Machine learning-based systems and methods for two-dimensional measurements in magnetic resonance imaging (MRI) images

Abstract

A system may receive an MRI image. A system may receive a user input comprising a requested anatomical measurement. A system may input, into a trained supervised machine learning model, the MRI image and the requested anatomical measurement. A system may segment, using the trained supervised machine learning model, a plurality of endpoint regions associated with the requested anatomical measurement from the MRI image. A system may determine a two-dimensional (2D) measurement based on the plurality of endpoint regions.

Description

MACHINE LEARNING-BASED SYSTEMS AND METHODS FOR TWO-DIMENSIONAL MEASUREMENTS IN MAGNETIC RESONANCE IMAGING (MRI) IMAGESCross-Reference To Related Applications

[0001] This application claims the benefit of U.S. provisional patent application No.63 / 670,178, filed on July 12, 2024, and titled "MACHINE LEARNING-BASED SYSTEMS AND METHODS FOR TWO-DI MENSIONAL MEASUREMENTS IN MAGNETIC RESONANCE IMAGING (MRI) IMAGES," the disclosure of which is expressly incorporated herein by reference in its entirety.BACKGROUND

[0002] Magnetic Resonance Imaging (MRI) is a critical diagnostic tool in modern medicine, offering detailed images of the internal structures of the body. It is widely used to diagnose and monitor various conditions. Accurate measurement of structures within MRI images is crucial for diagnosis, treatment planning, and monitoring the progression of diseases.

[0003] Traditional methods for making two-dimensional (2D) measurements in MRI images involve manual annotation by radiologists, which is time-consuming and subject to human error. These manual processes can lead to variability in measurements, affecting the reliability of the diagnosis and treatment plans.

[0004] There is therefore a need in the art for a machine learning-based technique for making 2D measurements in MRI images.SUMMARY

[0005] In some aspects, the techniques described herein relate to a computer- implemented method for measurement in magnetic resonance imaging (MRI) images including: receiving an MRI image; receiving a user input including a requested anatomical measurement; inputting, into a trained supervised machine learning model, the MRI image and the requested anatomical measurement; segmenting, using the trained supervised machine learning model, a plurality of endpoint regions associated with the requested anatomical measurement from the MRIimage; and determining a two-dimensional (2D) measurement based on the plurality of endpoint regions.

[0006] In some aspects, the method further includes detecting respective centers of each of the plurality of endpoint regions, wherein the 2D measurement is determined based on the respective centers of each of the plurality of endpoint regions.

[0007] In some aspects, the method further includes generating display data for the 2D measurement.

[0008] In some aspects, the method further includes overlaying the display data for the 2D measurement on the MRI image.

[0009] In some aspects, the plurality of endpoint regions includes at least two endpoint regions.

[0010] In some aspects, each of the plurality of endpoint regions has a diameter of greater than 3 pixels and less than 12 pixels.

[0011] In some aspects, each of the plurality of endpoint regions has a circular shape.

[0012] In some aspects, the 2D measurement is at least one of ascending aorta, descending aorta, main pulmonary artery, right pulmonary artery, left pulmonary artery, inferior vena cava, right ventricular end diastolic dimension, left ventricular end diastolic dimension, intraventricular septal thickness, left ventricular end systolic dimension, posterior wall thickness, left atrial dimension, left atrial length, right atrial length, or aortic root dimension.

[0013] In some aspects, the method further includes generating, using a language model, a report associated with the 2D measurement.

[0014] In some aspects, the trained supervised machine learning model is a convolutional neural network (CNN).

[0015] In some aspects, the MRI image is a cardiac MRI image.

[0016] In some aspects, the techniques described herein relate to a computer- implemented method for training a machine learning model including: receiving a plurality ofmagnetic resonance imaging (MRI) images; creating a labeled dataset by labeling, in each of the plurality of MRI images, a plurality of endpoint regions associated with anatomical measurements; and training a supervised machine learning model using the labeled dataset, wherein the trained supervised machine learning model is configured to segment the plurality of endpoint regions associated with the anatomical measurements from a new MRI image.

[0017] In some aspects, each of the plurality of endpoint regions has a diameter of greater than 3 pixels and less than 12 pixels.

[0018] In some aspects, each of the plurality of endpoint regions has a circular shape.

[0019] In some aspects, the supervised machine learning model is a convolutional neural network (CNN).

[0020] In some aspects, the MRI images are cardiac MRI images.

[0021] It should be understood that the above-described subject matter may also be implemented as a computer-controlled apparatus, a computer process, a computing system, or an article of manufacture, such as a computer-readable storage medium.

[0022] Other systems, methods, features and / or advantages will be or may become apparent to one with skill in the art upon examination of the following drawings and detailed description. It is intended that all such additional systems, methods, features and / or advantages be included within this description and be protected by the accompanying claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The components in the drawings are not necessarily to scale relative to each other. Like reference numerals designate corresponding parts throughout the several views.

[0024] FIGURE 1 is a flowchart illustrating example operations for measurement in MRI images according to implementations described herein.

