Heart MRI image dynamic segmentation and function evaluation system

Abstract

The invention discloses a heart MRI image dynamic segmentation and function evaluation system, and relates to the technical field of image semantic segmentation. The method comprises the following steps: firstly, segmenting a dynamic sequence based on a U-Net model integrated with a Monte Carlo Dropout strategy, and generating a segmentation contour, an uncertainty heat map and a global quality evaluation score; secondly, when interaction conditions are met, a correction process is started, a user instruction is received through an uncertainty sensing pen brush, a contour is updated through a deformation model, and a mass score is recalculated; thirdly, morphological and dynamic characteristics are extracted, and heart function parameters are calculated; and finally, performing credibility correction on the functional parameters through Monte Carlo sampling and a weighting function based on the mass fraction, and generating a credibility interval. According to the method, quantification of segmentation uncertainty and credibility transmission of a function evaluation result are realized, and the reliability and clinical applicability of heart MRI analysis are improved.

Description

Technical Field

[0001] This invention relates to the field of image semantic segmentation technology, specifically to a dynamic segmentation and functional assessment system for cardiac MRI images. Background Technology

[0002] In the field of medical image analysis, cardiac magnetic resonance imaging (MRI) has become an important examination tool for obtaining information on cardiac structure and function due to its non-invasive nature and excellent soft tissue resolution. Accurate segmentation of cardiac structures within these images and subsequent functional assessment based on the segmentation results are core technical steps in assisting the diagnosis of various cardiovascular diseases.

[0003] However, existing technologies face significant challenges in this process. Currently, cardiac structure analysis mainly relies on two modes: one is completely manual segmentation by physicians, which is not only time-consuming and inefficient, but also largely dependent on the operator's subjective experience, making it difficult to guarantee the repeatability and consistency of the results. The other is segmentation using fully automated deep learning models, which improves speed, but the inherent black-box nature of the models makes their decision-making process lack transparency. Crucially, the models lack the ability to quantify the reliability of their segmentation results. This forces clinicians to still perform tedious manual reviews of the automated segmentation results to ensure diagnostic safety, failing to substantially liberate human resources. Furthermore, in the subsequent functional assessment stage, the inability to perceive and quantify the inherent uncertainties in the initial segmentation results leads to doubts about the confidence level of key functional parameters such as ejection fraction and ventricular volume, potentially affecting the accuracy of clinical decisions. Summary of the Invention

[0004] This invention addresses the technical problems of low efficiency and insufficient reliability in existing cardiac magnetic resonance imaging analysis due to reliance on manual segmentation or unreliable fully automated segmentation models, by providing a dynamic segmentation and functional assessment system for cardiac MRI images.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] This invention provides a dynamic segmentation and functional assessment system for cardiac MRI images, comprising:

[0007] The initial contour segmentation module is used to process MRI dynamic sequence images based on a preset deep learning segmentation model, obtain the segmentation contour of the heart structure and the segmentation certainty of each contour point, generate an overall uncertainty heatmap, and calculate the initial global quality assessment score.

[0008] The interactive correction module is used to initiate the interactive correction process when the uncertainty heatmap and the initial global quality assessment score meet the preset interaction conditions, receive the user's correction instructions, update the segmentation contour and uncertainty heatmap in real time, obtain the contour matching degree of the segmentation contour, and update the initial global quality assessment score.

[0009] The functional parameter acquisition module is used to extract the morphological and dynamic features of the updated segmented contour, and input the morphological and dynamic features into the preset functional parameter calculation module to obtain N functional parameters of the heart.

[0010] The parameter range limitation module is used to perform confidence correction on the calculated N functional parameters based on the updated global quality assessment score, and obtain the confidence range of the N functional parameters.

[0011] The beneficial effects of this invention are:

[0012] Compared to existing technologies, this invention first quantifies the uncertainties in the segmentation process and generates intuitive heatmaps and global quality scores, providing a clear basis for assessing the reliability of initial segmentation results. Secondly, it introduces an interactive correction mechanism that intelligently guides users to prioritize corrections of high-uncertainty regions and updates segmentation results and quality assessments in real time, improving the efficiency and relevance of human-computer interaction. Thirdly, it links segmentation quality with subsequent functional parameter calculations, ensuring a more reliable data foundation for the process from morphological features to dynamic parameter extraction. Finally, it performs confidence correction on the finally calculated cardiac function parameters based on the updated global quality score, outputting statistically significant confidence intervals rather than single, definitive values. This allows clinicians to intuitively grasp the confidence level of functional assessment results, effectively reducing the risk of erroneous clinical decisions due to segmentation errors, and achieving controllable quality and reliable results throughout the entire process from image segmentation to functional assessment. Attached Figure Description

[0013] Figure 1 A schematic diagram of the structure of the dynamic segmentation and functional assessment system for cardiac MRI images provided by the present invention;

[0014] Figure 2 This is a schematic diagram illustrating the principle of the dynamic segmentation and functional assessment system for cardiac MRI images provided by the present invention.

