System and method for segmenting images

Abstract

The invention provides methods and systems for identifying shape-based atypical segmentations. In one example, a method includes receiving a segmentation of a region of interest (ROI) of a medical image, the segmentation output by a segmentation model; computing a confidence measure for the segmentation, the confidence measure indicating how well a shape of the segmentation can be encoded by one or more dominant shape variation patterns of a set of predetermined segmentations of the ROI; and in response to the confidence measure satisfying a predetermined condition with respect to a threshold, displaying the segmentation, storing the segmentation, and / or using the segmentation for one or more downstream processes; otherwise, prompting a user to perform a manual segmentation.

Description

Technical Field

[0001] The embodiments of the subject matter disclosed herein relate to medical imaging, and more specifically, to the segmentation of medical images. Background Technology

[0002] Medical imaging, such as ultrasound, can be used non-invasively to probe the internal structures of a patient's body and generate corresponding images. These medical images of internal structures can be saved for later analysis by clinicians, aiding in diagnosis and / or displayed on a display device in real-time or near real-time. In some examples, computerized tools can be used to identify internal structures, provide suggested diagnoses, perform automated measurements, etc. Summary of the Invention

[0003] In one implementation, a method includes: receiving a segmentation of a region of interest (ROI) of a medical image, the segmentation being output by a segmentation model; calculating a confidence metric for the segmentation, the confidence metric indicating how well the shape of the segmentation can be encoded by encoding one or more major shape variation patterns of a pre-determined set of segmentations of the ROI; and, in response to the confidence metric satisfying a predetermined condition relative to a threshold, displaying the segmentation, storing the segmentation, and / or using the segmentation for one or more downstream processes; otherwise, prompting a user to perform manual segmentation.

[0004] The above-described advantages, as well as other advantages and features, of this specification will become apparent, either alone or in connection with the accompanying drawings, from the following detailed description. It should be understood that the above summary is provided to present a simplified version of the selected concepts further described in the detailed description. This is not intended to identify key or essential features of the claimed subject matter, the scope of which is uniquely defined by the claims following the detailed description. Furthermore, the claimed subject matter is not limited to embodiments that address any of the disadvantages mentioned above or in any part of this disclosure. Attached Figure Description

[0005] A better understanding of the various aspects of this disclosure can be achieved by reading the following detailed description and referring to the accompanying drawings, in which:

[0006] Figure 1 A block diagram illustrating an implementation scheme for an ultrasound system is shown;

[0007] Figure 2 This is a block diagram illustrating an exemplary image processing system;

[0008] Figure 3 An exemplary segmentation of the region of interest (ROI) is schematically illustrated;

[0009] Figure 4This is a flowchart illustrating a method for training a segmentation model to generate a segmentation confidence metric using principal component analysis (PCA);

[0010] Figure 5 An example of Procrustes registration performed to reduce non-shape variations in segmentation is shown;

[0011] Figure 6 An exemplary PCA variation pattern is shown;

[0012] Figure 7 This is a flowchart illustrating a method for implementing a segmentation model; and

[0013] Figure 8 The histogram of confidence metrics calculated for multiple segments is shown. Detailed Implementation

[0014] Medical images, such as ultrasound images, can be used to diagnose or rule out patient conditions in a non-invasive manner. To facilitate the analysis of patient conditions, computerized tools can be applied to medical images to provide: automated or semi-automated measurements of anatomical features, identification or labeling of tissue characteristics, or even suggestive diagnoses of the patient's condition. As an example, during echocardiography, which uses ultrasound to image a patient's heart, automated functional imaging (AFI) can be applied to perform 2D speckle tracking to measure the deformation (strain) of the myocardial wall. However, these strain measurements may lack reproducibility due to variations in the initialization of regions of interest (ROIs) used to track cardiac tissue. Therefore, to improve reproducibility, deep learning segmentation networks (also known as segmentation models) can be used to automatically initialize the ROIs.

[0015] However, while deep learning models are generally highly accurate in generating ROIs, they are prone to generating outlier outputs when the segmented shape does not match the expected shape of the relevant anatomical structure. When the segmentation model outputs an atypical ROI shape, strain measurements may be inaccurate, requiring remeasurement or manual measurement. This remeasurement or manual measurement delays diagnosis and increases the cognitive load on clinicians.

[0016] Therefore, according to the embodiments disclosed herein, atypical segmentation outputs can be identified by determining a confidence metric for each segmentation output, which indicates the gap between the segmentation shape and the expected shape of the anatomical ROI. A large gap indicates that the segmentation shape is outside the expected shape range of the anatomical ROI, and thus the user can be notified of the atypical segmentation, allowing the user to perform manual segmentation or input different images for segmentation. The confidence metric is calculated by comparing the segmentation with the mean shape of the anatomical ROI identified from multiple confirmed segmentations of the anatomical ROI (e.g., a labeled dataset generated by one or more experts). Procrustes analysis can be used to register the segmentation output by the model with the mean shape to reduce any non-shape variations in the segmentation, such as size and rotation. The registered segmentation can be transformed into a low-dimensional shape and then reconstructed using one or more variation patterns identified by principal component analysis (PCA) of the labeled dataset. The gap between the shape of the reconstructed segment and the shape of the model-output segment can be determined, and segments with a high degree of gap can be flagged without further processing. In this way, a confidence metric for the segment can be calculated, and this confidence metric can be used to determine whether the output segment should be used for further processing. The confidence metric indicates how well the shape of the segment can be encoded by encoding one or more major shape variation patterns of a pre-determined set of segments for the ROI.

[0017] Figure 1 An exemplary ultrasound system is illustrated, comprising an ultrasound probe, a display device, and an image processing system. Ultrasound data can be acquired via the ultrasound probe, and ultrasound images can be displayed on the display device. The ultrasound images can be processed by an image processing system, such as... Figure 2 The image processing system is used to segment the anatomical ROI and calculate the confidence metric of the segmentation. Figure 3 Exemplary segmentations of anatomical ROIs are shown, including model-generated segmentation and expert-generated segmentation. Figure 4 An exemplary method for training a segmentation model is shown, which is configured to generate a confidence metric for the segments generated by the segmentation model. According to... Figure 7 The confidence metric, as shown, can be generated as follows: by reducing non-shape variations in the segmentation, such as by applying... Figure 5 The Procrustes registration is shown; and by applying one or more major PCA variation patterns of multiple expert-generated segmentations, such as Figure 6 The PCA pattern shown is used to reconstruct the segment. Figure 8 A histogram of the segment confidence metric is shown, illustrating the difference between high-confidence and low-confidence segmentations.