[0025] FIGURE 2 is a flowchart illustrating example operations for training a machine learning model according to implementations described herein.

[0026] FIGURE 3 is an example computing device.

[0027] FIGURE 4 is a flowchart illustrating example operations for training and deploying a supervised machine learning model for measurement in MRI images according to the Examples described herein.

[0028] FIGURE 5 illustrates a graphical user interface of MATLAB's medical image labeling system for deep learning.

[0029] FIGURE 6 illustrates a graphical user interface of an application in MATLAB for measurement in MRI images according to the Examples described herein.

[0030] FIGURE 7 illustrates an example system for automatically determining 2D measurements of MRI images, according to implementations described herein.

[0031] FIGURE 8A illustrates a method of determining measurements of a pulmonary artery bifurcation slice and inferior vena cava slice, according to the Examples described herein.

[0032] FIGURE 8B illustrates methods of determining measurements of an end-diastolic frame and end-systolic frame, according to the Examples described herein.

[0033] FIGURE 8C illustrates methods of determining measurements of an end-systolic frame, according to the Examples described herein.

[0034] FIGURE 8D illustrates methods of determining measurements of an end systolic frame, according to the Examples described herein.DETAILED DESCRIPTION

[0035] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. Methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present disclosure. As used in the specification, and in the appended claims, the singular forms "a," "an," "the" include plural referents unless the context clearly dictates otherwise. The term "comprising" and variations thereof as used herein are used synonymously with the term "including" and variations thereof and are open, non-limiting terms. The terms "optional" or "optionally" usedherein mean that the subsequently described feature, event or circumstance may or may not occur, and that the description includes instances where said feature, event or circumstance occurs and instances where it does not. Ranges may be expressed herein as from "about" one particular value, and / or to "about" another particular value. When such a range is expressed, an aspect includes from the one particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by use of the antecedent "about," it will be understood that the particular value forms another aspect. It will be further understood that the endpoints of each of the ranges are significant both in relation to the other endpoint, and independently of the other endpoint.

[0036] Definitions

[0037] As used herein, the terms "about" or "approximately" when referring to a measurable value such as an amount, a percentage, and the like, is meant to encompass variations of ±20%, ±10%, ±5%, or ±1% from the measurable value.

[0038] The term "subject" is defined herein to include animals such as mammals, including, but not limited to, primates (e.g., humans), cows, sheep, goats, horses, dogs, cats, rabbits, rats, mice and the like. In some embodiments, the subject is a human.

[0039] A convolutional neural network (CNN) is a type of deep neural network that has been applied, for example, to image analysis applications. Deep neural networks are computing systems including multiple layers of interconnected neurons (e.g., also referred to as "nodes"). Unlike a traditional neural network, each layer in a CNN has a plurality of nodes arranged in three dimensions (width, height, depth). CNNs can include different types of layers, e.g., convolutional, pooling, and fully-connected (also referred to herein as "dense") layers. A convolutional layer includes a set of filters and performs the bulk of the computations. A pooling layer is optionally inserted between convolutional layers to reduce the computational power and / or control overfitting (e.g., by downsampling). A fully-connected layer includes neurons, where each neuron is connected to all of the neurons in the previous layer. The layers are stacked similar to traditional neuralnetworks. Graph convolutional neural networks (GCNNs) are CNNs that have been adapted to work on structured datasets such as graphs.

[0040] Example Methods

[0041] Recent advances in machine learning have shown promise in automating various aspects of medical image analysis, thereby reducing the workload on radiologists and increasing the accuracy and consistency of measurements. Machine learning algorithms, particularly deep learning models, have significant potential in recognizing patterns and making precise measurements from medical images.

[0042] There is no existing machine learning-based solution for determining 2D linear measurements in cardiac magnetic resonance imaging (MRI) images. Conventional image segmentation techniques are not configured for determining 2D measurements. One technical challenge is training a supervised machine learning model such as a CNN to segment endpoints of a line for making 2D linear measurement. Currently, no cardiac imaging software exists that can automate that process. Conventionally, CNN training is used to segment heart tissue from other structures, and these segments are used to produce a 3D model. However, cardiac imaging and quantification has been based on 2D measurements since the 1970's. One potential solution would be to attempt to fit the cardiac segmentation to a 2D shape, such as a half-oval. However, such solutions do not provide a true measurement, just a close fit.

[0043] The techniques described herein address technical problems associated with existing solutions. Implementations of the present disclosure configure systems and computer- implemented methods for automated endpoint identification, allowing computerized 2D measurements of cardiac structures. Specialized machine learning models described herein are trained and / or configured to allow segmentation models to be used for linear measurements. Using specialized machine learning models increases the accuracy of cardiac measurements, and allows the automation of cardiac measurements that were not possible to accurately perform using existing automated techniques.

