Diagnostic modeling method and device

Abstract

The present disclosure relates to a system (100) for generating a diagnostic model. The system (100) includes a processor (108) configured to analyze a plurality of reference data sets. Each reference data set includes at least one image (230, 240). The analysis identifies at least one feature (230, 240) in each image. A metric is calculated based on at least one identified feature. Result data associated with at least some of the reference data sets is obtained. A diagnostic model is compiled based on at least one calculated metric and the associated result data. The present disclosure also relates to a method for generating a diagnostic model and a non-transitory computer readable medium.

Description

Technical Field

[0001] The present disclosure relates to diagnostic modeling methods and apparatus. More particularly, but not exclusively, the present disclosure relates to systems and methods for generating diagnostic models. Background Art

[0002] Two-dimensional (2D) echocardiography is an imaging technique by which the motion of the heart can be assessed in different states, such as at rest or under stress. The analysis can be performed in other states, including, for example, during an intermediate stress phase and / or a recovery phase. This can highlight areas of poor or abnormal cardiac function, thereby identifying patients who may need medical intervention. A typical model of the left ventricle includes sixteen (16) segments, which can be seen using different 2D images of the heart. Other models of the left ventricle can, for example, include seventeen (17) segments. The apical inferior segment, mid inferior segment, basal inferior segment, basal inferior segment, basal inferior segment, mid anterior segment, and basal anterior segment can be seen in an apical two chamber image. The apical septum segment, mid lateral segment, basal septum segment, apical lateral segment, mid lateral segment, and basal lateral segment can be seen in the apical four chamber image. The anteroseptum segment, anteroseptum segment, mid inferior segment, mid anterior segment, anterolateral segment, and inferolateral segment can be seen in the parasternal short axis image. The apical lateral segment, mid inferolateral segment, basal inferolateral segment, apical septum segment, midseptum segment, and basal septum segment can be seen in the apical two chamber image (or parasternal long axis image).The behavior of each segment can be viewed in different parts of the left ventricle. Currently, the movement of each segment of the myocardium under different states (e.g., at rest and under stress) is determined by interpretation of 2D echocardiographic data by a specialized cardiologist. This is performed in a classified manner. For example, each portion of the myocardial wall may be classified as having one of the following reference wall motion scores: normal ("1"), hypokinetic ("2"), akinetic ("3"), dyskinesia ("4"), and unscoreable ("X"). Other classifications may be used, such as defining five (5) or seven (7) discrete scores for each segment. Known techniques rely on subjective classification and can prove to be a time consuming exercise.

[0003] It is recognized that image quantification tools need to account for the following factors: (i) changes in underlying disease pathophysiology over time; (ii) disease pathology varies by geographic location and the nature of the patient population being referred for care changes; and (iii) evolving understandings of what is defined as disease or causes disease.

[0004] The present invention, at least in certain embodiments, seeks to provide an improved method and apparatus for generating diagnostic models. Summary of the invention

[0005] Aspects of the invention relate to a system for generating a diagnostic model, a method for generating a diagnostic model and a non-transitory computer readable medium as claimed in the appended claims.

[0006] According to another aspect of the present invention, a system for generating a diagnostic model is provided, the system comprising a processor, the processor being configured to:

[0007] analyzing a plurality of reference data sets, each reference data set comprising at least one image, the analyzing comprising identifying at least one feature in each image;

[0008] calculating at least one metric based on at least one identified feature;

[0009] obtaining result data associated with at least some of the reference data sets; and

[0010] A diagnostic model is compiled based on at least one calculated metric and associated outcome data. The use of outcome data establishes a feedback loop that can be used to refine and develop the diagnostic model to reflect different outcome scenarios. The outcome data can be used in a machine learning algorithm to adjust thresholds and / or weights. The outcome data may potentially help identify new biomarkers for diagnosis. The outcome data can be used to modify or adjust the algorithm used to generate the diagnostic model.

[0011] In at least some embodiments, the reference data sets may each include multiple images. Each reference data set may include a first image and a second image. The processor may be configured to analyze each reference data set to identify at least one first feature in the first image and to identify at least one second feature in the second image, each at least one first feature being paired with a corresponding one of the at least one second features. The processor may be configured to analyze each reference data set to compare each pair of corresponding first features and second features to identify one or more differences therebetween. Each pair of corresponding first features and second features may relate to the same features in the first image and the second image. The first image and the second image may relate to the same area but may be acquired at different times.

[0012] The one or more differences identified between the first feature and the second feature may, for example, include one or more of: opacity, brightness, contrast, cross-sectional area, size (in one, two, or three dimensions), position (in one, two, or three dimensions), and orientation (about one, two, or three axes).

[0013] The processor may be configured to calculate at least one metric based on the one or more differences identified between each corresponding pair of the first feature and the second feature.

[0014] The system may be adapted to generate a diagnostic model for diagnosing a cardiac condition. The first image may include a first end-systolic image, and the second image may include a second end-diastolic image. One or more differences identified between each pair of corresponding first and second features may represent cardiac cycle variations. At least one metric may be calculated based on the identified cardiac cycle variations with respect to each reference data set.

[0015] The processor may be configured to label the image with a label based on the result data. The labels may distinguish different classifications. Each label may, for example, include a classification indicating the presence or absence of a condition or disease. Each label may include a grade indicating the severity of the condition, for example including an indication of the severity of a stenosis.

[0016] The result data may include diagnostic information. The diagnostic information may include a record of a cardiac event (eg, myocardial infarction).

[0017] The result data may include records of stenosis greater than a threshold. The threshold may be defined as a percentage of stenosis. The threshold may be defined, for example, as 60%, 70%, 80% or 90%.

[0018] The outcome data may be generated after the echocardiographic image is acquired. For example, the outcome data may be generated 6 months, 12 months, 18 months, 24 months, or more after the echocardiographic image is acquired.

[0019] The processor may be configured to update the diagnostic model when result data becomes available or when result data is updated.

[0020] The processor may be configured to determine weights for at least some of the calculated metrics.Each weight may be determined based on result data associated with one of the given reference data sets.

[0021] A diagnostic model may be compiled from the calculated metrics and associated weights.

[0022] The processor may be configured to analyze other reference data sets. The additional reference data sets may each include at least one image. The processor may be configured to update the diagnostic model based on the analysis of the additional reference data sets.

[0023] At least one image may include an ultrasound image. At least one image may include an echocardiogram image. Alternatively or additionally, at least one image may include a magnetic resonance image (MRI) or a computed axial tomography (CAT) image.

[0024] According to another aspect of the present invention, there is provided a method for generating a diagnostic model, the method comprising:

[0025] analyzing a plurality of reference data sets, each reference data set comprising at least one image, the analyzing comprising identifying at least one feature in each image;

[0026] calculating at least one metric based on the at least one identified feature;

[0027] obtaining result data associated with at least some of the reference data sets; and

[0028] A diagnostic model is compiled based on the calculated metrics and the associated outcome data.

[0029] In at least some embodiments, the reference data sets may each include multiple images. Each reference data set may include a first image and a second image. The method may include analyzing each reference data set to identify at least one first feature in the first image, and to identify at least one second feature in the second image, each at least one first feature being paired with a corresponding one of the at least one second feature. The method may include comparing each pair of corresponding first features and second features to identify one or more differences therebetween. Each pair of corresponding first features and second features is associated with the same feature identified in the first image and the second image. The first image and the second image may relate to the same area, but may be acquired at different times. The one or more differences identified between the first feature and the second feature may, for example, include one or more of: opacity, brightness, contrast, cross-sectional area, size (in one, two, or three dimensions), position (in one, two, or three dimensions), and orientation (about one, two, or three axes).

[0030] The method may include calculating at least one metric based on one or more differences identified between each corresponding pair of first and second features.