[0018] The generated segmentation model and associated confidence metrics can be applied to medical images to identify anatomical ROIs for display and / or further processing. Figure 1 An exemplary ultrasound imaging system is shown that can be used to generate medical images that can be input into a segmentation model as disclosed herein. However, it should be understood that the ultrasound imaging system is presented herein as an exemplary medical imaging system, and other medical images, such as computed tomography (CT) images, magnetic resonance (MR) images, X-ray images, and visible light images, can be used to implement the segmentation model and the generated confidence metric without departing from the scope of this disclosure.

[0019] See Figure 1 A schematic diagram of an ultrasound imaging system 100 according to an embodiment of the present disclosure is shown. The ultrasound imaging system 100 includes a transmit beamformer 101 and a transmitter 102 that drives elements (e.g., transducer elements) 104 within a transducer array (referred to herein as probe 106) to transmit pulsed ultrasound signals (referred herein as transmit pulses) into a body (not shown). According to one embodiment, probe 106 may be a one-dimensional transducer array probe. However, in some embodiments, probe 106 may be a two-dimensional matrix transducer array probe. As further explained below, transducer element 104 may be made of a piezoelectric material. When a voltage is applied to a piezoelectric crystal, the piezoelectric crystal physically expands and contracts, thereby emitting ultrasound waves. In this way, transducer element 104 can convert an electronic emission signal into an acoustic emission beam.

[0020] After element 104 of probe 106 transmits a pulsed ultrasound signal into the patient's body, the pulsed ultrasound signal is reflected from internal structures (such as blood cells or muscle tissue) to produce an echo returning to element 104. The echo is converted into an electrical signal or ultrasound data by element 104, and the electrical signal is received by receiver 108. The electrical signal representing the received echo passes through receiver beamformer 110, which outputs ultrasound data.

[0021] The echo signals generated by the transmission operation are reflected from structures located at continuous distances along the emitted ultrasonic beam. The echo signals are sensed individually by each transducer element, and samples of the echo signal amplitude at specific time points represent the amount of reflection occurring at that specific distance. However, due to the difference in the propagation path between the reflection point P and each element, these echo signals are not detected simultaneously. Receiver 108 amplifies the individual echo signals, assigns a calculated reception time delay to each echo signal, and sums them to provide a single echo signal that approximately indicates the total ultrasonic energy reflected from point P located at a distance R along an ultrasonic beam oriented at angle θ.

[0022] During the reception of the echo, the time delay of each receiving channel is continuously varied to provide dynamic focusing of the received beam at a distance R, based on the assumed sound speed of the medium assuming that the echo signal is emitted from a distance R.

[0023] As instructed by processor 116, receiver 108 provides a time delay during scanning, such that the orientation of receiver 108 tracks the direction θ of the beam oriented by the transmitter, and samples the echo signal at consecutive distances R to provide time delay and phase shift for dynamic focusing along the beam at point P. Therefore, each transmission of the ultrasonic pulse waveform results in the acquisition of a series of data points representing the amount of sound reflected from a series of corresponding points P located along the ultrasonic beam.

[0024] According to some embodiments, probe 106 may include electronic circuitry to perform all or part of transmit beamforming and / or receive beamforming. For example, all or part of transmit beamformer 101, transmitter 102, receiver 108, and receive beamformer 110 may be located within probe 106. In this disclosure, the terms “scanning” or “under scanning” may also be used to refer to the process of acquiring data by transmitting and receiving ultrasound signals. In this disclosure, the term “data” may be used to refer to one or more datasets acquired using an ultrasound imaging system. User interface 115 may be used to control the operation of ultrasound imaging system 100, including for controlling the input of patient data (e.g., patient history), for changing scan or display parameters, for initiating probe repolarization sequences, etc. User interface 115 may include one or more of the following: a rotary element, a mouse, a keyboard, a trackball, hard keys linked to specific actions, soft keys configurable to control different functions, and a graphical user interface displayed on display device 118.

[0025] The ultrasound imaging system 100 also includes a processor 116 for controlling the transmitting beamformer 101, the transmitter 102, the receiver 108, and the receiving beamformer 110. The processor 116 communicates electronically (e.g., is communicatively connected) with the probe 106. For the purposes of this disclosure, the term "electronic communication" may be defined to include both wired and wireless communication. The processor 116 can control the probe 106 to acquire data according to instructions stored in the processor's memory and / or memory 120. The processor 116 controls which of the elements 104 are active and the shape of the beam emitted from the probe 106. The processor 116 also communicates electronically with a display device 118 and can process data (e.g., ultrasound data) into images for display on the display device 118. The processor 116 may include a central processing unit (CPU) according to one embodiment. According to other embodiments, the processor 116 may include other electronic components capable of performing processing functions, such as a digital signal processor, a field-programmable gate array (FPGA), or a graphics board. According to other embodiments, processor 116 may include multiple electronic components capable of performing processing functions. For example, processor 116 may include two or more electronic components selected from a list of electronic components, including: a central processing unit, a digital signal processor, a field-programmable gate array, and a graphics board. According to another embodiment, processor 116 may also include a composite demodulator (not shown) that demodulates real RF (radio frequency) data and generates composite data. In another embodiment, demodulation may be performed earlier in the processing chain. Processor 116 is adapted to perform one or more processing operations based on multiple selectable ultrasound modalities on the data. In one example, data may be processed in real time during a scanning session because echo signals are received by receiver 108 and transmitted to processor 116. For the purposes of this disclosure, the term "real time" is defined as including processes performed without any intentional delay. For example, embodiments may acquire images at a real-time rate of 7 frames / second to 20 frames / second. Ultrasonic imaging system 100 is capable of acquiring 2D data of one or more planes at significantly faster rates. However, it should be understood that the real-time frame rate may depend on the length of time spent acquiring each frame of data used for display. Therefore, when acquiring relatively large amounts of data, the real-time frame rate may be slow. Consequently, some embodiments may have a real-time frame rate significantly faster than 20 frames per second, while others may have a real-time frame rate lower than 7 frames per second. Data may be temporarily stored in a buffer (not shown) during a scanning session and processed in a less real-time manner during real-time or offline operation. Some embodiments of the invention may include multiple processors (not shown) to handle the processing tasks handled by processor 116 according to the exemplary embodiments described above.For example, before displaying an image, a first processor can be used to demodulate and extract the RF signal, while a second processor can be used to further process the data (e.g., by augmenting the data as further described herein). It should be understood that other embodiments may use different processor arrangements.