[0044] For example, the techniques described herein include methods of automatically identifying endpoint regions within cardiac MRI images, where the endpoint regions can be used to computationally calculate linear distances of structures in heart that have been imaged using cardiac MRI. The techniques described herein further include methods of training machine learning models to automatically identify endpoint regions, for example by training a supervised machine learning model with cardiac MRI images having labeled endpoint regions, which are approximately circular. For example, as described below, each of the circular endpoint regions may have a diameter between 3 pixels and 12 pixels, optionally about 6 pixels. As described herein, the supervised learning model trained on such a labeled dataset can accurately find the endpoints of a line, which can be used to make 2D linear measurements. Accordingly, the systems and methods described herein are configured to overcome the limitations of conventional technologies, and thereby automate 2D linear measurements within 3D MRI images.

[0045] Fig. 1 is a flowchart of an example method for measurement in MRI images. This disclosure contemplates that these operations can be performed using one or more computing devices. A computing device includes at least one processor and memory, (e.g. as shown by box 302 in Fig. 3).

[0046] At step 110, the method includes receiving an MRI image. Optionally, in some implementations, the MRI image is a cardiac MRI image. It should be understood that cardiac MRI images are provided only as an example. This disclosure contemplates that the methods described herein can be applied for measurements in images capturing anatomies other than a heart.

[0047] At step 120, the method includes receiving a user input comprising a requested anatomical measurement. An example user interface for receiving user inputs is shown in Fig. 6.

[0048] At step 130, the method includes inputting, into a trained supervised machine learning model, the MRI image and the requested anatomical measurement. The trained supervised machine learning model is operating in inference mode. The trained supervised machine learning model has therefore been trained with a labeled data set (or "dataset"). As described herein, thetrained supervised machine learning model "learns" a function that maps an input (also known as feature or features) to an output (also known as target or targets) during training with a labeled data set. Machine learning model training is discussed in further detail below, for example with regard to Fig. 2 and in the Examples. In the implementations described herein, the input is the MRI image and the requested anatomical measurement, and the output is a segmented image (see step 140 below). Optionally, in some implementations, the trained supervised machine learning model is a convolutional neural network (CNN). It should be understood that CNNs are provided only as an example. This disclosure contemplates that machine learning models other than CNNs may be used with the methods described herein.

[0049] At step 140, the method includes segmenting, using the trained supervised machine learning model, a plurality of endpoint regions associated with the requested anatomical measurement from the MRI image. Image segmentation is the process of partitioning the MRI image (e.g. a digital image) into multiple segments or regions to simplify its representation and make it more meaningful and easier to analyze. In the implementations described herein, the trained supervised learning model is configured to segment (e.g. partition) the plurality of endpoint regions associated with the requested anatomical measurement. Optionally, the plurality of endpoint regions include at least two endpoint regions. Alternatively or additionally, each of the plurality of endpoint regions has a circular shape. It should be understood that circular shape is provided only as an example. This disclosure contemplates endpoint regions having other than circular or approximately circular shapes. In some implementations, each of the plurality of endpoint regions has a diameter of greater than 3 pixels. Alternatively or additionally, in some implementations, each of the plurality of end point regions has a diameter of less than 12 pixels. For example, each of the plurality of endpoint regions may have a diameter between 3 pixels and 12 pixels. Optionally, each of the plurality of endpoint regions may have a diameter of about 6 pixels.

[0050] At step 150, the method includes determining a 2D measurement based on the plurality of endpoint regions. For example, the 2D measurement is a length of a line betweenendpoint regions. The 2D measurement can be aortic root right coronary cusp ( RCC) to commissure, aortic root left coronary cusp (LCC) to commissure, aortic root non-coronary cusp (NCC) to commissure, ascending aorta, descending aorta, main pulmonary artery, right pulmonary artery, left pulmonary artery, inferior vena cava, left upper pulmonary vein, left lower pulmonary vein, right upper pulmonary vein, right lower pulmonary vein, left ventricular end diastolic dimension, left ventricular end systolic dimension, left ventricular septum, left ventricular posterior wall, right ventricular end diastolic dimension, left atrial length, right atrial length, left atrial length (3ch), aortic valve, mitral valve, tricuspid valve, pulmonic valve, sinotubular junction, proximal transverse arch, distal transverse arch, aortic isthmus. Optionally, the 2D measurement can be at least one of ascending aorta, descending aorta, main pulmonary artery, right pulmonary artery, left pulmonary artery, inferior vena cava, right ventricular end diastolic dimension, left ventricular end diastolic dimension, interventricular septal thickness, left ventricular end systolic dimension, posterior wall thickness, left atrial dimension, left atrial length, right atrial length, or aortic root dimension. It should be understood that the 2D measurements above are provided only as examples. This disclosure contemplates making 2D measurements other than those provided as examples, for example by training the supervised machine learning model to segment different endpoint regions. In some implementations, the method optionally further includes detecting respective centers of each of the plurality of endpoint regions, where the 2D measurement is determined based on the respective centers of each of the plurality of endpoint regions. In other words, the 2D measurement is a length of a line between respective centers of the endpoint regions.

[0051] In some implementations, the method optionally further includes generating display data for the 2D measurement. Additionally, the method optionally further includes overlaying the display data for the 2D measurement on the MRI image. This is shown, for example, in Fig. 6.

