing to the first pattern determination system a first set of data representing a distribution of matter in the measured sample, determining a second set of data, by use of the first pattern determination system and based on the first set of data, representing a pattern in the distribution of matter, transferring the first and the second sets of data to the second pattern determination system, and determining, by use of the second pattern determination system and based on at least the first set of data, a fourth set of data representing a pattern in the distribution of matter, and - comparing the second set of data and the fourth set of data and determine whether the two sets of data are representing the same pattern of matter in order to judge whether the pattern of matter determined by the first pattern determination system is different from the pattern of matter determined by the second pattern system or vice versa.
2. A method according to claim 1 , which method further comprises the steps of providing a third set of data representing characteristics of the first set of data ι and - transferring the third set of data to the second pattern determination system.
3. A method according to claim 1 or 2, wherein the comparing of the second and fourth set of data comprises calculating the difference(s) between the two sets and wherein judgement on the two sets are based on the(these) difference(s).
4. A method according to any of the claims, wherein the two set, the second and fourth set, are judged to represent the same pattern of matter if thefthese) differences is(are) within certain limit(s).
5. A method according to any of the preceding claims, further comprising the steps of determining, by use of the second pattern determination system, a fifth set of data indicating changes to be applied to the first pattern determination system and/or indicating changes to be applied to a method used for providing a distribution of matter, and - transferring the fifth set of data to the first pattern determination system.
6. A method according to any of the preceding claims , which method further comprises the step of determining by use of the second pattern determination system a set of normalised data representing a normalisation of the third [sic: "third" ?] set of data, the determination being based on the second set of data.
7. A method according to any of the preceding claims, wherein a pattern is determined by inputting data representing a distribution of matter to a neural network, whereby the neural network returns a set of data representing the pattern represented in said set of data.
8. A method according to claim 7, wherein the neural network is a Bayesian network.
9. A method according to any of the claims 1-6, wherein a pattern is determined by matching, if possible, the first set of data with data of same type stored in a pre-compiled database storing data items each comprising data of said type and a pattern corresponding thereto.
10. A method according to any of the preceding claims, wherein patterns determined are patterns of antibody distributions in the measured sample.
11. A method according to any of the claims 1-9, wherein the patterns determined are patterns of abnormal (e.g. malignant) tissue in a human body to be treated by radiation.
12. A method according to claim 11 , wherein the patterns determined comprises a diagnosis and the method further comprises the step of determination of treatment, such as a dose plan, e.g. determination of volume(s) to be treated and radiation dose(s) to be exposed to said volume(s).
13. A method according to claim 12, further comprising the steps of transmitting the determined treatment to the second pattern determination system and comparing said transmitted treatment with a treatment determined by use of the second pattern determination system so as to judge whether those treatments are different from each other.
14. A method according claim 13, wherein the comparison and judgement are performed by the second pattern determination system.
15. A method according to any of the preceding steps, which method further comprises the step of transmitting acceptance data to the first pattern determination system, if the second set of data determined by first pattern determination system is judged not to be different from second set of date determined by the second pattern determination system or if these set are judged to be different from each other then transmitting rejection data to the first pattern determination system.
16. A method according to claim 15, wherein the method further comprises the steps of, if the acceptance data has not been received, such as received successfully, by the first pattern determination system from the pattern determination system, then transmitting criticality data to the second computer/pattern determination system and outputting a criticality response.
17. A method according to claim 16, wherein the criticality data is outputted to a second transmitter means for transmitting the criticality data to the second computer system/pattern determination system.
18. A method of determining a dose plan for radio treatment of a human body, said method is utilising a first pattern determinations system, which method comprises providing to the first pattern determination system, at least one patient image or image sequence wherein each image or image in said sequence is a 5 digitised/pixelised image of a cross section of a body, detect and classify, if possible, pattern(s) represented in the patient image or image sequence, said detection and classification being performed based on utilisation of a pattern recognition means run on a computer.
10 19. A method according to claim 18, which method further comprises and if a pattern is detected and classified determining by use of a planning means run on computer a dose plan for radio treatment of a human body, which dose plan comprises the space volume(s) to be included in the radio treatment focus and the radiation dose(s) to be exposed to
15 said volume(s) .
20. A method according to claim 19, wherein at least one patient image or image sequence is provided to the pattern determination system, said at least one patient image or image sequence is obtained by functional imaging modalities, such as PET, fMRI. 20
PCT/DK2001/000396 2000-06-09 2001-06-08 Quality and safety assurance of medical applications WO2001095253A2 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
AU2001263784A AU2001263784A1 (en) 2000-06-09 2001-06-08 Quality and safety assurance of medical applications

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
DKPA200000897 2000-06-09
DKPA200000897 2000-06-09

Publications (2)

Publication Number Publication Date
WO2001095253A2 true WO2001095253A2 (en) 2001-12-13
WO2001095253A3 WO2001095253A3 (en) 2002-05-10

Family

ID=8159550

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/DK2001/000396 WO2001095253A2 (en) 2000-06-09 2001-06-08 Quality and safety assurance of medical applications

Country Status (2)