[0026] The ultrasound imaging system 100 can continuously acquire data at frame rates, for example, from 10 Hz to 30 Hz (e.g., 10 to 30 frames per second). Images generated from the data can be refreshed on a display device 118 at a similar frame rate. Other embodiments are capable of acquiring and displaying data at different rates. For example, depending on the frame size and the intended application, some embodiments may acquire data at frame rates less than 10 Hz or greater than 30 Hz. A memory 120 is included for storing frames of processed acquired data. In an exemplary embodiment, the memory 120 has sufficient capacity to store at least several seconds of ultrasound data frames. The data frames are stored in a manner that facilitates retrieval based on their acquisition order or time. The memory 120 may include any known data storage medium.

[0027] In various embodiments of the invention, processor 116 can process data through different mode-related modules (e.g., B-mode, color Doppler, M-mode, color M-mode, spectral Doppler, elastography, TVI, strain, strain rate, etc.) to form 2D or 3D data. For example, one or more modules can generate B-mode, color Doppler, M-mode, color M-mode, spectral Doppler, elastography, TVI, strain, strain rate, and combinations thereof. As an example, one or more modules can process color Doppler data, which may include conventional color flow Doppler, power Doppler, HD flow, etc. Image lines and / or frames are stored in memory and may include timing information indicating the time when image lines and / or frames are stored in memory. These modules may include, for example, a scan transformation module to perform a scan transformation operation to convert the acquired images from beam space coordinates to display space coordinates. A video processor module may be provided that reads the acquired images from memory and displays the images in real time while performing procedures on a patient (e.g., ultrasound imaging). The video processor module may include a separate image memory, and ultrasound images may be written to the image memory for reading and display by the display device 118.

[0028] In various embodiments of this disclosure, one or more components of the ultrasound imaging system 100 may be included in a portable handheld ultrasound imaging device. For example, a display device 118 and a user interface 115 may be integrated into the external surface of the handheld ultrasound imaging device, which may also include a processor 116 and a memory 120. A probe 106 may include a handheld probe that electronically communicates with the handheld ultrasound imaging device to collect raw ultrasound data. Transmit beamformer 101, transmitter 102, receiver 108, and receive beamformer 110 may be included in the same or different parts of the ultrasound imaging system 100. For example, transmit beamformer 101, transmitter 102, receiver 108, and receive beamformer 110 may be included in the handheld ultrasound imaging device, the probe, and combinations thereof.

[0029] After performing a two-dimensional ultrasound scan, a data block containing scan lines and their samples is generated. Following the application of a back-end filter, a process called scan transformation is performed to convert the two-dimensional data block into a displayable bitmap image with additional scan information, such as depth, the angle of each scan line, etc. During scan transformation, interpolation techniques are applied to fill in any missing holes (i.e., pixels) in the resulting image. These missing pixels occur because each element of the two-dimensional block should typically cover many pixels in the resulting image. For example, in current ultrasound imaging systems, bicubic interpolation is used, which utilizes the adjacent elements of the two-dimensional block. Therefore, if the two-dimensional block is relatively small compared to the size of the bitmap image, the scan-transformed image will include areas with lower or no resolution than optimal, especially for deeper regions.

[0030] refer to Figure 2 An image processing system 202 according to an exemplary embodiment is illustrated. In some embodiments, the image processing system 202 is integrated into an ultrasound imaging system 100. For example, the image processing system 202 may be provided in the ultrasound imaging system 100 as a processor 116 and a memory 120. In some embodiments, at least a portion of the image processing system 202 is included in a device (e.g., an edge device, server, etc.) communicatively coupled to the ultrasound imaging system via a wired and / or wireless connection. In some embodiments, at least a portion of the image processing system 202 is included at a separate device (e.g., a workstation) that can receive images from the ultrasound imaging system or from a storage device storing images / data generated by the ultrasound imaging system. The image processing system 202 may be operatively / communically coupled to a user input device 232 and a display device 234. In one example, the user input device 232 may include a user interface 115 of the ultrasound imaging system 100, while the display device 234 may include a display device 118 of the ultrasound imaging system 100.

[0031] Image processing system 202 includes processor 204 configured to execute machine-readable instructions stored in non-transitory memory 206. Processor 204 may be a single-core or multi-core processor, and the program executing on it may be configured for parallel or distributed processing. In some embodiments, processor 204 may optionally include individual components distributed across two or more devices, which may be located remotely and / or configured for collaborative processing. In some embodiments, one or more aspects of processor 204 may be virtualized and executed by remotely accessible, networked computing devices configured in a cloud computing configuration.

[0032] Non-transitory memory 206 may store segmentation model 208, confidence module 209, ultrasound image data 210, and training module 212. Segmentation model 208 may include one or more machine learning models (such as deep learning networks) including multiple weights and biases, activation functions, loss functions, gradient descent algorithms, and instructions for implementing one or more deep neural networks to process input ultrasound images for segmenting regions of interest. Segmentation model 208 may include trained and / or untrained neural networks and may also include training routines or parameters (e.g., weights and biases) associated with one or more neural network models stored therein. As will be explained herein, confidence module 209 may be configured to generate a confidence metric for each segment output by segmentation model 208.

[0033] Ultrasound image data 210 may include Figure 1 Ultrasound images captured by ultrasound imaging system 100 or another ultrasound imaging system. Ultrasound image data 210 may include 2D images and / or 3D volumetric data from which 2D images / slices can be generated. Ultrasound image data 210 may include B-mode images, Doppler images, color Doppler images, M-mode images, etc., and / or combinations thereof. In some embodiments, ultrasound image data 210 may store ultrasound images and ground reality outputs in an ordered format, such that each ultrasound image is associated with one or more corresponding ground reality outputs. However, in an example where training module 212 is not located at image processing system 202, the images / ground reality outputs used to train segmentation model 208 may be stored elsewhere.

[0034] The non-transitory memory 206 may also include a training module 212, which includes instructions for training one or more machine learning models stored in the segmentation model 208. In some embodiments, the training module 212 is not located at the image processing system 202. Therefore, the segmentation model 208 includes a trained and validated network. Furthermore, in some examples, the segmentation model 208 may not be a deep learning model. Instead, the segmentation model 208 may be a computer vision-based model that uses computer vision techniques such as edge detection to generate segmentation.