[0052] In some implementations, the method optionally further includes generating, using a language model, a report associated with the 2D measurement. Optionally, the methodfurther includes generating, using a language model and scripting techniques, a report associated with the 2D measurement. For example, the 2D measurement can be correlated with a text label. The text label and 2D measurement can then be compared to a table of known normal values. It should be understood that multiple publications have been made regarding normal values in cardiac MRI. Optionally, users may choose a table of normal values for comparison. For basic quantitative measurements, scripting can be used to generate written descriptions such as, "Left ventricular posterior wall thickness is (mild, moderately, or severely) thickened." The current published normal values guidelines include ranges for classifying measurements such as mild, moderate, or severe. For qualitative descriptions, LLMs can be used such as, "The inferior wall is hypokinetic." Reports can then be generated using an LLM and optionally displayed on the user interface. In other words, the 2D measurements can be formatted into a verbal report for the medical records using LLMs. An example report output is shown, for example, in Fig. 6.

[0053] Fig. 2 is a flowchart of an example method for training a machine learning model. This disclosure contemplates that these operations can be performed using one or more computing devices. A computing device includes at least one processor and memory, e.g. as shown by box 302 in Fig. 3.

[0054] At step 210, the method includes receiving a plurality of MRI images. Optionally, in some implementations, the MRI images are cardiac MRI images. It should be understood that cardiac MRI images are provided only as an example. This disclosure contemplates that the methods described herein can be applied for measurements in images capturing anatomies other than a heart.

[0055] At step 220, the method includes creating a labeled dataset by labeling, in each of the plurality of MRI images, a plurality of endpoint regions associated with anatomical measurements. Optionally, the endpoint regions can be circular regions of interest ("RO Is"). The centroid of each ROI can be selected as an endpoint for a linear measurement. The circular ROI can optionally be a 6-pixel diameter circle, where the center of the circle is an endpoint of a linearmeasurement. Alternatively or additionally, the endpoint regions can be defined as regions of interest other than circular, such as line segments, linear measurement segments, and / or any other shapes. This can be accomplished using a medical image labeling tool running on a computing device, e.g., the MATLAB Medical Image Labeling application shown in Fig. 5. For example, a plurality of short axis slices (e.g. MRI images) can be imported into the medical image labeling application. In the Examples below, 200 images are imported (and endpoints marked as described below). It should be understood that the number of imported images (i.e. 200) is provided only as an example. This disclosure contemplates importing more or less than 200 images, particularly more than 200 images.

[0056] Thereafter, a plurality of endpoint regions are marked on each of the MRI images. Endpoints can be marked at each of the following locations at the tip of the papillary muscles: anterior epicardium, anterior endocardium, inferior endocardium, and inferior epicardium. It should be understood that there are typically 2 papillary muscles in the left ventricle (1 anterior, 1 posterior) and 3 papillary muscles in the right ventricle (1 anterior, 1 posterior, 1 septal). Thus, for each papillary muscle, endpoint regions can be marked at the outer layer (epicardium) and inner lining (endocardium) of the heart wall in each of the MRI images. As described above, each of the plurality of endpoint regions has a circular shape. Alternatively or additionally, each of the plurality of endpoint regions may have a diameter between 3 pixels and 12 pixels. Optionally, each of the plurality of endpoint regions may have a diameter of about 6 pixels.

[0057] Optionally, one or more data augmentation steps can be performed to improve the training of the machine learning models described herein. For example, the images can be resized (e.g. to 256 x 256 pixels). Alternatively or additionally, the endpoint regions can be treated as image masks, and also resized (e.g., to 256 x 256 pixels). Additional data augmentations that can be applied include random horizontal flip, random vertical flip, random 10-degree rotation, random brightness adjustment, random contrast adjustment, random cropping to focus on boundaries.Alternatively or additionally, class weight optimization can be applied.

[0058] These data augmentation steps have the benefits of allowing the model to be applicable to different patient position or size. In addition, the data augmentation improves model performance across different models and configurations of MRI scanner (e.g., allowing the model to be used across a range of different machines from different manufacturers).

[0059] An example architecture that can be used is the DeepLab v3 architecture with any or all of the modifications described below.

[0060] Implementations of the present disclosure include a Modified ASPP (Atrous Spatial Pyramid Pooling) Module. The modified ASPP includes reduced dilation rates for finer detail preservation. Example modifications to the dilation rates include: first atrous convolution: changed from (2,2) to (1,1); second atrous convolution: changed from (4,4) to (2,2); and / or third atrous convolution: changed from (6,6) to (3,3).

[0061] Alternatively or additionally, the modified ASPP can include additional ASPP branches configured for enhanced feature extraction: Global average pooling branch with lxl convolution; and / or attention mechanism branch with sigmoid activation for feature refinement

[0062] Alternatively or additionally, implementations of the present disclosure can further include an Enhanced Classifier Head. A multi-layer classifier can replace the standard single convolution, including: 3x3 convolution with batch normalization and ReLU; Dropout2D (0.1) for regularization; and / or final lxl convolution for class prediction. Optionally, the same modifications can be applied to the auxiliary classifier.