[0031] The system may be adapted to generate a diagnostic model for diagnosing a cardiac condition. The first image includes a first end-systolic image, and the second image includes a second end-diastolic image. One or more differences identified between each pair of corresponding first and second features represent cardiac cycle variations.

[0032] The result data can be used to label the corresponding reference data set with a label. The labels can distinguish different classifications. Each label can, for example, include a classification indicating the presence or absence of a condition or disease. Each label can include a grade indicating the severity of the condition, for example including an indication of the severity of stenosis.

[0033] The result data may include diagnostic information. The diagnostic information may include a record of a cardiac event (eg, myocardial infarction).

[0034] The resulting data may include records of stenosis greater than a threshold value.

[0035] The outcome data may be generated after the echocardiographic image is acquired. For example, the outcome data may be generated 6 months, 12 months, 18 months, 24 months, or more after the echocardiographic image is acquired.

[0036] The method may include updating the diagnostic model when the outcome data is updated and / or when new outcome data becomes available.

[0037] The method may include determining weights for at least some of the calculated metrics. Each weight may be determined based on result data associated with one of the given reference data sets. A diagnostic model may be compiled based on the calculated metrics and associated weights.

[0038] The method may include incrementally adding further reference data sets. The method may include analyzing further reference data sets. The method may include updating the diagnostic model based on the analysis of other reference data sets.

[0039] At least one image may include an ultrasound image. At least one image may include an echocardiogram image. Alternatively or additionally, at least one image may include a magnetic resonance image (MRI) or a computed axial tomography (CAT) image.

[0040] According to another aspect of the present invention, there is provided a system for generating a diagnostic model for diagnosing a cardiac condition, the system comprising a processor configured to:

[0041] analyzing a plurality of reference data sets, each reference data set comprising at least a first echocardiographic image and a second echocardiographic image, the analyzing comprising identifying an end-systolic image and an end-diastolic image within each reference data set;

[0042] comparing the end-systolic and end-diastolic images in each reference dataset to identify cardiac cycle changes;

[0043] calculating at least one metric based on the identified cardiac cycle variations with respect to each reference data set;

[0044] obtaining result data associated with at least some of the reference data sets; and

[0045] A diagnostic model is compiled based on the calculated metrics and associated outcome data.

[0046] The result data may be acquired within a period of time after the at least one image is acquired. For example, the result data may be generated 6 months, 12 months, 18 months, 24 months, or more after the at least one image is acquired.

[0047] According to one aspect of the present invention, there is provided a method for generating a diagnostic model for diagnosing a cardiac condition, the method comprising:

[0048] analyzing a plurality of reference data sets, each reference data set comprising at least a first echocardiographic image and a second echocardiographic image, the analyzing comprising identifying an end-systolic image and an end-diastolic image within each reference data set;

[0049] comparing the end-systolic and end-diastolic images in each reference dataset to identify cardiac cycle changes;

[0050] With respect to each reference data set, at least one metric is calculated based on the identified cardiac periodic changes;

[0051] obtaining result data associated with at least some of the reference data sets; and

[0052] A diagnostic model is compiled based on the calculated metrics and the associated outcome data.

[0053] According to another aspect of the present invention, a non-transitory computer-readable medium having a set of instructions stored therein is provided, which when executed causes a processor to perform the method described herein.

[0054] Any control unit or controller described herein may suitably include a computing device having one or more electronic processors. The system may include a single control unit or electronic controller, or alternatively, the different functions of the controller may be embodied in or carried in different control units or controllers. As used herein, the term "controller" or "control unit" will be understood to include a single control unit or controller and a plurality of control units or controllers that operate together to provide any of the control functions. In order to configure a controller or control unit, an appropriate instruction set may be provided, and when the instruction is executed, the control unit or computing device implements the control technology specified herein. The instruction set may be suitably embedded in the one or more electronic processors. Alternatively, the instruction set may be provided as software to be executed on the computing device stored in one or more memories associated with the controller. The control unit or controller may be implemented with software running on one or more processors. One or more other control units or controllers may be implemented with software running on one or more processors (optionally, one or more processors identical to the first controller). Other suitable arrangements may also be used.

[0055] Within the scope of the present application, it is expressly intended that the various aspects, embodiments, examples and alternatives, and in particular the individual features thereof, set forth in the preceding paragraphs, claims and / or in the following description and drawings may be used independently or in combination. That is, all embodiments and / or features of any embodiment may be combined in any manner and / or combination, unless such features are incompatible. The applicant reserves the right to change any originally filed claim or to file any new claim accordingly, including the right to amend any originally filed claim to be dependent on and / or combined with any feature of any other claim, even though not originally filed in this manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] One or more embodiments of the present invention will now be described, by way of example only, with reference to the accompanying drawings, in which:

[0057] Figure 1 is a schematic diagram of an echocardiography system according to an embodiment of the present invention;

[0058] Figure 2 A four-chamber view of the heart is schematically shown;

[0059] Figure 3 It is shown by Figure 1 A flow chart of the main steps of the diagnostic method performed by the system;

[0060] Figure 4A Shown by Figure 1 End-systolic image captured by the echocardiography system shown;

[0061] Figure 4B Shown by Figure 1 End-diastolic image captured by the echocardiography system shown;

[0062] Figure 5A A contour data set consisting of end-systolic contour points and end-diastolic contour points is shown;

[0063] Figure 5B Shows from Figure 5A The multiple elements generated by the contour data set shown;

[0064] Fig. 6A Shown for Figure 5B Generation of area measures for each element shown;

[0065] Figure 6B Shown for Figure 5B Generation of the average distance metric for each element shown;

[0066] Figure 6C Shown for Figure 5B Generation of rectangularity measures for each element shown;

[0067] Fig.6D Shown for Figure 5B Generation of solidity measures for each element shown;

[0068] Fig. 7A Analysis of the area metric for the resting state is shown;

[0069] Figure 7B Analysis of the mean distance metric for the resting state is shown;

[0070] Figure 7C An analysis of the rectangularity metric for the resting state is shown;

[0071] Fig.7D Analysis of the solidity measure for the resting state is shown;

[0072] Fig. 8A Analysis of area metrics for stress states is shown;

[0073] Figure 8B An analysis of the distance metric for stress states is shown;

[0074] Figure 8C An analysis of the rectangularity metric for stress conditions is shown;

[0075] Fig.8D Analysis of the robustness measures for stress conditions is shown;

[0076] Fig.9A Reference data models based on bivariate analysis of normalized area and mean distance metrics are shown;

[0077] Fig. 9B Shown is applied to Fig.9A Normally distributed data set for the reference data model shown;

[0078] Fig. 10A Multivariate analysis for each metric of the first resting state in the apical 2-chamber image is shown;

[0079] Fig. 10B Multivariate analysis for each metric of the second resting state in the apical 4-chamber image is shown;

[0080] Fig.11A Multivariate analysis for each measure of the first stress state in the apical 2-chamber image is shown;

[0081] Fig. 11B Multivariate analysis for each measure of the second stress state in the apical 4-chamber image is shown;

[0082] Fig. 12A Scoring of elements in the resting state is shown;

[0083] Fig. 12B shows the scoring of the elements under stress conditions; and

[0084] Fig.13 A decision tree consisting of a series of decision points defining thresholds is shown. DETAILED DESCRIPTION

[0085] An echocardiography system 100 according to an embodiment of the present invention will now be described with reference to the accompanying drawings. The echocardiography system 100 is operable to analyze images of a heart 200 and score cardiac cyclic motion.