[0035] In some embodiments, nontransitory memory 206 may include components included in two or more devices that can be remotely located and / or configured for coordinated processing. In some embodiments, one or more aspects of nontransitory memory 206 may include remotely accessible networked storage devices configured in a cloud computing configuration.

[0036] User input device 232 may include one or more of the following: a touchscreen, keyboard, mouse, touchpad, motion-sensing camera, or other devices configured to enable a user to interact with and manipulate data within image processing system 202. In one example, user input device 232 may enable a user to select ultrasound images for training a machine learning model, to indicate or mark the boundaries of anatomical ROIs in ultrasound image data 210, or to perform further processing using the trained machine learning model.

[0037] Display device 234 may include one or more display devices utilizing virtually any type of technology. In some embodiments, display device 234 may include a computer monitor and be capable of displaying ultrasound images. Display device 234 may be combined with processor 204, nontransitory memory 206, and / or user input device 232 in a shared housing, or it may be a peripheral display device and may include a monitor, touchscreen, projector, or other display device known in the art, which enables a user to view ultrasound images generated by the ultrasound imaging system and / or interact with various data stored in nontransitory memory 206.

[0038] It should be understood that Figure 2 The image processing system 202 shown is for illustrative purposes and not for limitation. Another suitable image processing system may include more, fewer, or different components.

[0039] Therefore, the image processing system 202 can be configured to acquire medical images, such as ultrasound images, including anatomical ROIs, such as the left ventricle of the heart, and input the medical images into the segmentation model 208. The segmentation model 208 can generate segments of the anatomical ROI from the input medical images. The confidence metric of the segmentation can be calculated by the confidence module 209 as follows: based on the degree of matching between the shape of the segment and the mean shape of previous segments confirmed by experts for the anatomical ROI, while considering the main variation patterns between the previous segments confirmed by experts. The previous segments confirmed by experts can be ground-based segments used to train the segmentation model 208 to generate segments, such as ground-based segments stored with the ultrasound image data 210, or segments confirmed by individual experts. To determine the mean shape and main variation patterns, Procrustes analysis can be performed using the segment confirmed by experts to find the mean shape and register the segment confirmed by experts with the mean shape to reduce non-shape variations in the segment confirmed by experts, such as variations due to different sizes of anatomical ROIs and different image orientations. Then, PCA can be performed on the registered shape to find the main shape variation patterns between segments confirmed by experts.

[0040] Figure 3 A set of images 300 with exemplary segmentation is shown, including a first image 302 of the left ventricle of a first imaging subject and a second image 310 of the left ventricle of a second imaging subject. Both the first image 302 and the second image 310 can be input to a trained segmentation model that outputs a segmentation of the left ventricle. As shown on the first image 302, the segmentation output by the segmentation model is indicated by line 304 (dashed dotted line). The segmentation output by the model may have a different shape than a segmentation performed by an expert, as shown by line 306 (dashed line). Conversely, the segmentation output by the model for the second image 310 (shown by line 312) has a shape similar to an expert-generated segmentation (shown by line 314).

[0041] The embodiments disclosed herein allow for the automatic detection of atypical shapes, such as the segmentation output by the model for the first image 302, without requiring experts to generate a confirmed or known segmentation shape for each image. As will be explained in more detail below, the model output segmentation can be registered with the mean shape of the left ventricle, and reconstruction can be performed in a low-dimensional space using the main variation patterns used for left ventricular segmentation. The reconstructed shape of the segmentation for the first image 302 is shown by lines 308 (solid lines) and 316 in the second image 310. A confidence metric can be calculated by determining the difference between the shape of the model output segmentation (e.g., line 304) and the reconstructed shape (e.g., line 308). Compared to the difference between the shape of the model output segmentation and the reconstructed shape of the second image 310 (e.g., 0.070), the difference between the shape of the model output segmentation and the reconstructed shape of the first image 302 is relatively high (e.g., 0.235, which may be higher than the threshold indicating atypical shapes), thus confirming that the segmentation output by the model for the first image 302 does not conform to the shape distribution previously seen by the expert annotator, and can therefore be classified as atypical and marked as possibly an incorrect segmentation.

[0042] Now go to Figure 4 The flowchart illustrates an exemplary method 400 for training a segmentation model and constructing a confidence module configured to generate a confidence metric for the segmentation output by the segmentation model. (See also...) Figures 1 to 2 Method 400 is described using systems and components, but it should be understood that method 400 can be implemented with other systems and components without departing from the scope of this disclosure. Method 400 may be implemented using non-transitory memory stored in a computing device (such as... Figure 1 memory 120 or Figure 2 The instructions in the memory 206 are executed, and the processor of the computing device (such as...) is... Figure 1 Processor 116 or Figure 2 The processor 204) is used to execute.

[0043] At 402, method 400 includes: obtaining a training image that includes the anatomical ROI. For example, the training image could be obtained using... Figure 1The training images are ultrasound images obtained using an ultrasound imaging system, and the anatomical ROI can be the left ventricle of the heart. However, the training images can include images obtained using different imaging modalities (e.g., CT images, MR images, etc.) and / or include different anatomical ROIs (e.g., different features of the heart, or different anatomical structures such as the liver, kidney, etc.). However, the training images can all be acquired using the same imaging modality and can all include the same anatomical ROI, such that the segmentation model is trained to segment a specific anatomical ROI in a specific medical image. At 404, method 400 includes: obtaining ground truth segmentation of the anatomical ROI in the training images. The ground truth segmentation can be expert-generated labels / annotations. For example, one or more experts (e.g., clinicians) can evaluate the training images and generate ground truth by manually (e.g., via a user input device) labeling / annotating the training images using ground truth segmentation indicating, for example, the boundaries of the anatomical ROI.

[0044] At position 406, a segmentation model is trained using the training image and ground reality segmentation. Before training, the segmentation model can be... Figure 2 The segmentation model 208. Training the segmentation model may include: taking each training image as input to the segmentation model; and updating one or more weights, biases, gradients, etc. of the segmentation model based on one or more missing values ​​between the segmentation model output and the associated ground truth segmentation. In some examples, the segmentation model can be trained to output a classification for each pixel in the input image based on the characteristics of the input image, where the classification indicates whether the pixel belongs to an anatomical ROI or whether the pixel belongs to the background (or another suitable classification). This allows the segmentation model to be trained in a way that does not understand the shape of the anatomical ROI, which may result in the segmentation model occasionally outputting atypical shapes.