[0063] Implementations of the present disclosure can further include measurementspecific Enhancements. An advanced centroid calculation algorithm can include any or all of the following: multi-scale analysis with adaptive scale selection based on object size; and Enhanced boundary detection (e.g., using one or more of: gradient-based edge detection, morphological erosion-based boundaries, and / or combined boundary weighting). Alternatively or additionally, distance transform weighting can be applied for center-biased calculations, validation andrefinement of centroid positions can be performed, and / or confidence scoring can be used for reliability assessment.

[0064] Alternatively or additionally, implementations of the present disclosure can further include a boundary-Aware Loss Function. The boundary- A ware Loss Function can optionally include any or all of: Sobel edge detection for identifying boundaries; Adaptive boundary weighting that increases loss contribution at object edges; Dynamic loss balancing between Dice and CrossEntropy components; and / or confidence-based weighting that focuses on uncertain predictions

[0065] Alternatively or additionally, implementations of the present disclosure can further include an input size validation step. Input size validation can optionally include automatic resizing of inputs smaller than 8x8 pixels to prevent processing errors and / or a forward pass modification with size checking.

[0066] At step 230, the method includes training a supervised machine learning model using the labeled dataset. Optionally, the supervised machine learning model is a CNN. The trained supervised machine learning model is configured to segment the plurality of endpoint regions associated with the anatomical measurements from a new MRI image.

[0067] An example CNN includes an encoder / decoder architecture with a residual network (e.g. a Resnet50 backbone with 50 layers). The training procedure can include a 60 / 20 / 20 train / validation / test data split. In some implementations, an Adam optimizer is used. Example hyperparameters used in training include:

[0068] Learning Rate: 0.0001

[0069] Batch Size: 8

[0070] Epochs: 300

[0071] Optimizer: Adam

[0072] Output Stride: 8

[0073] Patience: 40

[0074] The optimization can be based on minimization of a combined loss function, where the combined loss function includes DICE loss, cross entropy loss, and / or boundary-aware loss components. For example, training can be completed when a DICE score of greater than .75 is reached. As described above, example architecture that can be used is the DeepLab v3 architecture. The segmentation models used for endpoints of 2D measurements are optimized for mean centroid distance < 3 pixels.

[0075] The supervised machine learning models can be trained to output tensors. The tensors include logits (raw scores) for each class of pixels. The tensors can be converted to probabilities by softmax and processed to obtain a final segmentation mask using argmax, which can select the class with the highest probability being selected for each pixel. For example, the result can be a mask file where each pixel is assigned an integer representing a predicted class of each pixel in the image.

[0076] Additional example model architecture, parameters, and hyperparameters are discussed in the Examples below. The trained supervised machine learning model can then be deployed for measurement in MRI images, for example as described with regard to Fig. 1.

[0077] Example Computing Device

[0078] It should be appreciated that the logical operations described herein with respect to the various figures may be implemented (1) as a sequence of computer implemented acts or program modules (i.e., software) running on a computing device (e.g., the computing device described in Fig. 3), (2) as interconnected machine logic circuits or circuit modules (i.e., hardware) within the computing device and / or (3) a combination of software and hardware of the computing device. Thus, the logical operations discussed herein are not limited to any specific combination of hardware and software. The implementation is a matter of choice dependent on the performance and other requirements of the computing device. Accordingly, the logical operations described herein are referred to variously as operations, structural devices, acts, or modules. These operations, structural devices, acts and modules may be implemented in software, in firmware, in specialpurpose digital logic, and any combination thereof. It should also be appreciated that more or fewer operations may be performed than shown in the figures and described herein. These operations may also be performed in a different order than those described herein.

[0079] Referring to Fig. 3, an example computing device 300 upon which the methods and computer-implemented methods described herein may be implemented is illustrated. It should be understood that the example computing device 300 is only one example of a suitable computing environment upon which the methods described herein may be implemented. Optionally, the computing device 300 can be a well-known computing system including, but not limited to, personal computers, servers, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, network personal computers (PCs), minicomputers, mainframe computers, embedded systems, and / or distributed computing environments including a plurality of any of the above systems or devices. Distributed computing environments enable remote computing devices, which are connected to a communication network or other data transmission medium, to perform various tasks. In the distributed computing environment, the program modules, applications, and other data may be stored on local and / or remote computer storage media.

[0080] In its most basic configuration, computing device 300 typically includes at least one processing unit 306 and system memory 304. Depending on the exact configuration and type of computing device, system memory 304 may be volatile (such as random access memory (RAM)), non-volatile (such as read-only memory (ROM), flash memory, etc.), or some combination of the two. This most basic configuration is illustrated in Fig. 3 by box 302. The processing unit 306 may be a standard programmable processor that performs arithmetic and logic operations necessary for operation of the computing device 300. The computing device 300 may also include a bus or other communication mechanism for communicating information among various components of the computing device 300.

