An artificial intelligence-based mitral valve orifice area detection method

Abstract

This invention discloses an artificial intelligence-based method for mitral valve orifice area detection, relating to the field of intelligent ultrasound image analysis technology. The method includes: acquiring a continuous mitral valve ultrasound frame sequence, performing standardized preprocessing, and forming a sliding window frame sequence with a fixed length and step size; inputting the sliding window frame sequence into a phase embedding neural network to obtain a phase embedding vector sequence; calculating inter-frame similarity based on the phase embedding vector sequence; determining the cardiac cycle boundary based on the inter-frame similarity and extracting a complete cardiac cycle; calculating an open phase score within the complete cardiac cycle; selecting the most open phase frame of the mitral valve orifice based on the open phase score; and extracting a short temporal segment centered on the most open phase frame of the mitral valve orifice. This invention achieves consistent selection of the most open phase frame of the mitral valve orifice under different individual and imaging quality conditions through automatic and stable localization of key phases of the cardiac cycle, thereby reducing the differences in manual frame selection and fluctuations in repeated measurements.

Description

Technical Field

[0001] This invention relates to the field of intelligent ultrasound image analysis technology, and in particular to an artificial intelligence-based method for detecting the mitral valve orifice area. Background Technology

[0002] With the increasing use of echocardiography in valvular disease assessment, the quantitative detection of mitral valve orifice area has gradually shifted from manual recording and semi-automatic measurement to intelligent analysis. In recent years, the introduction of deep learning and representation learning has enabled algorithms to extract more discriminative anatomical and temporal features under complex echo textures, low-contrast boundaries, and different imaging windows, and to capture the dynamic changes of valve opening and closing in multi-frame sequences, achieving automated identification of key phases and boundaries. At the same time, the improvement of computing platforms and data annotation systems has promoted the model from offline scientific research validation to clinical decision support applications, forming an automated measurement process in collaboration with workstation / cloud systems, and improving the consistency and reproducibility of mitral valve orifice area assessment.

[0003] Existing methods have shortcomings. The determination of key phase frames lacks characterization of the phase structure of the complete cardiac cycle, and is prone to cycle boundary shifts and instability in the selection of the maximum open phase under time-series fluctuations, resulting in insufficient sensitivity and repeatability of area measurement to frame selection. In addition, valve orifice boundary acquisition is prone to contour breaks or non-closure in complex echo backgrounds, and the consistency constraints between the boundary coordinates and the original coordinate system after ROI clipping and scaling transformation are insufficient, which can easily lead to area conversion errors and superimposed label offsets, reducing the reliability and interpretability of the results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based method for detecting the mitral valve orifice area to solve the problems of phase frame instability and back-mapping deviation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an artificial intelligence-based method for detecting the mitral valve orifice area, which includes: acquiring a continuous frame sequence of mitral valve ultrasound, performing standardized preprocessing, forming a sliding window frame sequence with a fixed length and step size, and inputting the sliding window frame sequence into a phase embedding neural network to obtain a phase embedding vector sequence.

[0008] Inter-frame similarity is calculated based on phase embedding vector sequences. The cardiac cycle boundary is determined based on the inter-frame similarity, and the complete cardiac cycle is extracted. The open phase score is calculated within the complete cardiac cycle. The most open phase frame of the valve orifice is selected based on the open phase score, and a short time segment centered on the most open phase frame of the valve orifice is extracted. Valve orifice ROI localization and cropping are performed on the short time segment to obtain the valve orifice ROI frame sequence and back-mapping information. The ROI frame sequence is input into a neural implicit boundary network, which outputs an implicit probability function and extracts the topologically closed mitral valve orifice boundary through isosurface extraction. The valve orifice area of ​​each frame in the short time segment is calculated based on the topologically closed mitral valve orifice boundary, and the mitral valve orifice area value is obtained based on the valve orifice area of ​​each frame. The mitral valve orifice area value is superimposed with the closed boundary of the corresponding phase frame according to the back-mapping information to form a boundary superposition map.