[0045] Therefore, a confidence module can be generated to evaluate the shape of each segment output by the segmentation pattern. To generate the confidence module, at 408, method 400 includes: determining one or more variation patterns of the ground reality segmentation. Determining the variation patterns may include: eliminating non-shape variations in these segments, as shown at 410. Non-shape variations are eliminated by determining the mean shape of the segments and registering each segment to the mean shape using Procrustes analysis. Procrustes analysis allows for a rotation / size-invariant version of shape analysis by calculating the mean shape of the entire dataset and registering each new shape to minimize rotation and size differences between the new shape and the mean shape. Procrustes analysis works by normalizing two arrays (current mean A and sample B) to have a mean of 0 and a norm of 1.

[0046] Find the optimal rotation matrix (R) and scale (s) U, W, V using singular value decomposition (SVD). T=svd((B T A) T ); R = VU T ;s=∑W.

[0047] Then use B=BR T *s is used to normalize B (sample). Note that A and B are m×2 arrays, where m is the number of points (12 in the example figure) or [[x0,y0],[x1,y1],...[xm,ym]]. Although the segmentation shown herein includes 12 points, the segmentation may include more or fewer points without departing from the scope of this disclosure.

[0048] To generate the mean shape for ground truth segmentation, the mean shape is first initialized. The initial mean shape can be one of randomly selected ground truth segmentations, or it can be generated by determining the average position of each point segmented across the entire set of ground truth segmentations. Each shape of the ground truth segmentation (e.g., the labeled training dataset) is Procrustes registered with the initial mean shape. The registered shapes are then used to recalculate the mean shape. If the mean shape has changed relative to the initial mean shape, the ground truth segmentation shape is registered with the new mean shape, and a further updated mean shape is calculated. This process can be repeated until the mean shape stops changing. Once the mean shape stops changing, it can be saved as part of a confidence module (such as confidence module 209) or as part of the segmentation model.

[0049] Determining variation patterns may also include performing principal component analysis (PCA) to identify one or more major shape variation patterns in the ground reality segmentation, as shown in 412. PCA can be performed to identify a relatively small (e.g., 2-3) number of dimensions that can be used to encode the major shape variations seen in the ground reality segmentation. As will be explained in more detail below, once the segmentation model has been trained and deployed as the output segmentation, PCA patterns can be applied to reconstruct the shape / segmentation. After the segmentation has been registered with the mean shape determined in 412, PCA can be performed on the ground reality segmentation. After identifying the major variation patterns, these variation patterns (which may be referred to as PCA patterns) can be saved as part of the confidence module, as shown in 414. However, in other examples, these PCA patterns can be saved as part of the trained segmentation model. Furthermore, although... Figure 4 This paper describes the determination of PCA patterns using ground-based segmentation that was also used to train the segmentation model. However, in some examples, different sets of expert-confirmed segmentations of the anatomical ROI can be used to generate PCA patterns. In such examples, PCA patterns can be generated independently of the training of the segmentation model.

[0050] Figure 5 An exemplary plot 500 is shown, showing Procrustes registration with the mean shape of an anatomical ROI such as the left ventricle. Plot 500 includes a horizontal x-axis with values ​​increasing along the arrow direction and a vertical y-axis with values ​​increasing along the arrow direction. Plot 500 shows the original segmentation 502, for example, output by a segmentation model or generated by an expert, plotted on the x and y axes. Plot 500 also includes a mean shape 504, generated via the aforementioned Procrustes analysis (e.g., based on previous segmentations confirmed by multiple experts for the anatomical ROI), plotted on the x and y axes. As shown, the mean shape 504 is plotted such that its center point is located at point 0,0 on the plot. Plot 500 also includes a registration shape 506 plotted on the x and y axes, which includes the original segmentation 502 registered with the mean shape 504 via the Procrustes analysis / registration described above. As understood from graph 500, registration with the mean shape results in adjustments to the size and rotation of the original segment, but preserves the shape of the original segment. The mean shape 504 and the original segment 502 are each shown as including the shape plotted on... Figure 5 The 12 points on the coordinate system, as shown in the reference above. Figure 4 However, without departing from the scope of this disclosure, more or fewer points may be used. Thus, the mean shape 504 can be a set of points plotted in any coordinate system. The mean shape 504 is the mean shape of multiple segments of the anatomical ROI in a given imaging plane, not directly determined by the anatomical atlas, and does not necessarily represent the actual 3D shape of the anatomical feature.

[0051] Figure 6An exemplary set 600 of PCA patterns is shown, which are identified by performing PCA on previously segmented anatomical ROIs such as the left ventricle, confirmed by multiple experts. For example, set 600 can be generated by PCA performed as part of method 400. Set 600 includes a first pattern 610, referred to as pattern 0. The first pattern 610 may be an orientation. Line 612 shows a mean shape, while other lines (e.g., line 614) show variations in previously segmented ROIs confirmed by experts (e.g., ground truth segmentation) relative to the first pattern (e.g., orientation). Set 600 includes a second pattern 620, referred to as pattern 1. The second pattern 620 may be sphericity. Line 622 shows a mean shape, while other lines (e.g., line 624) show variations in previously segmented ROIs confirmed by experts (e.g., ground truth segmentation) relative to the second pattern (e.g., sphericity). Set 600 also includes a third pattern 630, referred to as pattern 2. The third pattern 630 may be a base point location (e.g., the endpoint location of a shape). Line 632 shows the mean shape, while other lines (e.g., line 634) show the variation of a previous segmentation (e.g., ground reality segmentation) confirmed by experts relative to a third pattern (e.g., relative to the endpoints of the shape).

[0052] Figure 7 A flowchart is shown illustrating an exemplary method 700 for executing (e.g., deploying) a trained segmentation model during the inference phase and generating a confidence metric for the segmentation output by the segmentation model. (Refer to...) Figures 1 to 2 Method 700 is described using systems and components, but it should be understood that other systems and components may be used to implement method 700 without departing from the scope of this disclosure. Method 700 may be implemented using non-transitory memory stored in a computing device (such as...). Figure 1 memory 120 or Figure 2 The instructions in the memory 206 are executed, and the processor of the computing device (such as...) is... Figure 1 Processor 116 or Figure 2 The processor 204) is used to execute.