[0086] like Figure 1As shown, the echocardiography system 100 includes: a transducer array 102, which is arranged to be located near the body of a patient 104, usually as close to the heart as possible; a processing unit 106, which includes a processor 108, which can be a digital electronic processor, a memory 110 such as a hard disk, and a display 112 such as a flat-screen display or an LED display. The system may further include a user input device, such as a touch screen 114 integrated into the display 112, which provides user input, allowing the user to provide input to the echocardiography system 100. Of course, other user inputs can be used, such as a mouse, a touch pad, or a keyboard. The processing unit 106 is connected to the transducer array 102 and is arranged to control the transducer array as a phased array so as to transmit an ultrasound beam that scans the entire patient in a series of pulses and detects the reflected ultrasound waves from the heart of each pulse. A single scan of the heart creates a single image, and the scan is usually repeated at a rate of 25 to 50 images per second to create a real-time video image showing the movement of the heart during the cardiac cycle. Each image may be stored in memory 110 as an image data set, which may include, for example, an intensity value for each pixel making up the image. Although the system is described herein in general terms, suitable echocardiography systems include, for example, a Philips Epic iE33, a GE Vivid e9, or a portable or handheld system such as a Philips CX50.

[0087] The process of echocardiography is well known and will not be described in detail herein. There are several different imaging methods, but the echocardiography system 100 according to the present embodiment uses two-dimensional imaging. It is known to provide images in several different planes through the heart, which images show different aspects of the four main chambers of the heart, namely the left ventricle (LV), the right ventricle (RV), the left atrium (LA) and the right atrium (RA). These views include, for example, apical 4-chamber images, apical 2-chamber images, apical 3-chamber images, and parasternal long-axis views and parasternal short-axis views. In each case, although a single still image can be obtained, it is generally necessary to obtain a series of views throughout the cycle of the heart so that its movement can be recorded and analyzed. The echocardiography system 100 can score the cardiac cyclic motion of the heart 200 using one or more of the above views.

[0088] As an example, in Figure 22 shows an apical 4-chamber view of a heart 200. The image includes a 2D plane of the heart 200. The image shows a left ventricle (LV) 202, a right ventricle (RV) 204, a left atrium 206, a right atrium 208, and a septum 210. In the apical 4-chamber view, an apex 212, a side wall 214, a base 216, and an inner wall 218 of the left ventricle 202 are also visible. A longitudinal axis 220 of the left ventricle 202 extends through the apex 212. The left ventricle 202 has a first side 222 and a second side 224 disposed on opposite sides of the longitudinal axis 220.

[0089] The processing unit 106 analyzes the apical 4-chamber image to implement the scoring techniques described herein. Alternatively or in addition, the processing unit 106 may utilize one or more of the following: an apical 2-chamber image, a parasternal short-axis image, and an apical 3-chamber image. The processing unit 106 may use other echocardiographic images to implement the scoring techniques described herein. The processing unit 106 may use various combinations of echocardiographic images to score sixteen (16) segments of the left ventricle. The processing unit 106 may analyze multiple images and score the cardiac cyclic motion based on metrics of the multiple images. The processing unit 106 may qualitatively evaluate the available images and prioritize images that are determined to provide a clearer representation of the cardiac cyclic motion of a particular portion of the heart wall. A Cartesian coordinate system is defined as including a vertical axis (herein referred to as the y-axis) extending through the apex 212 of the left ventricle 202 and extending along its longitudinal axis and a horizontal axis (herein referred to as the x-axis) passing through the midpoint of the left ventricle 202 and located midway between the apex 212 and the base 216.

[0090] exist Figure 3 A block diagram representing the operation of the echocardiographic system 100 is shown in . The echocardiographic system 100 is arranged to acquire a sequence of 2D images and store them in the memory 110 (box 300). The images may be acquired over a single cardiac cycle and may include, for example, between ten (10) and fifty (50) images covering one cycle. The echocardiographic system 100 may perform a single scan or more than one scan. For example, the echocardiographic system 100 may perform a first scan and a second scan. The first scan may be performed when the patient is at rest and the second scan may be performed when the patient is under stress. The echocardiographic system 100 may optionally perform one or more intermediate scans between the resting state and the stressed state, for example during a recovery phase, i.e., during a period when the heart rate returns to normal after being stressed. The acquisition of the images may be performed on a conventional echocardiographic system. Subsequent analysis of the images may be performed using a device that constitutes Figure 1The images may be processed by the same processing unit 106 which is part of the echocardiography system shown. However, the images may be downloaded to a computer (e.g. a laptop or PC) having a processor, memory, user input and display, which operates in the same manner as the processing unit 106 for this purpose, and further analysis of the images may be performed on this computer under the control of dedicated software. It will be appreciated that the images may be retrieved from a PACS (Picture Archiving and Communication System). Alternatively or additionally, the images may be sent to an external server for processing. The images may be anonymized prior to transmission.

[0091] For the left ventricle 202, an image closest to end systole (i.e., maximum contraction during the cardiac cycle) and an image closest to end diastole (i.e., largest volume during the cardiac cycle) are identified (block 302). This may be accomplished by a user viewing the images on the display 112 and selecting a first image 230 closest to end systole (referred to herein as the end systolic image 230), and selecting a second image 240 closest to end diastole (referred to herein as the end diastolic image 240). The end systolic image 230 and the end diastolic image 240 are acquired at a first time and a second time, respectively, during the cardiac cycle. Figure 4A An exemplary end-systolic image 230 is shown in Figure 4B 240. The selection of the end-systolic image 230 and the end-diastolic image 240 may be made by the user based on an assessment and comparison of the volume of the left ventricle 202 in each image as judged by eye, or by noting the opening and closing points of the mitral valve, or by using the QRS complex on an ECG plot, or by any combination of these. Alternatively, the processor 108 may be arranged to use image processing techniques to identify the end-systolic image 230 and the end-diastolic image 240. The image processing techniques may, for example, determine the volume of the left ventricle 202 in each image. The processor may identify the image with the smallest left ventricle volume as the end-systolic image 230; and identify the image with the largest left ventricle volume as the end-diastolic image 240. Alternatively, the image processing techniques may identify and track the motion of image elements that persist across multiple images to identify the end-systolic image 230 and the end-diastolic image 240. For example, a reversal in the direction of motion of a continuous image element may be used to identify end-systole and end-diastole.End-systolic image 230 and end-diastolic image 240 are identified in memory 110, such as marked with appropriate markings, so that they may be selected and viewed by a user.

[0092] The inner wall 218 of the left ventricle 202 is identified at the end-systole in the end-systole image 230 and at the end-diastole in the end-diastole image 240 (block 304). The left ventricle 202 is contoured (or mapped) at the end-diastole in the end-systole image 230 and at the end-systole in the end-diastole image 240 (block 306). The contouring of the left ventricle 202 includes identifying a plurality of end-systolic contour points 232-n around the inner wall 218 in the end-systolic image 230 and identifying a plurality of end-diastolic contour points 242-n around the inner wall 218 in the end-diastolic image 240. A first continuous curve is drawn between the end-systolic contour points 232-n to form an end-systolic contour line 233; a second continuous curve is drawn between the end-systolic contour points 242-n to form an end-diastolic contour line 243. The end-systolic contour line 233 and the end-diastolic contour line 243 may include straight lines and / or curved lines. End-systolic contour line 233 and end-diastolic contour line 243 may, for example, be contoured to match boundaries identified in end-systolic image 230 and end-diastolic image 240 , respectively.

[0093] The end-systolic contour points 232-n and the end-systolic contour line 233 form an end-systolic contour data set 234; and the end-diastolic contour points 242-n and the end-diastolic contour line 243 form an end-diastolic contour data set 244. Each end-systolic contour point 232-n in the end-systolic contour data set 234 is paired with one of the corresponding end-diastolic contour points 242-n in the end-diastolic contour data set 244. The resulting pairing of the end-systolic contour points 232-n and the end-diastolic contour points 242-n represents the change in the motion of the wall of the heart 200 during the cardiac cycle. The pairing of the end-systolic contour points 232-n and the end-diastolic contour points 242-n may correspond to the same feature of the left ventricle 202, although they are located at different locations in the end-systolic image 230 and the end-diastolic image 240 due to wall motion during the cardiac cycle. In the present embodiment, thirteen (13) end-systolic contour points and end-diastolic contour points are identified in the end-systolic image 230 and the end-diastolic image 240. The end-diastolic contour points and the end-systolic contour points are labeled 1 to 13 (i.e., n=1, 2, 3, ... 13) according to their locations along the endocardium. Figure 5A As shown, the end-systolic data set 234 and the end-diastolic contour data set 244 are combined.