[0009] As a preferred embodiment of the artificial intelligence-based mitral valve orifice area detection method of the present invention, the following steps are taken: The acquired mitral valve ultrasound continuous frame sequence is subjected to standardized preprocessing, and a sliding window frame sequence is formed according to a fixed length and step size.

[0010] Acquire a continuous frame sequence of mitral valve ultrasound data, and number and sort the frames according to the acquisition order;

[0011] The mitral valve ultrasound continuous frame sequence was subjected to frame-by-frame normalization preprocessing to obtain the normalized mitral valve ultrasound continuous frame sequence.

[0012] The standardized mitral valve ultrasound continuous frame sequence is divided into sliding window frame sequences by using a fixed window length and step size.

[0013] As a preferred embodiment of the artificial intelligence-based mitral valve orifice area detection method of the present invention, the specific steps of inputting the sliding window frame sequence into the phase embedding neural network to obtain the phase embedding vector sequence are as follows.

[0014] A phase embedding neural network was constructed using ResNet and a long short-term memory neural network, and the phase embedding neural network was pre-trained with phase self-supervised training.

[0015] The sliding window frame sequence is input into the phase embedding neural network in chronological order. ResNet extracts features from the sliding window frame sequence frame by frame. The long short-term memory neural network outputs the phase embedding vector at each time step, forming a phase embedding vector sequence that corresponds to each frame of the sliding window frame sequence.

[0016] As a preferred embodiment of the artificial intelligence-based mitral valve orifice area detection method of the present invention, the steps of calculating inter-frame similarity based on phase embedding vector sequences, determining cardiac cycle boundaries based on inter-frame similarity, and truncating the complete cardiac cycle are as follows.

[0017] The phase embedding vector of each frame in the phase embedding vector sequence is normalized to obtain the normalized phase embedding vector sequence.

[0018] Inter-frame similarity is calculated based on the normalized phase embedding vector sequence to form an inter-frame similarity matrix;

[0019] The cardiac cycle length is determined based on the inter-frame similarity matrix, and the cardiac cycle boundary is located based on the cardiac cycle length to extract the complete cardiac cycle.

[0020] As a preferred embodiment of the artificial intelligence-based mitral valve orifice area detection method of the present invention, the steps of calculating the open phase score within a complete cardiac cycle, selecting the most open phase frame of the valve orifice based on the open phase score, and extracting a short time segment centered on the most open phase frame of the valve orifice are as follows.

[0021] The average value of the phase embedding vector of the starting frame and the phase embedding vector of the ending frame of the complete cardiac cycle boundary is calculated and normalized to generate a reference phase embedding vector.

[0022] Calculate the cosine similarity between the phase embedding vector and the reference phase embedding vector, and invert the cosine similarity to obtain the open phase score;

[0023] The frame corresponding to the maximum open phase score is selected as the most open phase frame of the valve orifice, and a short time segment is extracted from the complete cardiac cycle with the most open phase frame of the valve orifice as the center.

[0024] As a preferred embodiment of the artificial intelligence-based mitral valve orifice area detection method of the present invention, the steps of locating and cropping the orifice ROI of short time segments to obtain the orifice ROI frame sequence and back-mapping information are as follows.

[0025] Select the center frame of a short temporal segment, perform contrast enhancement and threshold segmentation, and combine morphological opening and closing operations to obtain a low-echo region cleanup mask;

[0026] Select the target connected region of the valve orifice ROI from the low echo region purification mask and calculate the bounding rectangle. Generate the valve orifice ROI clipping box and record the clipping parameters to generate back-mapping information.

[0027] The short temporal segment is cropped frame by frame using the ROI clipping box to form the ROI frame sequence. The size of the ROI frame sequence is uniformly processed frame by frame, and the clipping and scaling parameters are written back to the mapping information.