Country Link
AU (1) AU2001263784A1 (en)
WO (1) WO2001095253A2 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114678117A (en) * 2022-05-26 2022-06-28 成都与睿创新科技有限公司 Management method and device for standardizing operating behaviors of operating room personnel

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1991018364A1 (en) * 1990-05-21 1991-11-28 Board Of Regents, The University Of Texas System Method for predicting the future occurrence of clinically occult or non-existent medical conditions
US5214715A (en) * 1991-01-31 1993-05-25 Trustees Of Boston University Predictive self-organizing neural network
US5689632A (en) * 1994-06-14 1997-11-18 Commissariat A L'energie Atomique Computing unit having a plurality of redundant computers
US5937202A (en) * 1993-02-11 1999-08-10 3-D Computing, Inc. High-speed, parallel, processor architecture for front-end electronics, based on a single type of ASIC, and method use thereof

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO1991018364A1 (en) * 1990-05-21 1991-11-28 Board Of Regents, The University Of Texas System Method for predicting the future occurrence of clinically occult or non-existent medical conditions
US5214715A (en) * 1991-01-31 1993-05-25 Trustees Of Boston University Predictive self-organizing neural network
US5937202A (en) * 1993-02-11 1999-08-10 3-D Computing, Inc. High-speed, parallel, processor architecture for front-end electronics, based on a single type of ASIC, and method use thereof
US5689632A (en) * 1994-06-14 1997-11-18 Commissariat A L'energie Atomique Computing unit having a plurality of redundant computers

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114678117A (en) * 2022-05-26 2022-06-28 成都与睿创新科技有限公司 Management method and device for standardizing operating behaviors of operating room personnel
CN114678117B (en) * 2022-05-26 2022-08-02 成都与睿创新科技有限公司 Management method and device for standardizing operating behaviors of operating room personnel

Also Published As

Publication number Publication date
AU2001263784A1 (en) 2001-12-17
WO2001095253A3 (en) 2002-05-10

Similar Documents

Publication Publication Date Title
Midtvedt et al. Quantitative digital microscopy with deep learning
Alsuliman et al. Machine learning and artificial intelligence in the service of medicine: Necessity or potentiality?
Chakraborty et al. Modified cuckoo search algorithm in microscopic image segmentation of hippocampus
CN105096225B (en) The analysis system of aided disease diagnosis and treatment, device and method
US8831327B2 (en) Systems and methods for tissue classification using attributes of a biomarker enhanced tissue network (BETN)
Thakoor et al. Robust and interpretable convolutional neural networks to detect glaucoma in optical coherence tomography images
JP2023501126A (en) Multi-instance learner for tissue image classification
CN109564617A (en) The method of characterization and imaging micro-object
Hu et al. Tumor tissue classification based on micro-hyperspectral technology and deep learning
Deshpande et al. A review of microscopic analysis of blood cells for disease detection with AI perspective
Hajirasouliha et al. Precision medicine and artificial intelligence: overview and relevance to reproductive medicine
Khang et al. Application of Computer Vision (CV) in the Healthcare Ecosystem
Fareed et al. ADD-Net: an effective deep learning model for early detection of Alzheimer disease in MRI scans
Alhazmi Detection of WBC, RBC, and platelets in blood samples using deep learning
Monteiro et al. Deep learning methodology proposal for the classification of erythrocytes and leukocytes
Sengupta et al. Intracranial hemorrhages segmentation and features selection applying cuckoo search algorithm with gated recurrent unit
Lapierre-Landry et al. Nuclei detection for 3d microscopy with a fully convolutional regression network
Olayah et al. Blood slide image analysis to classify WBC types for prediction haematology based on a hybrid model of CNN and handcrafted features
Li et al. Density center-based fast clustering of widefield fluorescence imaging of cortical mesoscale functional connectivity and relation to structural connectivity
Heni et al. Blood Cells Classification Using Deep Learning with customized data augmentation and EK-means segmentation
Bae et al. Transfer learning for predicting conversion from mild cognitive impairment to Dementia of Alzheimer’s type based on 3D-convolutional neural network
Mayala et al. Gubs: Graph-based unsupervised brain segmentation in mri images
Rathore et al. Prediction of stage of Alzheimer's disease Densenet Deep learning model
Monteiro et al. An artificial intelligent cognitive approach for classification and recognition of white blood cells employing deep learning for medical applications
WO2001095253A2 (en) Quality and safety assurance of medical applications

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A2

Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY BZ CA CH CN CO CR CU CZ DE DK DM DZ EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX MZ NO NZ PL PT RO RU SD SE SG SI SK SL TJ TM TR TT TZ UA UG US UZ VN YU ZA ZW

AL Designated countries for regional patents

Kind code of ref document: A2

Designated state(s): GH GM KE LS MW MZ SD SL SZ TZ UG ZW AM AZ BY KG KZ MD RU TJ TM AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE TR BF BJ CF CG CI CM GA GN GW ML MR NE SN TD TG

121 Ep: the epo has been informed by wipo that ep was designated in this application
DFPE Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101)
REG Reference to national code

Ref country code: DE

Ref legal event code: 8642

122 Ep: pct application non-entry in european phase
NENP Non-entry into the national phase

Ref country code: JP