[0094] like Figure 5B As shown, the end-systolic contour points 232-n and the end-diastolic contour points 242-n form a plurality of elements E representing the cardiac cycle motion of the inner wall. i (where i is an integer) (block 308). As described herein, for element E iThe analysis and scoring are performed to grade the cardiac cycle motion of the corresponding portion of the inner wall 218 of the left ventricle 202. In the illustrated arrangement, three (3) pairs of end-systolic contour points 232-n and end-diastolic contour points 242-n (i.e., three (3) end-systolic contour points 232-n and three (3) end-diastolic contour points 242-n from the corresponding end-systolic data set 234 and end-diastolic contour data set 244-n) define each element E. i In the example shown, the end-systolic image 230 and the end-diastolic image 240 are apical 2-chamber images. The changes in the cardiac cyclic motion of the inner wall are represented by six (6) elements E 1-6 Indicates. Element E 1-6 Each of the elements E is in the form of a planar (two-dimensional) polygon. 1-6 Corresponding to the various segments of the left ventricle model. In particular, the first element E1 corresponds to the basal lower segment; the second element E2 corresponds to the middle lower segment; the third element E3 corresponds to the apical lower segment; the fourth element E4 corresponds to the apical anterior segment; the fifth element E5 corresponds to the anterior middle segment; and the sixth element E6 corresponds to the basal anterior segment. It will be appreciated that the cardiac cycle variation may be represented by different numbers of elements E i Represents, for example, less than six (6) elements or more than six (6) elements.

[0095] Analysis element E i To generate at least one wall motion metric to score (ie, classify or grade) the cardiac cycle motion of a corresponding portion of the heart 200 (block 310). Element E is described in more detail herein. i The generated metrics are compared to a predefined reference data model to score the wall motion (box 312). The scoring results are then output, for example, to a screen or display (box 314). The score can be reviewed by a clinician.

[0096] The contouring of the left ventricle 202 will now be described in more detail. The contouring may be performed by an echocardiographer; or using suitable image processing techniques. Figure 4A and 4B FIG. 2 shows an echo image of the left ventricle 202 acquired with a contrast agent. Figure 4A The end-systolic image 230 is shown in FIG. Figure 4B24. The end-diastolic image 240 is shown in FIG. The apex 212 of the left ventricle 202 can be located at the top of the left ventricle 202, and the base 216 of each side 222, 224 can be located by the shape of the inner wall 218. The longitudinal axis (Y) is defined as a reference line passing through the apex 212 and the midpoint between the base points of the two sides 222, 224. The x-axis can then be defined as a line perpendicular to the y-axis that is midway between the apex and the midpoints on both sides of the base 216. The midpoint on each side 222, 224 can be identified as the point where the x-axis intersects the sidewall on that side 222, 224. The intermediate end-systolic contour points 232-n and the end-diastolic contour points 242-n can be identified by dividing the area between the apex 212 and the midpoint on each side 222, 224, and by dividing the area between the midpoint and the base point on each side 222, 224.

[0097] As described above, each of these end-systolic contour points 232-n and end-diastolic contour points 242-n can be identified by a user. Alternatively, image processing can be used to identify the end-systolic contour points 232-n and end-diastolic contour points 242-n. If image processing is used, the contour of the left ventricle 202 is first identified as the boundary between the brighter area within the left ventricle 202 and the darker area of ​​the myocardium that forms the wall around it (or vice versa for images obtained without contrast agents). Suitable algorithms for identifying such boundaries are well known. Once the boundary is identified, the algorithm can be arranged to identify the highest point (maximum y value) of the boundary as the vertex 212, and the point where the boundary changes direction at the lower end as the base point 216. Similarly, algorithms for analyzing the radius and direction of curvature and how it changes around the boundary can be used to identify these points and the points at the lower end of the vertex 212. The coordinates of each of the end-systolic contour points 232-n and end-diastolic contour points 242-n are determined with reference to a coordinate system. The scale of the images acquired by the echocardiography system 100 is known. Therefore, the coordinates of each end-systolic contour point 232-n and end-diastolic contour point 242-n define the location of the point in the corresponding image plane. The distance between each pair of contour points represents the distance that the corresponding portion of the heart 200 moves between end-systole and end-diastole.

[0098] Now refer to Figures 6A to 6D Describe the element E i to generate wall motion metrics. Fig. 6A As shown, each element E i is a polygon with n sides. Fig. 6A As shown, each element E i The element area A of is calculated based on the first contour data set 234 and the second contour data set 244 by the shoelace formula:

[0099]

[0100] where n = the number of sides of the polygon; and

[0101] (X i ,Y i ) = vertices of polygon (i = 1, 2, ..., 6).

[0102] Then each element E i The calculated area A is normalized to the total area of ​​the end-diastolic contour points and represents a fraction of the total area.

[0103] like Figure 6B As shown, the Euclidean distance (d) between each pair of end-diastolic and end-systolic contour points 232-n and the end-diastolic contour point 242-n is calculated using the following equation:

[0104]

[0105] Where: n = number of dimensions;

[0106] p = coordinates of the end-diastolic contour point; and

[0107] q = coordinates of the end-systolic contour point.

[0108] Then calculate each element E i The average distance (d) of the first element E1 For the second element E2 And so on). The mean distance was then normalized to a fraction of the total circumference of the end-diastolic contour points.

[0109] like Figure 6C As shown, each element E i The rectangularity is calculated as the area of ​​each element (A i ) and its minimum bounding rectangle (R i )’s area ratio:

[0110]

[0111] Among them: A i =Each element E i area; and

[0112] R i = Area of ​​the minimum bounding rectangle.

[0113] like Fig.6D As shown, each element E i The firmness S iCalculated as the area of ​​the element (A i ) and the area of ​​its convex hull (H i ) is the ratio between:

[0114]

[0115] Among them: A i =Each element E i area; and

[0116] H i = the area of ​​the corresponding convex hull.

[0117] To evaluate the correlation between the calculated metrics and the reference wall motion scores, a reference data set comprising raw (i.e., unprocessed) two-dimensional echocardiographic data was analyzed. The reference data set consisted of historical data including end-diastolic images and end-systolic images for a group of patients. The end-diastolic images and end-systolic images were analyzed according to the techniques described herein to identify end-systolic contour points 232-n and end-diastolic contour points 242-n. Elements E corresponding to various segments of the standard model of the left ventricle 202 were thereby identified. i . Using the techniques described in this article, element E i The following metrics are calculated: normalized area A, normalized average distance d, rectangularity and solidity Si. For each element E i Resting and stressed state measures are generated. Element E i The scores were also scored separately by two cardiologists using a standard scoring system consisting of the reference wall motion scores: normal ("1"), hypokinetic ("2"), akinetic ("3"), dyskinetic ("4"), and unscoreable ("X"). Any scores that differed between the two reference datasets were reviewed and a consensus was reached. Elements with a wall motion score of "X" were removed from the reference dataset (n = 2). Due to the small number of elements in the available reference dataset with a wall motion score of "4" (n = 2), these elements were also removed from the analysis. Therefore, in this example, each element E from the reference dataset was scored as i Scores are normal ("1"), hypokinetic ("2"), and akinetic ("3"). The analysis is repeated for each patient in both the resting state and the stressed state. As described herein, the metrics calculated from the analysis of the original reference data set are used to generate a reference data model, to which the calculated metrics can be compared. For each element E iA reference data model is generated. The reference data model may be a univariate model or a multivariate model. The reference data model may be stored in a memory 110 of the echocardiography system 100. Alternatively, the reference data set may be stored in the memory 110 and the reference data model generated by the processing unit 106. The method may enable dynamic comparison, such as dynamic comparison with respect to a particular metric or combination of metrics.