[0028] As a preferred embodiment of the artificial intelligence-based mitral valve orifice area detection method of the present invention, the specific steps of inputting the ROI frame sequence into a neural implicit boundary network, outputting an implicit probability function, and extracting the topologically closed mitral valve orifice boundary through isosurface are as follows.

[0029] A neural implicit boundary network is constructed by combining an image feature extraction network and an implicit probability function output network.

[0030] The valve orifice ROI frame sequence is input into the neural implicit boundary network in chronological order. The image feature extraction network extracts frame-by-frame features from the valve orifice ROI frame sequence, and the implicit probability function output network outputs implicit probability functions from the frame-by-frame features.

[0031] The mitral valve orifice boundary is obtained by performing isosurface extraction on the implicit probability function frame by frame, and the mitral valve orifice boundary is mapped back to a short time segment based on the back mapping information.

[0032] As a preferred embodiment of the artificial intelligence-based mitral valve orifice area detection method of the present invention, the specific steps for obtaining the topologically closed mitral valve orifice boundary by performing isosurface extraction on the implicit probability function frame by frame are as follows.

[0033] When multiple closed boundaries are obtained by isosurface extraction, region filling is performed on each of the multiple closed boundaries and the number of pixels inside each closed boundary is calculated.

[0034] The closed boundary with the largest number of internal pixels is selected as the mitral valve orifice boundary of topological closure.

[0035] As a preferred embodiment of the artificial intelligence-based mitral valve orifice area detection method of the present invention, the specific steps of calculating the orifice area of ​​each frame in a short time segment based on the topologically closed mitral valve orifice boundary, and obtaining the mitral valve orifice area value based on the orifice area of ​​each frame, are as follows.

[0036] Obtain the corresponding ultrasound image calibration information from short time segments, and calculate the conversion factor of the actual area of ​​a single pixel based on the ultrasound image calibration information;

[0037] The region of the mitral valve orifice boundary of the short temporal segment topological closure is filled frame by frame, and the number of internal pixels is counted. The orifice area value of each frame is calculated by combining the actual area conversion factor of a single pixel.

[0038] The largest mitral valve area value is selected from the valve area values ​​of each frame in the short time segment as the mitral valve area value, and the mitral valve valve boundary corresponding to the phase frame and the topological closure of the corresponding phase frame is determined.

[0039] As a preferred embodiment of the artificial intelligence-based mitral valve orifice area detection method of the present invention, the specific steps for forming a boundary overlay map by superimposing the mitral valve orifice area value with the closed boundary of the corresponding phase frame using back-mapping information are as follows.

[0040] Based on the back-mapping information, the mitral valve orifice boundary of the topological closure of the corresponding phase frame is mapped to the original coordinates of the corresponding phase frame;

[0041] The boundary of the mitral valve orifice is overlaid on the corresponding phase frame image and the area value of the mitral valve orifice is marked to generate a boundary overlay map.

[0042] The beneficial effects of this invention are as follows: By automatically and stably locating the key phases of the cardiac cycle, consistent selection of the most open phase frame of the mitral valve is achieved under different individuals and imaging quality conditions, thereby reducing the differences in manual frame selection and fluctuations in repeated measurements, and improving the stability and repeatability of mitral valve orifice area assessment; by reliably closing the valve orifice boundary to obtain and output superimposed and verifiable result images, the area calculation results and image evidence are presented synchronously, thereby reducing the risk of misjudgment caused by incomplete boundaries and misaligned annotations, and improving measurement accuracy, interpretability and clinical efficiency. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of an artificial intelligence-based method for detecting the mitral valve orifice area.

[0045] Figure 2 This is a flowchart for capturing the complete cardiac cycle.

[0046] Figure 3 A flowchart for determining the final closed boundary.