[0081] Computing device 300 may have additional features / functionality. For example, computing device 300 may include additional storage such as removable storage 308 and non-removable storage 310 including, but not limited to, magnetic or optical disks or tapes. Computing device 300 may also contain network connection(s) 316 that allow the device to communicate with other devices. Computing device 300 may also have input device(s) 314 such as a keyboard, mouse, touch screen, etc. Output device(s) 312 such as a display, speakers, printer, etc. may also be included. The additional devices may be connected to the bus in order to facilitate communication of data among the components of the computing device 300. All these devices are well known in the art and need not be discussed at length here.

[0082] The processing unit 306 may be configured to execute program code encoded in tangible, computer-readable media. Tangible, computer-readable media refers to any media that is capable of providing data that causes the computing device 300 (i.e., a machine) to operate in a particular fashion. Various computer-readable media may be utilized to provide instructions to the processing unit 306 for execution. Example tangible, computer-readable media may include, but is not limited to, volatile media, non-volatile media, removable media and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. System memory 304, removable storage 308, and non-removable storage 310 are all examples of tangible, computer storage media. Example tangible, computer-readable recording media include, but are not limited to, an integrated circuit (e.g., field-programmable gate array or application-specific IC), a hard disk, an optical disk, a magneto-optical disk, a floppy disk, a magnetic tape, a holographic storage medium, a solid-state device, RAM, ROM, electrically erasable program read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices.

[0083] In an example implementation, the processing unit 306 may execute program code stored in the system memory 304. For example, the bus may carry data to the system memory304, from which the processing unit 306 receives and executes instructions. The data received by thesystem memory 304 may optionally be stored on the removable storage 308 or the non-removable storage 310 before or after execution by the processing unit 306.

[0084] It should be understood that the various techniques described herein may be implemented in connection with hardware or software or, where appropriate, with a combination thereof. Thus, the methods and apparatuses of the presently disclosed subject matter, or certain aspects or portions thereof, may take the form of program code (i.e., instructions) embodied in tangible media, such as floppy diskettes, CD-ROMs, hard drives, or any other machine-readable storage medium wherein, when the program code is loaded into and executed by a machine, such as a computing device, the machine becomes an apparatus for practicing the presently disclosed subject matter. In the case of program code execution on programmable computers, the computing device generally includes a processor, a storage medium readable by the processor (including volatile and non-volatile memory and / or storage elements), at least one input device, and at least one output device. One or more programs may implement or utilize the processes described in connection with the presently disclosed subject matter, e.g., through the use of an application programming interface (API), reusable controls, or the like. Such programs may be implemented in a high level procedural or object-oriented programming language to communicate with a computer system. However, the program(s) can be implemented in assembly or machine language, if desired. In any case, the language may be a compiled or interpreted language and it may be combined with hardware implementations.

[0085] Examples

[0086] The following examples are put forth so as to provide those of ordinary skill in the art with a complete disclosure and description of how the compounds, compositions, articles, devices and / or methods claimed herein are made and evaluated, and are intended to be purely exemplary and are not intended to limit the disclosure. Efforts have been made to ensure accuracy with respect to numbers (e.g., amounts, temperature, etc.), but some errors and deviations shouldY1be accounted for. Unless indicated otherwise, parts are parts by weight, temperature is in °C or is at ambient temperature, and pressure is at or near atmospheric.

[0087] Example 1:

[0088] An example implementation of the present disclosure includes systems and methods to automate linear measurements in cardiac MRI images. Fig. 4 illustrates the example operations for training and deploying a supervised machine learning model. Fig. 5 illustrates a graphical user interface of MATLAB's medical image labeling system for deep learning. Fig. 6 illustrates a graphical user interface of an application in MATLAB for measurement in MRI images. A total of 200 short axis slices were imported into the MATLAB Medical Image Labeling application (see Fig. 5), and a plurality of points (e.g. endpoint regions) were placed on each image at the following locations at the tip of the papillary muscles: anterior epicardium, anterior endocardium, inferior endocardium, and inferior epicardium. As discussed below, each of the endpoint regions has a diameter greater than 3 pixels and less than 12 pixels. Optionally, each of the endpoint regions has a diameter of about 6 pixels. The labeled images were used to traina DeepLab v3+ with a ResNet-50 backbone (e.g. a supervised machine learning model, specifically a CNN) with the following parameters: trainingOptions('adam', lnitialLearnRate=0.001, GradientDecayFactor=0.9, L2Regularization=0.01, MaxEpochs=500, MiniBatchSize=64, Shuffle- 'every-epoch", LearnRateSchedule="piecewise", Verbose=false, ValidationData=dsVal, ExecutionEnvironment- 'auto", ResetlnputNormalization=false, Plots- 'training-progress"). Once the models were trained to detect the endpoint regions, the models were applied to validation images. The center of the detected endpoint regions was detected and listed as separate points. The points were built into final measurement lines, and the length of lines were exported as lengths: anterior wall thickness, left ventricular end-diastolic dimension, inferior wall thickness.