[0118] exist FIG. 7A to FIG. 7D The values ​​assigned by the cardiologists for each element E in the resting state are shown in i Graph of the calculated metric and wall motion score (normal "1", hypokinetic "2", akinetic "3"). The second quartile point of the metric is represented by a box plot containing the midline of the metric set for each reference wall motion score. Fig. 7A The normalized element area A for each reference wall motion score is shown in Figure 7B The normalized mean distance d for each reference wall motion score is shown in Figure 7C The calculated firmness for each reference wall motion score is shown in Fig.7D The rectangularity calculated for each reference wall motion score is shown in . The Wilks-Lambda nonparametric multivariate test statistic with P < 0.05 was determined in each wall motion score set. Statistically significant correlations were identified between the calculated metrics and the assigned wall motion scores.

[0119] exist Figures 8A-8D The values ​​assigned by the cardiologists for each element E under stress are shown in i A plot of the calculated metric and the wall motion score (i.e., normal "1", hypokinetic "2", akinetic "3"). The second quartile point of the metric is represented by a box plot containing the midline of the metric set for each reference wall motion score. Fig. 8A The normalized element area A for each reference wall motion score is shown in Figure 8B The normalized mean distance d for each reference wall motion score is shown in Figure 8C The calculated firmness for each reference wall motion score is shown in Fig.8D The rectangularity calculated for each reference wall motion score is shown in . The Wilks-Lambda nonparametric multivariate test statistic with P < 0.05 was determined in each wall motion score set. Statistically significant correlations were identified between the calculated metrics and the assigned wall motion scores.

[0120] Each element E can be calculated according to one of the calculated metrics iThe processing unit 106 may be configured to define a univariate distribution, such as a univariate normal distribution. For example, the score may correspond to a z-score (standard score) of one of the calculated metrics. The z-score indicates how many standard deviations the calculated metric is from the population mean in units of standard deviation. The processing unit 106 may be configured to assign a z-score to each element E corresponding to the determined z-score. i However, referring to two or more calculated measures can improve the performance of each element E i The accuracy of the calculated score. Processing unit 106 can be configured to define a multivariate distribution, such as a multivariate normal distribution. Processing unit 106 can be configured to define a bivariate distribution or a higher dimensional distribution. Processing unit 106 can be configured to calculate the "distance" of the calculated metric from a reference population. This technique allows analysis to be performed in higher dimensions. One approach is to use the distance from the first principal component of the data PC1. Alternatively or additionally, the Mahalanobis distance can be calculated by processing unit 106. Other statistical analysis techniques are also applicable.

[0121] By determining the average distance and each element E i The correlation between the areas of can be determined corresponding to each element E i In this embodiment, the score includes a continuous scale, rather than a traditional scoring system that relies on discrete values. A simulated representation of the correlation between the z-scaled element area and the average distance is as follows Fig.9A The lower left quadrant Q3 represents those elements E identified as having potential abnormal wall motion i . A set of three (3) ellipses E1 to E3 represent the confidence intervals (CI) of the distribution of the data points: the inner ellipse E1 represents CI = 68%, the middle ellipse E2 represents CI = 95%, and the outer ellipse E3 represents CI = 99%. The centroid O of the data is shown; the line PC1 represents the orthogonal regression line through the data (i.e., the first principal component of the data PC1). A set of markers M1-M3 represent the data points being scored. Fig. 9B A set of randomly generated, normally distributed data with a covariance of 0.56 (n=1,000) is shown in . Each data point is colored according to a continuous score determined by the analysis techniques described herein.

[0122] According to this embodiment, the processing unit 106 is implemented to process each element E i A continuous method of scoring. As described in this article, according to each element E i For each element E derived from the end-systolic image 230 and the end-diastolic image 240 i This description focuses on the six (6) elements E corresponding to the segments visible in the standard phantom of the apical 2-chamber image.i It will be appreciated that, for example, by analyzing the apical 3 chamber image and / or the apical 4 chamber image, additional elements E corresponding to other segments of the left ventricle 202 may be targeted. i The same technique is implemented. The analysis is performed independently with respect to the end-systolic image 230 and the end-diastolic image 240 acquired for the rest and stress states. The processing unit 106 may compare the analysis results with respect to the rest and stress states.

[0123] exist Fig.9A and 9B The bivariate analysis of the element E according to the two calculated measures is visualized in i In the current case, the metric is the score of each element E i The normalized area and average distance of the element E determined by analyzing the reference data set i The normalized area and average distance of E are used to generate the reference data model. In the present case, only those elements of the wall motion that are assigned a score of "1" are included in the generation of the reference data model. In order to generate a given element E i The processing unit 106 calculates the score of the element E i The processing unit 106 calculates the element E identified by analyzing the end-systolic image 230 and the end-diastolic image 240 of the patient. i The normalized area and average distance of this paper are Fig.9A and 9B The described embodiment utilizes a i It will be appreciated that other combinations of metrics can be used to identify each element E. i Scoring. For example, a bivariate analysis could combine mean distance and a solidity measure; or standardized area and rectangularity measures.

[0124] The processing unit 106 can be configured to perform a multivariate analysis. The processing unit 106 can be configured to combine each of the metrics described herein, namely: normalized area A, normalized mean distance d, rectangularity, and solidity S i Each element E can be calculated based on the multivariate analysis of the four (4) calculated metrics i The score of Fig. 10A , 10B , 11A and 11B show graphs of scores calculated from the first principal component PC1 and the assigned wall motion scores (ie, normal "1", hypokinetic "2", akinetic "3"). Fig. 10A A graph of the scores calculated from the first principal component PC1 of the multivariate analysis of the resting state in the apical 2 chambers is shown in FIG. Fig. 10BA graph of the scores calculated from the first principal component PC1 of the multivariate analysis of the resting state in the apical 4 chambers is shown in FIG. Fig.11A A graph of the scores calculated from the first principal component PC1 of the multivariate analysis of stress status in the apical 2 chambers is shown in FIG. Fig. 11B Graphs showing scores calculated from the first principal component PC1 of the multivariate analysis of stress status in the apical 4 chambers are shown.

[0125] For each element E i The calculated measure and the corresponding element E i Based on this comparison, each element E is calculated. i The score represents the value of the first principal component of the new data; that is, how far the calculated measure is from the centroid of the data, and therefore how different they are from the reference data. The score can be calculated on a continuous scale. Fig. 12A and 12B , where each element is shaded according to the continuous scoring scale described here. Fig. 12A The scoring images shown in represent 2-chamber data at rest; Fig. 12B The scoring images shown represent 2-chamber data under stress conditions.

[0126] The processing unit 106 may be configured to calculate for each element E i The processing unit 106 may calculate different metrics for scoring. These metrics may be in addition to or in place of the metrics described herein for univariate and multivariate analysis. The processing unit 106 may, for example, calculate one or more of the following metrics: shear strain; coefficient of variation of distances in an element; and aspect ratio. The processing unit 106 may also calculate distance metrics other than the distances between pairs of points described herein. For example, the processing unit 106 may calculate the distance between diametrically opposite points within the same element; or the distance between corresponding points in different elements (i.e., E i and E i+j between).