[0015] In the attached diagram, the components represented by each number are as follows:

[0016] Initial contour segmentation module 11, interactive correction module 12, function parameter acquisition module 13, parameter range limitation module 14. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0020] Example 1, as Figure 1 , Figure 2 As shown, this embodiment of the invention provides a dynamic segmentation and functional assessment system for cardiac MRI images, including:

[0021] The initial contour segmentation module 11 is used to process MRI dynamic sequence images based on a preset deep learning segmentation model, obtain the segmentation contour of the heart structure and the segmentation certainty of each contour point, generate an overall uncertainty heatmap, and calculate the initial global quality assessment score.

[0022] Specifically, based on a pre-defined deep learning segmentation model, dynamic MRI sequence images are processed to obtain the segmentation contours of the heart structure and the segmentation certainty of each contour point. An overall uncertainty heatmap is generated, and an initial global quality assessment score is calculated, including:

[0023] Construct a deep learning segmentation model based on the U-Net architecture and enable the Monte Carlo Dropout strategy during the model inference stage;

[0024] For each frame of the input MRI dynamic sequence, perform T forward propagation inferences to obtain T segmentation probability maps;

[0025] For each pixel in the segmentation probability map, the average of T prediction results is calculated as the final class probability of the corresponding pixel, and the final segmentation contour is generated by thresholding.

[0026] For each pixel, the standard deviation of the prediction results of T times is calculated as the segmentation certainty of the corresponding pixel. The uncertainty heatmap is constructed based on the segmentation certainty of all pixels.

[0027] The average segmentation certainty of all pixels within the segmented contour region is statistically analyzed and normalized to serve as the initial global quality assessment score.

[0028] First, a deep learning segmentation model based on the U-Net architecture is constructed. The U-Net architecture refers to a convolutional neural network that adopts an encoder-decoder structure. This structure fuses high-resolution features from the encoding stage with corresponding upsampled features from the decoding stage through skip connections. Specifically, the encoder extracts and compresses image features step by step through a series of convolution and pooling operations, while the decoder recovers the spatial resolution of the feature map step by step through a series of upsampling and convolution operations, ultimately outputting a segmentation probability map with the same size as the input.

[0029] Specifically, the deep learning segmentation model built on the U-Net architecture is an end-to-end image segmentation network. This deep learning segmentation model can predict the probability of each point in the input image belonging to different cardiac structure categories pixel by pixel, and is used to accurately identify key anatomical structures such as the endocardium and epicardium from single-frame images of dynamic sequences of cardiac magnetic resonance imaging.

[0030] Simultaneously, during the model inference phase, a Monte Carlo Dropout strategy is employed. This strategy refers to a method that randomly activates Dropout layers during the model prediction phase to approximate Bayesian inference. Specifically, during forward propagation, a specific proportion of neuron connections in the neural network are randomly dropped multiple times, ensuring that the same input yields slightly different outputs after multiple inferences, thus achieving an effective estimate of the uncertainty in model prediction.

[0031] Specifically, deep learning segmentation models based on the U-Net architecture include:

[0032] Construct a U-Net deep learning segmentation model with an encoder-decoder structure;

[0033] The input to the deep learning segmentation model is a single-frame image at a time point in a preprocessed MRI dynamic sequence;

[0034] The output of the deep learning segmentation model is a segmentation probability map with multiple channels, which is the same size as the input image, and each channel corresponds to a heart structure category.

[0035] Specifically, this deep learning segmentation model employs a typical encoder-decoder structure as its basic network framework. The encoder is responsible for feature extraction, progressively compressing the spatial dimension and increasing the number of feature channels through successive convolution and pooling operations, thereby capturing multi-level feature information from local details to global semantics in the image. The decoder is responsible for feature reconstruction and precise localization, progressively restoring the spatial resolution of the feature map through successive upsampling and convolution operations, and combining skip connections from features at the same level as the encoder to fuse high-resolution location information with deep semantic information, thereby achieving the utilization of high-level semantic features while maintaining localization accuracy.

[0036] The input data for this deep learning segmentation model is a single frame image at a specific time point from a preprocessed dynamic sequence of magnetic resonance imaging (MRI). The dynamic MRI sequence is a series of images continuously acquired at different time points within a single cardiac cycle. This sequence is obtained using cardiac cine imaging technology, which can completely record the periodic motion of the heart from end-diastole to end-systole. Preprocessing operations typically include intensity normalization and size standardization. Intensity normalization aims to eliminate image grayscale differences caused by different scanning devices and parameters, mapping image pixel values ​​to a uniform range. Size standardization adjusts the input image to a fixed size specified by the model through interpolation and other operations, ensuring that the network structure can process it correctly. Preprocessing reduces the interference of non-physiological variations on the deep learning segmentation model's learning, while improving the convergence speed and generalization performance of the deep learning segmentation model.