[0053] At point 702, a medical image of the anatomical ROI is obtained. The medical image may be an ultrasound image, and the anatomical ROI may be the left ventricle of the heart, but other imaging modalities (e.g., CT, MRI) and / or other anatomical ROIs (e.g., the brain, liver, other features of the heart, etc.) are also possible without departing from the scope of this disclosure. The medical image may be identified via user input (e.g., an operator may input user input indicating that the medical image includes an anatomical ROI and should be segmented) or automatically (e.g., a computing device may determine that the medical image includes a standard imaging plane containing the anatomical ROI; and that the medical image should be segmented according to an imaging protocol, etc.).

[0054] At position 704, the medical image is input into the trained segmentation model. The segmentation model can be a deep learning model trained for segmenting anatomical ROIs, such as... Figure 2 The segmentation model 208 and / or based on Figure 4 The segmentation model trained by the method is used. At 706, the segmentation output of the segmentation model is received. The segmentation output may include a visual indication of the anatomical ROI overlaid on the medical image, or another suitable representation of the location, shape, size, etc. of the anatomical ROI within the medical image.

[0055] At position 708, the confidence metric for the segmentation is determined. The confidence metric indicates the shape of the segmentation output by the segmentation model relative to the mean shape of the anatomical ROI (e.g., see above reference). Figure 5 How well the mean shape matches. As previously explained, the mean shape is determined by multiple segments (e.g., expert-annotated medical images), not by the current segment (e.g., not using the current medical image to determine the mean shape). To determine the confidence metric, non-shape variations in the segmentation are eliminated (or at least reduced), as shown in 710. Non-shape variations are eliminated / reduced by registering the segmentation with the mean shape of the anatomical ROI, such as registration performed via Procrustes registration. The mean shape can be in Figure 4 The mean shape is determined at 410 points.

[0056] Determining the confidence metric also includes using the saved PCA pattern to transform the segmentation to a lower-dimensional space and back to form a reconstructed segmentation, as shown in Figure 712. (See above for reference.) Figure 4 In essence, PCA identifies a set of principal dimensions of variation for the entire segmentation set (e.g., multiple previous expert-generated segments). Each segment output by the segmentation model (e.g., the current segmentation output at 706) can then be stored (encoded) as its position along only each PCA dimension. These dimensions and their positions along each dimension can then be used to reconstruct the segments to determine a confidence metric. Therefore, after registration with the mean shape to reduce non-shape variations, the registered segments are transformed to a low-dimensional space and reconstructed using the stored PCA patterns. The stored PCA patterns can be in... Figure 4 The PCA pattern is determined at position 412 and stored at position 414. This can be determined according to the equation Y = (X - μ)P. T The segmentation is transformed to a low-dimensional space, where P is the component calculated using PCA (e.g., the saved PCA pattern), and μ is the mean shape. The segmentation can then be transformed back to a high-dimensional space according to the following equation to generate the reconstructed shape. At position 714, the confidence metric is calculated as the original segment X (e.g., the segmentation output by the segmentation model at position 706) versus the reconstructed segment. The difference between them. For example, the original segmentation and the reconstructed segmentation may each include a set of points (e.g., 12 points) that can be plotted on a common 2D coordinate system, and according to the equation The difference can be the sum of the squared differences between the points of the reconstructed segment and the original segment.

[0057] At 716, method 700 determines whether the confidence measure meets a predetermined condition relative to a threshold. For example, at 716, the method may determine whether the confidence measure is below a threshold. Figure 3 As shown and explained above, the confidence metric can be the difference between the original segment and the reconstructed segment, and therefore, the lower the difference (e.g., the lower the confidence metric), the higher the shape match between the original and reconstructed segments is likely to be. Figure 3 As shown, the segmentation model's reconstruction of the segmentation for the first image 302 has a difference of 0.235 compared to the original segmentation, which is higher than the threshold. Conversely, the segmentation model's reconstruction of the segmentation for the second image 310 has a difference of 0.070 compared to the original segmentation, which is lower than the threshold. Thus, the segmentation of the first image can be identified as atypical.

[0058] The threshold for distinguishing between typical and atypical segmentations can be determined empirically based on confidence metrics calculated for multiple segmentations. For example, Figure 8 As shown in this article (reference 1) Figure 7 A histogram 800 is calculated for the confidence metric as explained herein. The first set of facets 802 of the histogram may include segments with a confidence metric in the range of 0–0.10, and the second set of facets 804 may include segments with a confidence metric greater than 0.20. Therefore, the threshold may be set between 0.10 and 0.20, such as 0.15. Although the confidence metric is described herein as equal to the calculated gap, in some examples, the confidence metric may be determined as 1 minus the calculated gap, such that a higher confidence metric corresponds to a smaller gap. In such examples, the threshold may be a different value (e.g., 0.85), and segments with a confidence metric below the threshold may be identified as atypical.

[0059] Return to Figure 7If the confidence metric does indeed satisfy a condition relative to a threshold (e.g., the confidence metric is below a threshold), method 700 proceeds to 718 to output a segment for display, storage, and / or use in downstream processes. Because the confidence metric satisfies the condition relative to the threshold, the segment output by the segmentation model can be considered a typical shape, and therefore the segment can be used in one or more downstream processes, such as calculating strain, ejection fraction, etc. In some examples, the segment can be used by different models to automatically calculate strain, ejection fraction, etc. In some examples, the confidence metric can be displayed on a display device, such as display device 234, along with the medical image and the segment. When the confidence metric satisfies the condition relative to the threshold, the original segment (the segment output by the segmentation model) can be displayed, stored, and / or used in downstream processes, and in at least some examples, the original segment may not be corrected or modified. Method 700 then terminates.

[0060] If the confidence metric does not meet the condition relative to a threshold, method 700 proceeds to 720 to output a notification indicating that the segmentation output by the segmentation model is a low-confidence segment (e.g., the segment has an atypical shape). This notification may, for example, be output on a display device. In some examples, the notification may include prompts for an operator to initiate manual segmentation or acquire different images to input into the segmentation model, and the original segment output by the segmentation model may be discarded. Thus, when the confidence metric indicates that a segment has an atypical shape, the segment can be discarded without being used in any further downstream processes. Method 700 then terminates.