[0127] The processing unit 106 has been described with particular focus on processing the elements E in an image. i It will be appreciated that the processing unit 106 may analyze multiple images. Different images may contain the same element E i The processing unit 106 may be configured to compare the images for a specific element E based on the analysis of the different images. i The scores generated. If a difference is detected between the scores, this can be flagged as a potential problem with image quality or similar. This can enable the quality of different images to be checked. Similarly, especially for fine-grained elements Ei , elements E arranged next to each other i The calculated scores are usually correlated. If an expected relationship is identified, it can be flagged as a potential image quality issue. These techniques can enable the identification of elements E that are more likely to be correct. i One or more images in .

[0128] The processing unit 106 may analyze the elements E visible in one or more images. i , to infer the invisible element E i For example, the behavior of the adjacent invisible element E i Or in the invisible element E i At least one element E arranged nearby i The score calculated is used to estimate the invisible element E i Further analysis may be performed to build a complete model of the left ventricle 202 based on the available views. The resulting model may enable the analysis of the left ventricle 202 from one or more nearby elements E. i Inference score.

[0129] The statistical analyses described in this article were performed in the R statistical computing environment (v3.4.1) using the ggplot2, dplyr, ggbeeswarm, Momocs, pathmapping, and ggpubr packages. Due to the imbalance in the number of observations between groups in each comparison, multivariate, nonparametric hypothesis tests were used to compare group means using the npmv package. A type I error rate (α) of 0.05 was used for all comparisons.

[0130] The processing unit 106 has been described herein as calculating the element E corresponding to a segment of the standard model of the left ventricle 102 i It will be appreciated that the techniques described herein do not require element E i Corresponds to a segment. For example, element E i Can be smaller than the standard model segment. Element E i can correspond to a subsection of the standard model. By reducing the element E i The score can provide a more precise indication of the location of the abnormality in the heart's cyclic motion. For example, it is envisioned that the score can indicate the location of the abnormal function within one of the segments of the standard model, such as highlighting the location near the border or in the center of the segment.

[0131] As described herein, a reference data model is generated by analyzing a reference data set comprising raw two-dimensional echocardiographic data. In the above-described embodiment, the reference data set comprises historical data comprising end-diastolic images and end-systolic images of a group of patients. In a further development, the reference data set used to generate the reference data model can be iteratively updated. For example, analysis of a new echocardiograph can be incorporated into the reference data set to increase the total amount of data available. Thus, the reference data model can continue to be refined as additional data becomes available. Iterative development of the reference data model can allow for pathological changes and patient evolution.

[0132] For each element E i The analysis described herein is performed independently. However, it will be appreciated that the analysis can be modified to take into account multiple elements E i For example, the analysis can be performed simultaneously on a first element and a second element E that are adjacent to or opposite to each other. i The motion of the left ventricle 202 (eg, a first element and a second element on opposite sides of the left ventricle 202) is scored.

[0133] The reference data model described herein may also be modified based on outcome data available for some or all of the reference data sets. The term "outcome data" is used herein to refer to diagnostic information. The outcome data is associated with a corresponding record or data set in the reference data set. The diagnostic information may, for example, be related to angiographic data and / or a cardiac event of a patient. The outcome data may indicate whether a patient has a positive or negative diagnosis of a cardiac condition (e.g., the presence or absence of coronary artery disease) within a time interval that has passed. For example, the outcome data may be generated one (1), two (2), or three (3) years after the acquisition of the echocardiographic data. The weight of the data in the reference data set may be adjusted based on the outcome data. For example, the weight applied to the data in the reference data set for which outcome data is available may be increased or decreased to change its statistical significance. The weight may be adjusted based on the time period that has passed between the acquisition of the echocardiographic image and the subsequent diagnostic event. In one variation, the reference data model may be generated based only on the data for which outcome data is available. The reference data model may be generated based only on the data for which outcome data is indicated whether a particular condition (e.g., coronary artery disease) is present. The outcome data may be used to filter the reference data set to generate a different reference data model.

[0134] The processing unit 106 may also be configured to provide a diagnostic function to generate a diagnostic output. A diagnostic system is disclosed in the applicant's international patent application PCT / GB2017 / 051720, the contents of which are incorporated herein by reference in their entirety. It is recognized that the diagnostic function may utilize the result data described herein. The diagnostic function may also rely on one or more metrics generated for the wall motion score. For example, each element E iThe rectangularity of the image may be used as a feature in a diagnostic model. Using outcome data in generating a diagnostic model may help to take into account different disease rates and characteristics over time and / or in different medical settings. For example, different settings may record different ratios of positive ("disease") to negative ("normal") results. By generating a diagnostic model using outcome data, allowance may be made for these types of variations. The results of a stress echo test, as determined by a cardiologist during / shortly after the test, may not always be accurate. An analysis conducted by applicant of one (1) year result accuracy showed an average error of 7.2% for stress echo results across multiple data sets. By referencing outcome data over a period of time, the accuracy of the diagnostic model may be improved, thereby enabling pattern-accurate predictions of whether an individual will go on to develop a disease (e.g., coronary artery disease). The use of outcome data is independently considered patentable. This enhanced diagnostic functionality is now described as an improvement over prior embodiments. The same reference numbers are used for the same components.

[0135] As described herein, each end-systolic contour point 232-n is paired with a corresponding one of the end-diastolic contour points 242-n in the end-diastolic contour data set 244. The resulting pairing of the end-systolic contour points 232-n and the end-diastolic contour points 242-n represents changes in the motion of the wall of the heart 200 during the cardiac cycle. Once the end-systolic contour points 232-n and the end-diastolic contour points 242-n have been identified, their x and y coordinates in a Cartesian coordinate system can be stored in the memory 110, for example as an end-systolic coordinate set including coordinates of each point on the end-systolic image and an end-diastolic coordinate set including coordinates of each point on the end-diastolic image. The processor can be configured to calculate the transition of the geometry of the left ventricle 202 between end-systole and end-diastole from the two coordinate sets.

[0136] The processing unit 106 is configured to calculate values ​​for various parameters used to quantify the motion of the left ventricle 202 between end-systole and end-diastole. The calculations may include calculating how far each point has moved in each of the x and y directions by calculating the change in position (end-diastole - end-systole) along the x-axis and y-axis. This gives a set of x-axis motions Δx and a set of y-axis motions Δy for each corresponding pair of end-systolic contour points 232-n and end-diastolic contour points 242-n. Each of these values ​​may be a simple distance, without an indication of direction. The average change for all points on the x-axis (ΔX) and y-axis (ΔY), respectively, may then be calculated to provide an average Δx value or x-direction motion ΔX and an average Δy value or y-direction motion ΔY for the entire left ventricle 202. If each individual motion value were purely a distance, without any indication of whether they were in the positive or negative x direction or the positive or negative y direction, these averages would describe the total amount of motion, but would not be able to indicate the direction or whether different portions of the LV wall were moving in the same or opposite directions.

[0137] Another parameter that can be calculated for each pair of end-systolic contour points 232-n and end-diastolic contour points 242-n is the average of the x-direction movement Δx and the y-direction movement Δy, where the average value for each point Δxy = (Δx + Δy) / 2. The average of all Δxy values ​​for all points can then be calculated as the ΔXY value for the entire ventricle. This calculation is similar to the calculation of shear strain and is therefore referred to herein as a shear transformation. It should be understood that for a given distance of movement, this parameter will be greatest for movement at 45 degrees relative to both the x and y axes, and smallest for movement along one of the axes.

[0138] Another parameter that can be calculated is similar to the principal transform that can be calculated from the x and y strain components, and is therefore referred to herein as the principal transform, given by

[0139] Main transformation=C1(ΔX+ΔY-√(ΔX+ΔY)^2+C2ΔXY^2).

[0140] Where C1 and C2 are constants. The constant C1 may be, for example, 1 / 2, and the constant C2 may be 4. These values ​​are used in the examples described below.