[0037] Furthermore, the final output of the deep learning segmentation model is a segmentation probability map with the exact same height and width as the input image. This segmentation probability map contains multiple channels, each output channel specifically responsible for predicting the probability distribution of a particular cardiac structure category. For example, channel one corresponds to the left ventricular endocardium, channel two to the left ventricular myocardium, and channel three to the right ventricular endocardium, etc. Each pixel location in each channel outputs a probability value between zero and one, representing the confidence that the pixel belongs to the corresponding cardiac structure category.

[0038] For example, due to the blurred boundaries of tissue structures and significant individual morphological differences in cardiac magnetic resonance imaging, and the powerful feature learning and pixel-level classification capabilities of deep learning models in image semantic segmentation tasks, a deep learning segmentation model based on the U-Net architecture was chosen. Furthermore, considering the need for quantitative evaluation of the reliability of segmentation results in clinical applications, a Monte Carlo Dropout strategy was employed during the model inference stage to estimate uncertainty.

[0039] Specifically, this deep learning segmentation model employs a classic encoder-decoder architecture. The encoder consists of four downsampling stages, each containing two 3×3 convolutional layers and a modified linear unit activation function, followed by a 2×2 max-pooling layer for downsampling. The decoder uses bilinear interpolation upsampling and feature skip connections to the corresponding stages of the encoder to gradually restore spatial resolution. The final output layer uses a 1×1 convolution and a sigmoid activation function to generate a multi-channel segmentation probability map with the same size as the input image. During model training, all convolutional layers are followed by a dropout layer with a uniform dropout rate of 0.3. This setting ensures both model regularization and provides a foundation for subsequent Monte Carlo uncertainty estimation.

[0040] During training, key hyperparameters included a learning rate of 0.0001, 200 training epochs, and a batch size of 8. Training data was sourced from multiple publicly available cardiac magnetic resonance imaging datasets, with all training samples containing segmentation labels for endocardium, epicardium, and other classes, annotated by cardiac imaging experts. All samples were divided into training, validation, and test sets in an 8:1:1 ratio. A weighted combination of the Dice loss function and cross-entropy loss function was used as the optimization objective, with parameters updated via the Adam optimizer. Model performance was continuously monitored using the validation set during training; training was terminated when the Dice coefficient on the validation set no longer improved after 20 consecutive epochs, yielding the optimal model parameters.

[0041] The Monte Carlo Dropout strategy employed during the model inference phase endows the trained deterministic segmentation model with uncertainty quantification capabilities similar to those of a Bayesian model. Specifically, it enables a quantitative assessment of the reliability of the segmentation results, rather than simply providing a single segmentation output. By calculating the standard deviation of the classification result for each pixel through the prediction distribution generated by multiple random forward propagations, the model's understanding of the confidence level for segmenting different regions can be accurately characterized spatially, providing clear guidance for subsequent interactive correction processes.

[0042] Furthermore, based on the trained deep learning segmentation model, for each frame of the input cardiac magnetic resonance imaging (MRI) dynamic sequence, T independent forward propagation inference processes are performed, resulting in T segmentation probability maps for the same input frame. A segmentation probability map is a multi-channel matrix with the same spatial size as the input image, where the value of each pixel position on different channels represents the probability confidence of the deep learning segmentation model in classifying that pixel as belonging to a specific cardiac anatomical structure. These T repeated inference processes aim to fully sample the prediction distribution generated by the deep learning segmentation model due to random dropout. By maintaining the activation state of the dropout layer during the inference phase, each forward propagation generates a network structure with subtle differences, ensuring that the T inference results effectively reflect the output fluctuations of the deep learning segmentation model due to internal randomness when facing the same input.

[0043] Secondly, after obtaining T segmentation probability maps, for each pixel in the segmentation probability map, its probability value corresponding to the same cardiac structure category in T predictions is statistically processed. Specifically, the arithmetic mean of the probability values ​​is calculated, and this average is used as the final category probability of the pixel belonging to that cardiac structure category. This final category probability is compared with a predefined probability threshold. When the final category probability is greater than or equal to the threshold, the pixel is determined to belong to the current cardiac structure category; when the final category probability is less than the threshold, the pixel is determined not to belong to the current cardiac structure category. The predefined probability threshold is a configurable numerical parameter, set according to the segmentation accuracy and recall requirements of the specific clinical application scenario; for example, it can be set to 0.5 as the default threshold.