[0061] Therefore, Method 700 provides a confidence metric for the segmentation generated by a segmentation model that indicates how well the shape of the segment can be encoded by encoding one or more pre-determined major shape variation patterns of the segment. Thus, the confidence metric indicates how well the small number of dimensions (e.g., 2–4) found using principal component analysis can be used to reconstruct the shape of the segment. If the shape / contour of the segment output by the segmentation model cannot be reconstructed with a high degree of similarity to the original segment, the original segment is determined to fall outside the major shape variation patterns established by segmentation confirmed by multiple experts, and therefore the segment can be considered atypical. In this way, atypical segments can be discarded and new segments can be obtained (e.g., manually), which avoids errors in downstream processes (e.g., strain measurements) that rely on accurate segmentation of the ROI. To determine how well the shape of the segment can be encoded by encoding one or more major shape variation patterns, the difference between the reconstructed segment and the original segment can be calculated for each point of the reconstructed segment, as explained above. However, in other examples, in order to determine how well the shape of the segment can be encoded by encoding one or more major shape variation patterns, other metrics can be used, such as the reconstructed segment region relative to the original segment region, the reconstructed segment centroid position relative to the original segment centroid position, etc.

[0062] The disclosed method for calculating the confidence metric for segmentation of a given anatomical region of interest (ROI) offers several advantages. As previously explained, the confidence metric is determined based on the mean shape and one or more major variation patterns of the anatomical ROI, each of which is determined by multiple expert-generated segmentations. Thus, the mean shape and major variation patterns can be based on the segmentation itself, rather than on the expected shape of the anatomical ROI or known anatomical features. This can be advantageous because segmentations used in various procedures (e.g., strain measurement, ejection fraction calculation) can be based on the boundaries of the anatomical ROI in a given imaging plane and may not fully represent the entire shape of anatomical features within the patient's body. Therefore, the determination of atypical segmentation shapes versus typical segmentation shapes can be more accurate than comparisons with actual anatomical features. Furthermore, by defining the mean shape and each segmentation as a set of points plotted on 2D axes, eliminating non-shape variations and calculating gaps (e.g., the gap between the original segmentation and the reconstructed segmentation) can be straightforward and requires minimal processing power, allowing the confidence module to be executed on a wide variety of devices.

[0063] The technical advantage of calculating the confidence metric of segments output by the segmentation model is that atypical segments can be automatically identified and discarded without expert confirmation of each segment output by the segmentation model, thereby reducing errors in downstream processes that use those segments. In doing so, the efficiency of the computing device in performing downstream processes is improved by avoiding unnecessary processing associated with utilizing atypical segments to perform various functions (e.g., strain measurement).

[0064] This disclosure also provides support for a method comprising: receiving a segmentation of a region of interest (ROI) of a medical image, the segmentation being output by a segmentation model; calculating a confidence metric for the segmentation, the confidence metric indicating how well the shape of the segmentation can be encoded by encoding one or more principal shape variation patterns of a pre-determined set of segmentations of the ROI; displaying the segmentation, storing the segmentation, and / or using the segmentation for one or more downstream processes in response to the confidence metric satisfying the predetermined conditions relative to a threshold, otherwise prompting a user to perform manual segmentation. In a first example of the method, the medical image includes an ultrasound image. In a second example of the method, optionally including a first example, the ROI includes the left ventricle of a patient's heart, and one or more downstream processes include strain measurement. In a third example of the method, optionally including one or both of the first and second examples, the segmentation model is a deep learning model trained to output segmentations based on medical images. In a fourth example of the method, optionally including one or more or each of the first to third examples, calculating the confidence metric includes registering the segmentation with the mean shape of the ROI to reduce non-shape variations of the segmentation relative to the mean shape. In a fifth example of the method, optionally including one or more examples from the first to fourth examples, registering the segmentation with the mean shape includes performing Procrustes registration. In a sixth example of the method, optionally including one or more examples from the first to fifth examples, the mean shape is determined based on a predetermined set of segmentations of the ROI. In a seventh example of the method, optionally including one or more examples from the first to sixth examples, computing a confidence metric for the segmentation includes generating a reconstructed segmentation by transforming the segmentation to a low-dimensional space and back using encoding of one or more principal shape change patterns after registration with the mean shape, the confidence metric indicating how well the shape of the segmentation can be encoded by encoding one or more principal shape change patterns of a predetermined set of segmentations of the ROI. In an eighth example of the method, optionally including one or more examples from the first to seventh examples, computing the confidence metric further includes calculating the gap between the segmentation output by the segmentation model and the reconstructed segmentation. In a ninth example of the method, one or more examples from the first through eighth examples may be optionally included, which determine the encoding of one or more major shape variation patterns based on principal component analysis performed on a set of predetermined segmentations of the ROI. In a tenth example of the method, one or more examples from the first through ninth examples may be optionally included, where the set of predetermined segmentations of the ROI includes ground reality segmentations applied to train the segmentation model.

[0065] This disclosure also provides support for a system comprising: a processor and a non-transitory memory storing instructions executable by the processor to: input a patient medical image including a region of interest (ROI) as input to a segmentation model; receive a segmentation of the ROI from the segmentation model; register the segmentation with the mean shape of the ROI determined by a plurality of previous segmentations of the ROI, thereby forming a registered segmentation; generate a reconstructed segmentation from the registered segmentation using one or more major variation patterns determined by the plurality of previous segmentations of the ROI; calculate a confidence metric for the segmentation based on the reconstructed segmentation; and display, store, and / or use the segmentation for one or more downstream processes in response to the confidence metric satisfying a predetermined condition relative to a threshold. In a first example of the system, the medical image includes an ultrasound image, wherein the ROI includes the left ventricle of a patient's heart, and wherein one or more downstream processes include strain measurement. In a second example of the system, optionally including the first example, the instructions are executable to discard the segmentation and / or prompt a user to manually generate a new segmentation in response to the confidence metric not satisfying a predetermined condition relative to a threshold. In a third example of the system, optionally including one or both examples from the first and second examples, the confidence metric for calculating the segmentation based on the reconstructed segmentation includes calculating the gap between the reconstructed segmentation and the segmentation from the segmentation model. In a fourth example of the system, optionally including one or more or each example from the first to third examples, multiple previous segmentations of the ROI include ground-based segmentations applied to train the segmentation model.