[0089] Testing showed that supervised machine learning models were poorly trained and reproducibility was difficult when using endpoint regions having a diameter of 3 pixels or less. Additionally, it was found that supervised machine learning models had a fair amount of variability inthe endpoint localization when using endpoint regions having a diameter of 12 pixels or more. In contrast, it was found that supervised machine learning models performed well when using endpoint regions having a diameter of 6 pixels. Specifically, such trained models were found to have acceptable accuracy and loss, and when used for inference, these models made measurements with lesser difference from human made measurements.

[0090] Example 2

[0091] As used herein, the following terms are defined:

[0092] A block diagram of an example implementation is shown in FIG. 7. The example implementation includes an image analysis system 700 configured to receive any or all of the MRI images 702. The image analysis system can include a flow module 712, a two-dimensional measurement module 714, and / or a volumetric module 716. The image analysis system 700 can be implemented using one or more computing devices 300 as described with reference to FIG. 3. The image analysis system 700 can be configured to perform any of the methods described herein, and output quantitative imaging values 720. Optionally, the quantitative imaging values 720 can be transmitted to another computing device, and / or output for display.

[0093] With reference to FIGS. 8A-8D, example methods of image analysis are disclosed according to implementations of the present disclosure. The methods of FIG. 8A-8D can optionally be performed by the two-dimensional measurement module 712 of FIG. 7.

[0094] FIG. 8A shows an example method 800. At step 802, an axial DICOM image stack is received. DICOM is intended only as a non-limiting example format, and the image received at step 802 can be any format of MRI image.

[0095] At steps 804a and 804b, the method includes selecting images from the image stack received at step 802. The images can optionally be selected by trained machine learning classifiers. As shown in FIG. 8A, step 804a can include inputting the image stack into a trained classifier configured to find the pulmonary artery bifurcation slice. Step 804b can include inputting the image stack into a trained classifier configured to find the inferior vena cava slice.

[0096] At step 806a, the pulmonary artery bifurcation slice can be input into the trained segmentation models described herein configured to find the endpoints of linear measurements of the pulmonary artery bifurcation slice. Example linear measurements include the ascending aortic diameter, descending aortic diameter, and main pulmonary artery diameter.

[0097] At step 806b, the inferior vena cava slice can be input into the trained segmentation models described herein. The trained segmentation model can be configured to determine an outline of a vessel (e.g., the inferior vena cava or aorta). The outline of the vessel can then be used to geometrically determine a smallest linear measurement that traverses through a centroid of the outlined vessel.

[0098] Example 2D measurements of the left ventricle include the interventricular septum, posterior wall, left-ventricular end-diastolic dimension, and right ventricular end-diastolic dimension, which are defined by the following endpoints: right ventricular endocardium, right ventricular septum, left ventricular septum, left ventricular posterior wall endocardium, and left ventricular posterior wall epicardium.

[0099] FIG. 8B illustrates another example method 820 according to implementations of the present disclosure. At step 822, a LVOT ("left ventricular outflow tract") DICOM image stack is received. Again, DICOM is intended only as a non-limiting example image format.

[0100] At step 824a and 824b, LV blood pool segmentation is performed. LV blood pool segmentation can be used to find both the end-diastolic frame in step 824a, and the end- systolic frame in step 824b. As used in the present disclosure, "blood pool" segmentation refers to any method of image segmentation configured identify the region of the left ventricle that is filled with blood.

[0101] At step 826a, the method can include applying a segmentation model trained according to implementations of the present disclosure to the end-diastolic frame of the LVOT image stack. The trained segmentation model can be configured to find endpoints corresponding to the endpoints of linear measurements in the left ventricle. Example 2D measurements of the left ventricle include the interventricular septum, posterior wall, left- ventricular end-diastolic dimension, and right ventricular end-diastolic dimension, which are defined by the following endpoints: right ventricular endocardium, right ventricular septum, left ventricular septum, left ventricular posterior wall endocardium, and left ventricular posterior wall epicardium.

[0102] At step 826b, the method can include applying a segmentation model trained according to implementations of the present disclosure to determine 2D measurements of the end-systolic frame identified in step 824b. Example 2D measurements include left ventricular end-systolic dimensions, left atrial length, and right atrial length. These measurements can be defined by the following endpoints: left ventricular septal endocardium, left ventricular posterior wall endocardium, Left atrial anterior wall, and left atrial posterior wall.

[0103] FIG. 8C illustrates another example method 840 according to implementations of the present disclosure. At step 842, the method includes receiving a four- chamber DICOM image stack. Again, DICOM is intended only as a non-limiting example image format.

[0104] At step 844, the method includes applying LV blood pool segmentation (as described with reference to FIG. 8B) to find the end-systolic frame.

[0105] At step 846, the method includes applying a segmentation model trained according to implementations of the present disclosure to the end-systolic frame of the four- chamber image stack. The trained segmentation model can be configured to find endpoints corresponding to linear measurements in a four-chamber view of the heart shown in the end- systolic frame. Example 2D measurements of the end-systolic frame include left atrial area (two chamber), left atrial length, and right atrial length, which are defined by the following endpoints and / or line segments: Left atrial anterior wall, center of mitral valve, right atrial posterior wall, center of tricuspid valve. The left atrial area can be calculated by summing the number of pixels enclosed by line segments defining the left atrial area. The 2-D dimensions of the pixels (e.g., length and width) can be known, so that the left atrial area can be determined by summing the pixels of the left atrial area.