[0044] Simultaneously, for each pixel, the standard deviation of its multiple predicted probability values ​​is calculated. This standard deviation quantifies the volatility of the deep learning segmentation model's classification of that pixel and is defined as the segmentation certainty for that pixel. The lower the segmentation certainty value, the more consistent the multiple predictions of the deep learning segmentation model, and the more reliable the segmentation result at that location; the higher the value, the more dispersed the prediction results and the greater the uncertainty.

[0045] Arranging the segmentation certainty of all pixels in the entire image according to their spatial location creates an uncertainty heatmap that reflects the spatial distribution of the segmentation result's uncertainty. An uncertainty heatmap is a two-dimensional spatial distribution map visually represented by color gradients, where the color intensity or hue variation of each pixel corresponds to the numerical value of its segmentation certainty. This uncertainty heatmap characterizes the reliability of the deep learning segmentation model's segmentation results for various anatomical regions in cardiac MRI images. It can be used to precisely guide clinicians' attention, prioritize the review and interactive correction of image regions with concentrated uncertainty, thereby optimizing workflows and improving the overall reliability assessment efficiency of segmentation results.

[0046] Finally, to further quantify the overall segmentation quality, statistics are performed on all pixels within the target region defined by the generated segmentation contours, and the average segmentation certainty of each pixel is calculated. This average segmentation certainty reflects the average reliability of the entire segmentation result. Next, this average segmentation certainty is normalized, for example, by linearly transforming it to a standardized scoring range of 0-100. The normalized value is defined as the initial global quality assessment score. This global quality assessment score serves as a single quantitative indicator of the overall reliability of the segmentation of the current frame image, providing a quantitative basis for subsequent interactive correction decisions.

[0047] The interactive correction module 12 is used to start the interactive correction process when the uncertainty heatmap and the initial global quality assessment score meet the preset interaction conditions, receive the user's correction instructions, update the segmentation contour and uncertainty heatmap in real time, obtain the contour matching degree of the segmentation contour, and update the initial global quality assessment score.

[0048] Specifically, when the uncertainty heatmap and the initial global quality assessment score meet preset interaction conditions, an interactive correction process is initiated. This process receives correction instructions from the user and updates the segmentation contour and uncertainty heatmap in real time. Simultaneously, the contour matching degree of the segmentation contour is obtained, and the initial global quality assessment score is updated, including:

[0049] Preset uncertainty threshold, quality assessment score threshold, and maximum number of highlighted areas X;

[0050] When the initial global quality assessment score is lower than the quality assessment score threshold, or the number of consecutive regions exceeding the uncertainty threshold in the uncertainty heatmap is greater than X, it is determined that the preset interaction condition is met.

[0051] In the uncertainty heatmap, the first X consecutive regions where the uncertainty value exceeds the uncertainty threshold are identified as highlighted areas;

[0052] Based on the uncertainty-aware brush tool, it receives contour correction operations from users in either the highlighted or non-highlighted areas. When performing correction operations in the highlighted area, the model response weight is higher than that in the non-highlighted area, and correction operations in the non-highlighted area will trigger a confirmation prompt.

[0053] Based on the user's contour correction operation, the segmentation contours of the current frame and adjacent time frames are updated in real time through the deformation model, and the uncertainty heatmap of the contour correction area is recalculated.

[0054] Based on the difference between the corrected segmented contour and the segmented contour initially predicted by the model, the contour fit of the segmented contour is calculated, wherein the smaller the difference, the higher the contour fit.

[0055] The average segmentation certainty is recalculated by combining the contour fit with the updated uncertainty heatmap, and the initial global quality assessment score is updated, wherein the contour fit is proportional to the updated global quality assessment score.

[0056] Before initiating the interactive correction process, several key parameters need to be preset as the basis for judgment. These include an uncertainty threshold for identifying high-uncertainty regions, a quality assessment score threshold for evaluating the overall segmentation quality, and a maximum number of highlighted regions (X) to limit the number of suggested regions. These parameters are set based on accuracy requirements in clinical practice, the anatomical characteristics of the heart structure, and cognitive load studies of human-computer interaction. For example, the uncertainty threshold can be set to 0.15 based on the prediction confidence distribution of the deep learning segmentation model on the validation set; the quality assessment score threshold can be set to 75 points based on the clinically acceptable minimum segmentation quality; and the maximum number of highlighted regions (X) can be set to 5 regions based on visual attention research recommendations. This series of parameter configurations provides clear quantitative standards for subsequent interactive decisions.

[0057] Specifically, when the initial global quality assessment score is lower than the preset quality assessment score threshold, or when the number of continuous regions exceeding the uncertainty threshold in the uncertainty heatmap is greater than the preset maximum number of highlighted regions X, it is determined that the current segmentation result requires manual intervention to meet the preset interaction conditions.

[0058] After determining that interactive correction is needed, the first step is to identify regions in the uncertainty heatmap. The first X consecutive regions whose uncertainty values ​​exceed a preset uncertainty threshold are then highlighted as key areas. These highlighted areas represent the set of spatial locations where the deep learning segmentation model's predictions show significant discrepancies and low confidence, providing clear priority targets for interactive correction operations.