[0141] This transform is closely related to the shear transform and therefore tends to vary in a similar manner to that parameter, but with negative values ​​indicative of cardiac contraction. However, as shown in the test results below, in some cases, particularly those with coronary artery disease (CAD), the main transform value can give a more reliable diagnosis.

[0142] It is understood that each of these parameters relates to the change between end-systole and end-diastole in a single coronary cycle. However, in a stress echocardiogram (or corresponding test using other imaging methods), there is one value for each parameter of the heart at rest and one value for each parameter of the heart under stress. Comparing these values ​​(e.g., determining the difference between them) can provide more information about cardiac function that can be used for diagnosis.

[0143] Once the x-motion and y-motion and the shear transform values ​​and the main transform values ​​have been calculated, the processor is configured to compare them with reference values ​​stored in the memory 110 to diagnose one or more specific cardiac conditions and generate a diagnostic output. The output may be a simple binary output indicating a positive or negative diagnosis. The processor unit 106 may be arranged to display the output on the display 112. Alternatively, or in addition, it may be arranged to store the output as data associated with the image on which it is based, for example by adding the output data indicating the diagnosis to a file storing the image.

[0144] The reference value may be determined by means of a learning algorithm, which may, for example, be run on the processor unit 106 and use a database of stress echo images and associated diagnoses determined by conventional methods, which may be stored in the memory 110. Specifically, the database may include a large number of image groups, each group of images including end-systolic images and end-diastolic images for resting states and stress states, and associated diagnoses for each group of images, such as a positive or negative diagnosis of coronary artery disease (CAD). The learning algorithm may be arranged to analyze the images to calculate the values ​​of the various parameters mentioned above, and then determine the correlation between the diagnosis and the values ​​of each of the various parameters.

[0145] Sample images from seventy (70) subjects were analyzed. All results generated were from apical 4 chamber images. First, the values ​​of positive and negative results determined based on DSE results were compared. The comparison was then repeated, correcting the DSE results for confirmed false positives in the DSE results.

[0146] Table 1 shows the values ​​of main shift (mm), shear shift value (mm) and average ΔX (mm) of DSE results (1=positive (Pos), 2=negative (Neg)) in apical 4-chamber images at rest and under stress.

[0147] Group Statistics

[0148]

[0149] Table 2 shows the means of the main transformed values ​​(in mm), shear transformed (in mm), and X-transformed (in mm) of the adjusted DSE results (1 = positive, 2 = negative) at rest and during stress.

[0150] Group Statistics

[0151]

[0152] Table 3 shows the independent samples T-tests of the variables and the adjusted DSE.

[0153] Independent Samples T Test

[0154]

[0155] Based on the values ​​of the various parameters obtained from the sample data, machine learning can be used to determine the accuracy of each parameter as an indicator of the adjusted dobutamine stress echo (DSE) result. Using the above data, the data was classified using a J48 pruned decision tree with a 10-fold cross validation method. The following table summarizes the accuracy of each parameter as an indicator of diagnostic outcome, with the following abbreviations used: TP = true positive; FP = false positive; FN = false negative; TN = true negative; PPV = positive predictive value; and NPV = negative predictive value.

[0156] Table 4 Accuracy of Consultant Interpretation

[0157]

[0158] Table 5 Accuracy of stressor main transformation of adjusted DSE results

[0159]

[0160] Table 6 Accuracy of the resting principal transformation of the adjusted DSE results

[0161]

[0162] Table 7 Accuracy of stress shear transformation of adjusted DSE results

[0163]

[0164] Table 8 Accuracy of resting shear transformation of adjusted DSE results

[0165]

[0166] Then use machine learning to extract from all the variables Fig.13 The decision tree shown is provided to provide an accurate diagnosis based on the data. The decision tree defines a series of decision points, each of which defines a reference value or threshold for a parameter. The decision tree outlines a simple algorithm that operates as follows. First, the main transformation of the left ventricle 202 as described above is determined for the stress condition of the heart. If the transformation is less than -5.95 mm (i.e., a negative value with an amplitude greater than 5.95 mm), the diagnosis is negative. If the value is greater than -5.95 mm (i.e., a negative value with an amplitude greater than 5.95 mm), the main variation between the resting state and the stressed state is greater than 12.278053 mm, the diagnosis is negative, but if it is less than this distance, the diagnosis is positive. It should be understood that the structure of the decision tree and the reference value or threshold for each decision point in the decision tree will depend on the diagnosis to be made.

[0167] The processing unit 106 implements a fixed (static) diagnostic model for diagnosing coronary artery disease. Fig.13 As shown, the decision tree defines a series of decision points, each of which defines a reference value or threshold value for a parameter. The processing unit 106 can implement a dynamic diagnostic model. The reference value or threshold value for each decision point in the decision tree can be dynamically modified, for example to reflect new echocardiographic data and / or outcome data. In at least some embodiments, this can provide improved diagnostic functionality.

[0168] New echocardiographic data can be incrementally added to the existing reference data set. The new data is used to expand the total amount of data and may gradually change the diagnostic model. The reference values ​​or thresholds used in the decision tree can be updated to reflect the available echocardiographic data. The iterative development of the reference data allows the diagnostic model to change over time. It will be understood that the decision trees described herein can be replaced by other analytical tools, such as supervised machine learning models.

[0169] The result data includes diagnostic information for each patient, such as information related to angiographic data and / or cardiac events. The result data in this embodiment indicates whether the presence of coronary artery disease was detected within the time interval that passed after the end-systolic image and the end-diastolic image used in the reference data set were acquired. For example, the result data can be generated one (1) year, two (2) years, three (3) years or more after the echocardiographic data are acquired. The result data in this embodiment is generated one (1) year after the echocardiographic data is acquired. The result data is compiled by taking into account any angiographic data and cardiac events that occur within the time interval that passes. It will be understood that the result data continues to develop with respect to time. Therefore, the result data can be continuously updated, such as updated annually or updated when the classification changes. By updating the result data, the diagnostic tools and diagnostic models generated based on the reference data can be dynamically adjusted to represent pathological changes and patient progression.

[0170] To implement a dynamic diagnostic model, a classification model is constructed using a supervised machine learning algorithm. The result data is used to label reference data accessed by the machine learning algorithm. The machine learning algorithm uses the labels to distinguish between different classifications. In the present embodiment, the classification corresponds to the presence or absence of coronary artery disease. Alternatively or additionally, the classification can, for example, grade a particular condition based on an identified level or percentage of stenosis. It will be understood that the classification can distinguish between other states, and when generating a diagnostic model, the machine learning algorithm can adjust the relative weight of the reference data based on the labels derived from the result data. In at least some embodiments, updating the reference data based on the result data can provide improved diagnostic accuracy based on stress echocardiography.

[0171] In order to build a diagnostic model, a set of features needs to be calculated from the contour data. i) and optionally the features are calculated relative to the entire left ventricle 202. The available feature set is analyzed to identify those features that are most relevant. The most relevant features can thus be identified to establish a diagnostic model. In the case of a random forest (composed of multiple decision trees), the identified features form decision nodes. The most relevant features may vary in different geographic regions and / or as the disease progresses, and the features identified for the model may change. Even if the features remain unchanged, the thresholds and weights may change. As shown in Tables 1 and 2 of this article, the top features remain unchanged as the ejection fraction at peak stress. However, the next most important feature varies with different states. In the first dataset (Table 1), the next most relevant is the volume change between end-systole and end-diastole. However, in the combined dataset, the area of ​​a specific segment in the 2-chamber image at rest is the next most relevant. In order to train the model, the reference data needs to be labeled. Given the potential inaccuracies, using the results of the stress echo (determined by a cardiologist) as labels does not necessarily lead to an accurate model. At least some of these deficiencies may be overcome or ameliorated using outcome data collected over a period of time after the reference data was acquired (eg, one (1) year after the echocardiographic images were acquired).