[0106] FIG. 8D illustrates yet another example method 860 according to implementations of the present disclosure. At step 862, the method includes receiving a two- chamber DICOM image stack. Again, DICOM is intended only as a non-limiting example imageformat. At step 864, an LV blood pool segmentation model is applied to find an end-systolic frame (as described with reference to FIG. 8B).

[0107] At step 866, the method includes inputting the end-systolic frame of the two-chamber DICOM image stack into a segmentation model trained according to implementations of the present disclosure. The segmentation model can be configured to determine endpoints corresponding to linear measurements. Example 2D measurements include left atrial area (two- chamber) and left atrial length. Left atrial area (two-chamber) can be calculated as described above by summing the number of pixels in the left atrial area and multiplying the number of pixels by the known real-world dimensions of each pixel. Left atrial length can be determined by the following endpoints: Left atrial posterior wall and center of the mitral valve.

[0108] The machine learning classifiers described with reference to FIGS. 8A-8D can optionally be trained using cardiac MRI scans. The images can be labeled by human cardiologists (e.g., to create image masks) and then the images and / or image masks can be resized (e.g., to 224 x 224 pixels). An example architecture for training the classifiers includes a ResNetSO model with the following example hyperparameters:

[0109] Learning Rate: 0.0001

[0110] Batch Size: 32

[0111] Epochs: 20

[0112] The training run can be completed upon determining that an accuracy of greater than 85% is reached.

[0113] Although the subject matter has been described in language specific to structural features and / or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

Claims

WHAT IS CLAIMED:

1. A computer-implemented method for measurement in magnetic resonance imaging (MRI) images comprising: receiving an MRI image; receiving a user input comprising a requested anatomical measurement; inputting, into a trained supervised machine learning model, the MRI image and the requested anatomical measurement; segmenting, using the trained supervised machine learning model, a plurality of endpoint regions associated with the requested anatomical measurement from the MRI image; and determining a two-dimensional (2D) measurement based on the plurality of endpoint regions.

2. The computer-implemented method of claim 1, further comprising detecting respective centers of each of the plurality of endpoint regions, wherein the 2D measurement is determined based on the respective centers of each of the plurality of endpoint regions.

3. The computer-implemented method of claim 1, further comprising generating display data for the 2D measurement.

4. The computer-implemented method of claim 1, further comprising overlaying a display of the 2D measurement on the MRI image.

5. The computer-implemented method of claim 1, wherein the plurality of endpoint regions comprise at least two endpoint regions.

6. The computer-implemented method of claim 1, wherein each of the plurality of endpoint regions has a diameter of greater than 3 pixels and less than 12 pixels.

7. The computer-implemented method of claim 1, wherein each of the plurality of endpoint regions has a circular shape.

8. The computer-implemented method of claim 1, wherein the 2D measurement is at least one of ascending aorta, descending aorta, main pulmonary artery, right pulmonary artery, left pulmonary artery, inferior vena cava, right ventricular end diastolic dimension, left ventricular enddiastolic dimension, intraventricular septal thickness, left ventricular end systolic dimension, posterior wall thickness, left atrial dimension, left atrial length, right atrial length, or aortic root dimension.

9. The computer-implemented method of claim 1, further comprising generating, using a language model, a report associated with the 2D measurement.

10. The computer-implemented method of claim 1, wherein the trained supervised machine learning model is a convolutional neural network (CNN).

11. The computer-implemented method of claim 1, wherein the MRI image is a cardiac MRI image.

12. The computer-implemented method of claim 11, wherein the MRI image is an image of an end-diastolic frame or an end-systolic frame.

13. The computer-implemented method of claim 1, wherein the method further comprises receiving an MRI image stack and selecting, from the MRI image stack, the MRI image.

14. The computer-implemented method of claim 13, wherein the MRI image is selected by a classifier model configured to select a pulmonary artery bifurcation slice or an IVC slice.

15. The computer-implemented method of claim 13, wherein the MRI image is selected by a left-ventricular blood pool segmentation model.

16. A computer-implemented method for training a machine learning model comprising: receiving a plurality of magnetic resonance imaging (MRI) images; creating a labeled dataset by labeling, in each of the plurality of MRI images, a plurality of endpoint regions associated with anatomical measurements; and training a supervised machine learning model using the labeled dataset, wherein the trained supervised machine learning model is configured to segment the plurality of endpoint regions associated with the anatomical measurements from a new MRI image.

17. The computer-implemented method of claim 16, wherein each of the plurality of endpoint regions has diameter of greater than 3 pixels and less than 12 pixels.

18. The computer-implemented method of claim 16, wherein each of the plurality of endpoint regions has a circular shape.

19. The computer-implemented method of claim 16, wherein the supervised machine learning model is a convolutional neural network (CNN).

20. The computer-implemented method of claim 16, wherein the MRI images are cardiac MRI images.

Images