[0059] Secondly, a specially designed uncertainty-aware brush tool is used for contour correction. This tool is an intelligent drawing tool that dynamically adjusts its interactive behavior based on the uncertainty value of the current location. When the user operates in a non-highlighted area, a confirmation prompt mechanism is triggered to prevent accidental operations in areas where the deep learning segmentation model is highly confident. When correcting in other areas, the uncertainty-aware brush tool assigns higher weights to the model's response, ensuring that expert opinions are fully adopted in uncertain areas of the deep learning segmentation model.

[0060] Based on the user's contour correction operations, the segmentation contours of the current frame and adjacent time frames are updated in real time using a deformation model. This mechanism can intelligently propagate expert single-frame correction results to the entire time series, effectively maintaining the temporal consistency of the cardiac dynamic sequence. Simultaneously, it is necessary to recalculate the uncertainty heatmap of the contour correction region to promptly reflect changes in the reliability of the corrected segmentation results.

[0061] Secondly, the contour fit is calculated based on the difference between the corrected segmentation contour and the initial segmentation contour predicted by the deep learning segmentation model. Specifically, the smaller the difference, the higher the contour fit, indicating that the initial segmentation result is closer to the result after expert correction. This contour fit quantifies the magnitude of expert correction and reflects the quality of the initial segmentation result.

[0062] Finally, by combining the calculated contour fit with the updated uncertainty heatmap, the average segmentation certainty is recalculated, and the initial global quality assessment score is updated accordingly. The contour fit and the updated global quality assessment score are directly proportional. When the contour fit is low, it indicates that the initial segmentation had a large error, but after effective correction by experts, the global quality assessment score will significantly improve. When the contour fit is high, it indicates that the initial segmentation was reliable to begin with, and the improvement in the global quality assessment score is relatively small.

[0063] The functional parameter acquisition module 13 is used to extract the morphological and dynamic features of the updated segmented contour, and input the morphological and dynamic features into the preset functional parameter calculation module to obtain N functional parameters of the heart.

[0064] Specifically, the morphological and dynamic features of the updated segmented contours are extracted, including:

[0065] Based on the ventricular cavity segmentation contour of each time frame, the ventricular cavity volume of each time frame is calculated, and the volume values ​​of all time frames are connected to generate a ventricular cavity volume curve.

[0066] In the end-diastolic frame, the volume of the myocardium is calculated based on the epicardial and endocardial contours, and multiplied by the myocardial density to obtain the myocardial mass.

[0067] In the end-diastolic and end-systolic frames, the ventricular wall is divided into several segments, and the wall thickness and its rate of change in each segment are calculated.

[0068] The ventricular cavity volume curve, myocardial mass, ventricular wall thickness and their rate of change are used as the morphological and dynamic features of the updated segmentation profile.

[0069] First, based on the ventricular cavity segmentation contours of each time frame in the dynamic sequence, the ventricular cavity volume at the corresponding time point is calculated using either the Simpson method or the voxel accumulation method. Specifically, the Simpson method divides the ventricular cavity into multiple circular slices at equal intervals along its long axis, calculates the slice area based on the endocardial contour at each slice, and then integrates the volume based on the slice thickness and the number of slices. The voxel accumulation method directly counts the total number of voxels contained within the ventricular cavity segmentation contours and multiplies the total number of voxels by the physical volume of a single voxel to obtain the volume value. By sequentially connecting the volume measurements from each time frame, a complete curve of ventricular cavity volume changing over time is constructed. This ventricular cavity volume curve clearly demonstrates the filling and emptying process of the ventricle during the cardiac cycle, providing core data support for assessing cardiac pump function.

[0070] During the end-diastolic timeframe, the volume data of myocardial tissue is obtained simultaneously using the epicardial and endocardial contours through a three-dimensional spatial volume calculation method. Specifically, the overall heart volume enclosed by the epicardial contour and the volume of the cardiac chambers enclosed by the endocardial contour are calculated separately. Subtracting the cardiac chamber volume from the overall heart volume yields the myocardial tissue volume. Then, multiplying the myocardial tissue volume value by myocardial density accurately calculates the myocardial mass parameter. This myocardial mass parameter is of significant reference value for the diagnosis of pathological changes such as myocardial hypertrophy. Myocardial density is a biological parameter characterizing the physical properties of myocardial tissue, obtained through histological measurements and consensus in medical literature; for example, 1.05 grams per cubic centimeter can be used as the standard myocardial density value.