[0172] The results data may be collected for different time periods. In at least some embodiments, the results data may provide an indication of how far in advance the effects of coronary artery disease may be identified. In addition, multiple categories of markers may be used to predict different disease severities. As more results data is accumulated, the diagnostic model will be updated to help ensure that the classification remains as accurate as possible due to the potential for disease progression and changes in total volume. This may be accomplished by retraining the entire model each time new results data is received. In practice, this may be time consuming. Alternatively, the processor may implement incremental machine learning techniques to continually update the diagnostic model.

[0173] The implementation of the classification model will now be described with reference to a first reference data set and a second reference data set. The first data set comprises a first set of one hundred twenty-four (124) stress echocardiograms (collected in Oxford between May 2011 and August 2013). The second data set comprises a set of three hundred ninety-nine (339) stress echocardiograms from another study (collected in six (6) different hospitals in the Thames Valley between March 2015 and August 2016). The outcome data were compiled one (1) year after the stress echocardiograms were obtained. The outcome data generates binary outcome values. In particular, the outcome is considered positive if one of the following events is found in the one (1) year interval that has passed:

[0174] (i) cardiac events (e.g. myocardial infarction);

[0175] (ii) Angiogram showing stenosis greater than 70%.

[0176] If neither of the above events (i) or (ii) occurred within the past one (1) year interval, the result was considered negative. In the first data set, ten (10) positive results were identified, and in the second data set, thirteen (13) positive results were identified.

[0177] The Boruta package in the R statistical computing environment evaluates the most relevant features for predicting an outcome. The Boruta package performs feature selection by comparing the importance of an attribute to the importance of a random attribute. A standard implementation of a random forest with 500 trees was implemented. Table 9 details the most important features for the first dataset and their average importance scores. The second dataset has been added to the first dataset. Table 10 details the most important features for the combined first and second datasets and their average importance scores. As more data becomes available for processing, the most relevant features also change. This suggests that classification models may change as more reference data is added. It is believed that these changes will be more pronounced if additional reference data is acquired at a later date and / or over a wider geographic area. Although the use of a random forest model has been described here, it should be understood that another model may be used, or indeed multiple models may be used.

[0178] Table 9 The most relevant features using the first dataset

[0179]

[0180] Table 10 Most relevant features of data using a combination of the first and second datasets

[0181]

[0182]

[0183] The implementation of a continuous learning strategy that is able to incorporate new reference data can provide more powerful and accurate diagnostic models that can be achieved. By gradually incorporating new reference data, the need to retrain the entire model every time new data becomes available can be reduced or avoided (which can prove to be a time-consuming process, especially as the size of the reference dataset increases). In addition, diagnostic models can adapt to disease characteristics that change over time. This is especially important because the most relevant biomarkers may change over time as the environment and population lifestyle change, so the model needs to adapt to these factors. Dynamic diagnostic models can adapt to the changing aspects and characteristics of cardiovascular disease, providing powerful and accurate prediction models.

[0184] The dynamic diagnostic model described herein utilizes outcome data acquired over a one (1) year period. It will be appreciated that outcome data may be accumulated over different time periods. By combining outcome data over longer time periods, the predictive power of the dynamic diagnostic model over longer time periods may be assessed.

[0185] The present application has been described with reference to cardiovascular disease. However, it will be appreciated that the methods and devices described herein may have other applications. For example, if the environment changes and the tumor size, appearance, or calcification changes, diagnostic tools may be developed to adapt to the changing tumor imaging biomarkers. Furthermore, the technology may be applicable to imaging systems other than echocardiographs.

[0186] It will be understood that various modifications may be made to the embodiments described herein without departing from the scope of the appended claims.

Claims

1. A system for generating a diagnostic model, the system comprising a processor, the processor being configured to: analyzing a plurality of reference data sets, each reference data set comprising a first image and a second image, the analyzing comprising identifying at least one first feature in the first image and identifying at least one second feature in the second image, each at least one first feature being paired with a corresponding one of the at least one second features; comparing each pair of corresponding first and second features to identify one or more differences therebetween, the one or more differences being indicative of cardiac cycle variations; calculating at least one metric based on the one or more differences identified between each pair of corresponding first and second features, wherein the at least one metric is a normalized area, a normalized mean distance, a rectangularity, and a solidity of elements representing cardiac cycle variations; obtaining result data associated with at least one corresponding reference data set; and compiling a diagnostic model based on the at least one calculated metric and the result data; wherein the first image comprises a first end-systolic image, and the second image comprises a second end-diastolic image, and wherein the result data includes cardiac diagnosis information and is associated with a corresponding data set in the reference data set; and The diagnostic model is updated when new result data associated with an existing corresponding data set in the reference data set becomes available, or when result data associated with an existing corresponding data set in the reference data set is updated.

2. The system according to claim 1, wherein: The processor is configured to label the image with a tag based on the result data.

3. The system according to claim 2, wherein: Tags distinguish different categories.

4. The system according to claim 1, wherein: Diagnostic information included documentation of cardiac events including myocardial infarction.

5. The system according to claim 1, wherein: The resulting data includes records of stenosis greater than a threshold value.

6. The system according to any one of claims 1 to 5, wherein: The processor is configured to determine a weight for at least one calculated metric, each weight being determined based on result data associated with a given one of the reference data sets.

7. The system according to claim 1, wherein: A diagnostic model is compiled based on the calculated metrics and associated weights.

8. The system according to claim 1, wherein: The processor is configured to implement a machine learning algorithm to generate a diagnostic model.

9. The system according to claim 1, wherein: The processor is configured to analyze the additional reference data set; and update the diagnostic model based on the analysis of the additional reference data set.

10. A method for generating a diagnostic model, the method comprising: analyzing a plurality of reference data sets, each reference data set comprising a first image and a second image, the analyzing comprising identifying at least one first feature in the first image and identifying at least one second feature in the second image, each at least one first feature being paired with a corresponding one of the at least one second features; comparing each pair of corresponding first and second features to identify one or more differences therebetween, the one or more differences being indicative of cardiac cycle variations; calculating at least one metric based on the one or more differences identified between each pair of corresponding first and second features, wherein the at least one metric is a normalized area, a normalized mean distance, a rectangularity, and a solidity of elements representing cardiac cycle variations; obtaining result data associated with at least one corresponding reference data set; and compiling diagnostic models based on the calculated metrics and outcome data; wherein the first image comprises a first end-systolic image, and the second image comprises a second end-diastolic image, and wherein the result data includes cardiac diagnosis information and is associated with a corresponding data set in the reference data set; and The diagnostic model is updated when new result data associated with an existing corresponding data set in the reference data set becomes available, or when result data associated with an existing corresponding data set in the reference data set is updated.

11. The method according to claim 10, wherein: The resulting data are used to label the corresponding reference datasets with labels.

12. The method according to claim 11, wherein: Tags distinguish different categories.

13. The method according to claim 10, wherein: Diagnostic information included documentation of cardiac events including myocardial infarction.

14. The method according to claim 10, wherein: The resulting data includes records of stenosis greater than a threshold value.

15. The method according to any one of claims 10 to 14, comprising: A weight is determined for at least one calculated metric, each weight being determined based on result data associated with one of the given reference data sets.

16. The method according to claim 15, wherein: A diagnostic model is compiled based on the calculated metrics and associated weights.

17. The method according to claim 10, wherein: Implement machine learning algorithms to generate diagnostic models.

18. The method according to claim 10, comprising: Additional reference data sets are incrementally added, the method comprising: analyzing the additional reference data sets; and updating the diagnostic model based on the analysis of the additional reference data sets.

19. A non-transitory computer readable medium having a set of instructions stored therein, which when executed causes a processor to perform the method of any one of claims 10 to 18.

Images