[0066] This disclosure also provides support for a method comprising: training a segmentation model by inputting training data as input to the segmentation model to segment a region of interest (ROI) in a medical image, the training data including multiple medical images and multiple expert-generated segments of the ROI, each segment being associated with a corresponding medical image in the multiple medical images; determining a mean shape of the ROI by the multiple expert-generated segments; determining one or more principal shape variation patterns of the ROI by the multiple expert-generated segments; and deploying the mean shape and one or more principal variation patterns during inference using the trained segmentation model to determine a confidence metric for the ROI segmentation output by the trained segmentation model. In a first example of the method, determining the mean shape of the ROI by the multiple expert-generated segments includes: identifying an initial mean shape; registering each of the multiple expert-generated segments to the initial mean shape using Procrustes registration; computing a new mean shape from the registered segments; and iteratively repeating Procrustes registration and computing the new mean shape until a stable mean shape is generated. In a second example of the method, the first example may be optionally included, wherein the medical image includes an ultrasound image. In a third example of the method, one or both of the first and second examples may be included, with the ROI comprising the left ventricle of the heart.

[0067] When describing elements of various embodiments of this disclosure, the terms “an,” “a,” and “the” are intended to refer to one or more of these elements. The terms “first,” “second,” etc., do not indicate any order, quantity, or importance, but are used to distinguish one element from another. The terms “comprising,” “including,” and “having” are intended to be inclusive and mean that additional elements may exist in addition to the listed elements. As used herein, the terms “connected to,” “linked to,” etc., indicate that an object (e.g., a material, element, structure, component, etc.) may be connected to or linked to another object, regardless of whether the object is directly connected to or linked to the other object, or whether one or more intervening objects exist between the object and the other. Furthermore, it should be understood that references to “an embodiment” or “an embodiment” of this disclosure are not intended to be construed as excluding the existence of additional embodiments that also include the referenced features.

[0068] In addition to any modifications previously indicated, those skilled in the art can devise many other variations and alternative arrangements without departing from the spirit and scope of this description, and the appended claims are intended to cover such modifications and arrangements. Therefore, although the information has been described above in a specific and detailed manner in conjunction with what is currently considered to be the most practical and preferred aspects, it will be apparent to those skilled in the art that many modifications can be made without departing from the principles and concepts set forth herein, including but not limited to changes in form, function, mode of operation, and use. Likewise, as used herein, in all respects, examples and embodiments are intended to be illustrative only and should not be construed as restrictive in any way.

Claims

1. A method for segmenting an image, the method comprising: Receive the segmentation of the region of interest (ROI) of a medical image, the segmentation being output by a segmentation model; Calculate a confidence metric for the segmentation, which indicates how well the shape of the segmentation can be encoded by encoding one or more major shape variation patterns of a pre-determined set of segments of the ROI; In response to the confidence metric satisfying a predetermined condition relative to a threshold, the segment is displayed, the segment is stored, and / or the segment is used in one or more downstream processes; Otherwise, prompt the user to perform manual splitting.

2. The method according to claim 1, wherein the medical image comprises an ultrasound image.

3. The method of claim 1, wherein the ROI comprises the left ventricle of the patient's heart, and the one or more downstream processes comprise strain measurement.

4. The method of claim 1, wherein the segmentation model is a deep learning model trained to output the segmentation based on the medical image.

5. The method of claim 1, wherein calculating the confidence metric includes registering the segment with the mean shape of the ROI to reduce non-shape variations of the segment relative to the mean shape.

6. The method of claim 5, wherein registering the segmentation with the mean shape includes performing Procrustes registration.

7. The method of claim 5, wherein the mean shape is determined based on a predetermined set of segments of the ROI.

8. The method of claim 5, wherein calculating the confidence metric for the segmentation comprises: A reconstructed segment is generated by transforming the segmentation to a low-dimensional space and back using encoding of one or more major shape variation patterns after registration with the mean shape. The confidence metric indicates how well the shape of the segment can be encoded by encoding the one or more major shape variation patterns of a pre-determined set of segments of the ROI.

9. The method of claim 8, wherein calculating the confidence metric further comprises: Calculate the difference between the segmentation output by the segmentation model and the reconstructed segmentation.

10. The method of claim 8, wherein the encoding of the one or more major shape change patterns is determined based on principal component analysis performed on a predetermined set of segments of the ROI.

11. The method of claim 10, wherein a predetermined set of segmentations of the ROI includes ground-based segmentations applied to train the segmentation model.

12. A system for segmenting an image, the system comprising: processor; and A non-transitory memory that stores instructions that can be executed by the processor to: The input includes a patient medical image containing a region of interest (ROI) as input to the segmentation model; Receive the segmentation of the ROI from the segmentation model; The segment is registered with the mean shape of the ROI as determined by a plurality of previous segments of the ROI, thereby forming a registered segment; The reconstructed segment is generated from the registration segment using one or more major variation patterns determined by multiple previous segmentations of the ROI; Based on the reconstructed segmentation, calculate the confidence metric of the segmentation; as well as In response to the confidence metric satisfying a predetermined condition relative to a threshold, the segment is displayed, the segment is stored, and / or the segment is used in one or more downstream processes.

13. The system of claim 12, wherein the medical image comprises an ultrasound image, wherein the ROI comprises the left ventricle of the patient's heart, and wherein the one or more downstream processes comprise strain measurement.

14. The system of claim 12, wherein the instructions are executable to discard the segment and / or prompt the user to manually generate a new segment in response to the confidence metric not satisfying the predetermined condition relative to the threshold.

15. The system of claim 12, wherein calculating the confidence metric of the segmentation based on the reconstructed segmentation comprises: Calculate the difference between the reconstructed segment and the segment from the segmentation model.

16. The system of claim 12, wherein the plurality of previous segmentations of the ROI include ground-based segmentation applied to train the segmentation model.

17. A method for segmenting an image, the method comprising: The segmentation model is trained by inputting training data as input to segment regions of interest (ROI) in medical images. The training data includes multiple medical images and multiple expert-generated segmentations of the ROI, each segmentation being associated with a corresponding medical image in the multiple medical images. The mean shape of the ROI is determined by the multiple expert-generated segmentations; One or more major shape variation patterns of the ROI are determined by the multiple expert-generated segmentations; as well as During inference using the trained segmentation model, the mean shape and one or more major variation patterns are deployed to determine a confidence metric for the segmentation of the ROI output by the trained segmentation model.

18. The method of claim 17, wherein determining the mean shape of the ROI by the plurality of expert-generated segmentations comprises: Identify the shape of the initial mean; The Procrustes registration is used to register each of the plurality of expert-generated segments with the initial mean shape; and a new mean shape is calculated from the registered segments; and the Procrustes registration is iteratively repeated and the new mean shape is calculated until a stable mean shape is generated.

19. The method of claim 17, wherein the medical image comprises an ultrasound image.

20. The method of claim 17, wherein the ROI comprises the left ventricle of the heart.

Images