[0071] For the two key time points of end-diastole and end-systole, several standard anatomical segments were divided along the ventricular wall. Within each anatomical segment, the ventricular wall thickness was obtained by calculating the shortest spatial distance between the endocardial and epicardial surfaces. Furthermore, the rate of change in ventricular wall thickness for each anatomical segment was calculated by comparing the thickness measurements at end-diastole and end-systole. This set of parameters can quantify the contractile function and motor coordination of the local myocardium.

[0072] Finally, the aforementioned ventricular cavity volume curves, myocardial mass, segmental wall thickness, and their rate of change were collectively used as morphological and dynamic features extracted from the updated segmented contours. These morphological and dynamic features comprehensively cover the global functional indicators and local mechanical properties of the heart, laying a data foundation for subsequent calculation of cardiac function parameters and clinical assessment.

[0073] Furthermore, the morphological and dynamic characteristics are input into a preset functional parameter calculation module to obtain N functional parameters of the heart, including:

[0074] The N functional parameters include at least ejection fraction and stroke volume;

[0075] From the ventricular volume curve, the maximum value is extracted as the end-diastolic volume, and the minimum value is extracted as the end-systolic volume.

[0076] Based on the end-diastolic volume and end-systolic volume, the ejection fraction and stroke volume are calculated.

[0077] After obtaining complete ventricular cavity volume curves, myocardial mass, and ventricular wall thickness and its rate of change parameters, these morphological and dynamic characteristics are input into a pre-defined functional parameter calculation module for further processing. This functional parameter calculation module contains a series of mathematical calculation models based on cardiac physiology principles, used to derive N types of clinically diagnostic cardiac function parameters from the basic characteristics.

[0078] Specifically, the N functional parameters include at least ejection fraction and stroke volume. Ejection fraction is the percentage of ventricular stroke volume to ventricular end-diastolic volume, a core indicator of cardiac pumping efficiency; stroke volume is the total amount of blood ejected by one ventricle during a single cardiac cycle, representing the actual blood-pumping capacity of the heart with each contraction.

[0079] In the specific calculation process, extreme value analysis is first performed on the ventricular cavity volume curve. The maximum volume value throughout the entire cardiac cycle is identified and extracted from this curve; this value corresponds to the volume when the ventricular cavity reaches its maximum filling state at end-diastole, i.e., the end-diastolic volume. Simultaneously, the minimum volume value is extracted from the curve; this value corresponds to the state when the ventricular cavity reaches its minimum emptying state at end-systole, i.e., the end-systolic volume.

[0080] Based on the obtained end-diastolic volume and end-systolic volume, the ejection fraction and stroke volume can be calculated. The ejection fraction is calculated by subtracting the end-systolic volume from the end-diastolic volume, dividing the difference by the end-diastolic volume, and multiplying by a percentage. This parameter reflects the proportion of blood pumped out by the ventricle with each systole and is an important indicator for assessing cardiac systolic function. Stroke volume is calculated by directly subtracting the end-systolic volume from the end-diastolic volume. This parameter characterizes the absolute amount of blood pumped out by the ventricle in a single heartbeat.

[0081] This calculation process can accurately extract two core clinical parameters, ejection fraction and stroke volume, from the basic ventricular volume curve, providing a quantitative basis for cardiac function assessment.

[0082] The parameter range limiting module 14 is used to perform confidence correction on the calculated N functional parameters based on the updated global quality assessment score, and obtain the confidence range of the N functional parameters.

[0083] Specifically, based on the updated global quality assessment score, the calculated N functional parameters are reliably corrected to obtain the reliability ranges for the N functional parameters, including:

[0084] Based on the uncertainty heatmap, M groups of possible profile variants conforming to the uncertainty distribution are generated by Monte Carlo sampling;

[0085] For each set of contour variants, recalculate all of its N functional parameters;

[0086] For each functional parameter, the corresponding M calculation results are used to form a statistical distribution. The P% quantile to (100-P)% quantile of the statistical distribution is taken as the initial confidence interval of the functional parameter, and N initial confidence intervals of N functional parameters are obtained, where P is greater than 0 and less than 100.

[0087] The widths of the N initial confidence intervals are weighted and corrected using the updated global quality assessment score to obtain N confidence intervals for the N functional parameters. The higher the updated global quality assessment score, the narrower the corrected confidence interval width.

[0088] First, based on the spatial uncertainty distribution reflected in the uncertainty heatmap, multiple sets of possible contour variants conforming to this distribution are generated using the Monte Carlo sampling method. These possible contour variants represent multiple alternative segmentation results that reasonably fluctuate the initial segmentation contour while maintaining the basic anatomical structure of the heart, characterizing the reasonable output range that the deep learning segmentation model might produce when considering its own prediction uncertainty. Each set of possible contour variants is a reasonable variation obtained by randomly perturbing the initial segmentation contour according to the segmentation certainty of each pixel. Pixel regions with lower segmentation certainty exhibit larger morphological fluctuations, while pixel regions with higher segmentation certainty maintain relatively stable contour features.

[0089] Secondly, for each set of contour variants, the complete function parameter calculation process is executed, and the N function parameters corresponding to that set of contour variants are recalculated. This process will generate M sets of function parameter calculation results, each set of results corresponding to one possible contour variant.

[0090] Furthermore, for each specific functional parameter, the M calculation results obtained from multiple profile variants are aggregated to form a statistical distribution. From this statistical distribution, a preset P% quantile to the (100-P)% quantile is extracted as the initial confidence interval for that functional parameter. Here, P is set according to the clinical diagnostic requirements for the parameter's confidence level, and the P value is a preset constant greater than 0 and less than 100. This method yields the initial confidence interval for each functional parameter, forming a set of N initial confidence intervals for N functional parameters.

[0091] Finally, the width of the obtained initial confidence interval is weighted and corrected using the updated global quality assessment score. Since the global quality assessment score is positively correlated with the overall reliability of the segmented contour, the initial confidence interval needs to be adjusted accordingly based on the global quality assessment score.

[0092] For example, the weighted adjustment of the confidence interval width can be implemented using the following linear function: Adjusted confidence interval width = Initial confidence interval width × [1 - α × (Global quality assessment score / 100)], where α is a weighting coefficient used to control the influence of the global quality assessment score on the interval width. This weighting coefficient α is set according to the stringency of the accuracy requirements for functional parameters in clinical applications, for example, it can be set to 0.5. When the global quality assessment score is a perfect score of 100, the adjusted interval width will narrow to (1 - α) times the initial width; when the global quality assessment score is 0, the interval width remains unchanged from its initial value. This linear weighting mechanism ensures that the adjustment range of the confidence interval maintains a continuous and quantifiable correspondence with the segmentation quality assessment results.

[0093] Specifically, the correction principle is as follows: the higher the updated global quality assessment score, the better the overall reliability of the segmentation result, and the narrower the corresponding confidence interval of the functional parameters should be; conversely, the lower the global quality assessment score, the wider the confidence interval should be. After this weighted correction process, N confidence intervals for N functional parameters, corrected by the quality assessment scores, are finally obtained.

[0094] In summary, the embodiments of this application have at least the following technical effects:

[0095] Compared to existing technologies, this application firstly introduces a Monte Carlo Dropout-based uncertainty quantification mechanism to achieve pixel-by-pixel evaluation of the reliability of cardiac magnetic resonance image segmentation results, generating uncertainty heatmaps and global quality scores with clear clinical guidance significance. Secondly, it designs an interactive correction process based on uncertainty perception, effectively guiding experts to prioritize key areas and prevent misoperation through highlighted prompts and intelligent brush tools, thus improving human-computer collaboration efficiency. Thirdly, it achieves intelligent propagation of single-frame correction results over time through a deformation model, ensuring the temporal consistency of segmentation results in dynamic analysis. Finally, it establishes a transfer mechanism from segmentation quality to functional parameter reliability, generating clinically valuable reliability intervals through multiple rounds of sampling and statistical inference. This allows cardiac function assessment results to retain the automation advantages of deep learning while possessing the interpretability and reliability of traditional methods, providing a more comprehensive reference for clinical decision-making.

[0096] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0097] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0098] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A dynamic segmentation and functional assessment system for cardiac MRI images, characterized in that, The system includes: The initial contour segmentation module is used to process MRI dynamic sequence images based on a preset deep learning segmentation model, obtain the segmentation contour of the heart structure and the segmentation certainty of each contour point, generate an overall uncertainty heatmap, and calculate the initial global quality assessment score. The interactive correction module is used to initiate the interactive correction process when the uncertainty heatmap and the initial global quality assessment score meet the preset interaction conditions, receive the user's correction instructions, update the segmentation contour and uncertainty heatmap in real time, obtain the contour matching degree of the segmentation contour, and update the initial global quality assessment score. The functional parameter acquisition module is used to extract the morphological and dynamic features of the updated segmented contour, and input the morphological and dynamic features into the preset functional parameter calculation module to obtain N functional parameters of the heart. The parameter range limitation module is used to perform confidence correction on the calculated N functional parameters based on the updated global quality assessment score, and obtain the confidence range of the N functional parameters.

2. The dynamic segmentation and functional assessment system for cardiac MRI images according to claim 1, characterized in that, Based on a pre-defined deep learning segmentation model, dynamic MRI sequence images are processed to obtain the segmentation contours of cardiac structures and the segmentation certainty of each contour point. An overall uncertainty heatmap is generated, and an initial global quality assessment score is calculated, including: Construct a deep learning segmentation model based on the U-Net architecture and enable the Monte Carlo Dropout strategy during the model inference stage; For each frame of the input MRI dynamic sequence, perform T forward propagation inferences to obtain T segmentation probability maps; For each pixel in the segmentation probability map, the average of T prediction results is calculated as the final class probability of the corresponding pixel, and the final segmentation contour is generated by thresholding. For each pixel, the standard deviation of the prediction results of T times is calculated as the segmentation certainty of the corresponding pixel. The uncertainty heatmap is constructed based on the segmentation certainty of all pixels. The average segmentation certainty of all pixels within the segmented contour region is statistically analyzed and normalized to serve as the initial global quality assessment score.

3. The dynamic segmentation and functional assessment system for cardiac MRI images according to claim 2, characterized in that, Deep learning segmentation models based on the U-Net architecture include: Construct a U-Net deep learning segmentation model with an encoder-decoder structure; The input to the deep learning segmentation model is a single-frame image at a time point in a preprocessed MRI dynamic sequence; The output of the deep learning segmentation model is a segmentation probability map with multiple channels, which is the same size as the input image, and each channel corresponds to a heart structure category.

4. The dynamic segmentation and functional assessment system for cardiac MRI images according to claim 1, characterized in that, When the uncertainty heatmap and the initial global quality assessment score meet preset interaction conditions, an interactive correction process is initiated. This process receives correction instructions from the user and updates the segmentation contour and uncertainty heatmap in real time. Simultaneously, the contour matching degree of the segmentation contour is obtained, and the initial global quality assessment score is updated, including: Preset uncertainty threshold, quality assessment score threshold, and maximum number of highlighted areas X; When the initial global quality assessment score is lower than the quality assessment score threshold, or the number of consecutive regions exceeding the uncertainty threshold in the uncertainty heatmap is greater than X, it is determined that the preset interaction condition is met. In the uncertainty heatmap, the first X consecutive regions where the uncertainty value exceeds the uncertainty threshold are identified as highlighted areas; Based on the uncertainty-aware brush tool, it receives contour correction operations from users in either the highlighted or non-highlighted areas. When performing correction operations in the highlighted area, the model response weight is higher than that in the non-highlighted area, and correction operations in the non-highlighted area will trigger a confirmation prompt. Based on the user's contour correction operation, the segmentation contours of the current frame and adjacent time frames are updated in real time through the deformation model, and the uncertainty heatmap of the contour correction area is recalculated. Based on the difference between the corrected segmented contour and the segmented contour initially predicted by the model, the contour fit of the segmented contour is calculated, wherein the smaller the difference, the higher the contour fit. The average segmentation certainty is recalculated by combining the contour fit with the updated uncertainty heatmap, and the initial global quality assessment score is updated, wherein the contour fit is proportional to the updated global quality assessment score.

5. The dynamic segmentation and functional assessment system for cardiac MRI images according to claim 1, characterized in that, Extract the morphological and dynamic features of the updated segmented contours, including: Based on the ventricular cavity segmentation contour of each time frame, the ventricular cavity volume of each time frame is calculated, and the volume values ​​of all time frames are connected to generate a ventricular cavity volume curve. In the end-diastolic frame, the volume of the myocardium is calculated based on the epicardial and endocardial contours, and multiplied by the myocardial density to obtain the myocardial mass. In the end-diastolic and end-systolic frames, the ventricular wall is divided into several segments, and the wall thickness and its rate of change in each segment are calculated. The ventricular cavity volume curve, myocardial mass, ventricular wall thickness and their rate of change are used as the morphological and dynamic features of the updated segmentation profile.

6. The dynamic segmentation and functional assessment system for cardiac MRI images according to claim 5, characterized in that, The morphological and dynamic characteristics are input into a preset functional parameter calculation module to obtain N functional parameters of the heart, including: The N functional parameters include at least ejection fraction and stroke volume; From the ventricular volume curve, the maximum value is extracted as the end-diastolic volume, and the minimum value is extracted as the end-systolic volume. Based on the end-diastolic volume and end-systolic volume, the ejection fraction and stroke volume are calculated.

7. The dynamic segmentation and functional assessment system for cardiac MRI images according to claim 1, characterized in that, Based on the updated global quality assessment score, the calculated N functional parameters are reliably corrected to obtain the reliability ranges for the N functional parameters, including: Based on the uncertainty heatmap, M groups of possible profile variants conforming to the uncertainty distribution are generated by Monte Carlo sampling; For each set of contour variants, recalculate all of its N functional parameters; For each functional parameter, the corresponding M calculation results are used to form a statistical distribution. The P% quantile to (100-P)% quantile of the statistical distribution is taken as the initial confidence interval of the functional parameter, and N initial confidence intervals of N functional parameters are obtained, where P is greater than 0 and less than 100. The widths of the N initial confidence intervals are weighted and corrected using the updated global quality assessment score to obtain N confidence intervals for the N functional parameters. The higher the updated global quality assessment score, the narrower the corrected confidence interval width.

Images