[0047] Figure 4 This is a flowchart for generating the boundary overlay map. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0051] Reference Figures 1-4 This is one embodiment of the present invention, which provides an artificial intelligence-based method for detecting the mitral valve orifice area, comprising the following steps:

[0052] S1. Acquire continuous frame sequences of mitral valve ultrasound, perform standardized preprocessing, and form a sliding window frame sequence with a fixed length and step size. Input the sliding window frame sequence into a phase embedding neural network to obtain a phase embedding vector sequence.

[0053] S1.1 Acquire a continuous frame sequence of mitral valve ultrasound, and number and sort the frames according to the acquisition order. Perform standardized preprocessing on each frame of the continuous frame sequence of mitral valve ultrasound to obtain a standardized continuous frame sequence of mitral valve ultrasound.

[0054] It should be noted that the mitral valve region is continuously scanned by an ultrasound probe, and continuous imaging acquisition is initiated on the ultrasound equipment to continuously acquire and save the mitral valve ultrasound continuous frame sequence; each frame of the mitral valve ultrasound continuous frame sequence is read sequentially according to the acquisition time and written with a frame number; the mitral valve ultrasound continuous frame sequence is sorted according to the frame number to obtain the sorted mitral valve ultrasound continuous frame sequence; the effective imaging area of ​​ultrasound fan-forming is extracted frame by frame in the sorted mitral valve ultrasound continuous frame sequence, and black borders and character annotation areas are removed.

[0055] The effective imaging region is scaled by interpolation to the input image size required by the phase embedding neural network and the neural implicit boundary network to complete the size unification process. The effective imaging region after size unification is linearly normalized by the minimum and maximum gray values, mapping the gray range to 0 to 1 to complete the gray intensity normalization process. The effective imaging region after gray intensity normalization is subjected to median filtering for noise reduction. The effective imaging regions after frame-by-frame size unification, normalization, and noise reduction are recombined in frame number order to form a standardized mitral valve ultrasound continuous frame sequence.

[0056] S1.2. The standardized mitral valve ultrasound continuous frame sequence is slidably segmented according to a fixed window length and step size to form a sliding window frame sequence.

[0057] It should be noted that the fixed window length is the number of input frames required for training and inference of the phase embedding neural network, and the fixed step size is the number of frames that the starting frame of the adjacent sliding window frame sequence moves forward along the frame number direction. Starting from the first frame of the standardized mitral valve ultrasound continuous frame sequence, continuous frames are truncated according to the fixed window length to form the first segment of the sliding window frame sequence. The starting frame of the window is moved forward according to the fixed step size, and continuous frames are truncated repeatedly to form subsequent sliding window frame sequences until the starting frame of the window moves to the end of the standardized mitral valve ultrasound continuous frame sequence. When the number of continuous frames at the end of the window is less than the fixed window length, the last frame at the end is used to repeatedly fill in the missing frames in the order of frame number to ensure that each segment of the sliding window frame sequence meets the fixed window length. The segments of the sliding window frame sequence are arranged in ascending order of the starting frame number to form the sliding window frame sequence.

[0058] S1.3 Construct a phase embedding neural network using ResNet and Long Short-Term Memory neural network, and perform phase self-supervised pre-training on the phase embedding neural network.

[0059] It should be noted that ResNet is used as the feature extraction network and connected to the Long Short-Term Memory Neural Network in chronological order to form a phase embedding neural network. The phase embedding neural network outputs phase embedding vectors frame by frame from the sliding window frame sequence. Sliding window frame sequence pairs are extracted from the sliding window frame sequence in chronological order as pre-training samples. For each set of pre-training samples, sliding window frame sequence pairs with adjacent time positions within the same cardiac cycle are constructed as positive sample pairs, and sliding window frame sequence pairs with different cardiac cycles or long time intervals are constructed as negative sample pairs. The positive sample pairs and negative sample pairs are input into the phase embedding neural network in sequence to obtain the corresponding phase embedding vector pairs. The phase self-supervised pre-training loss is calculated based on the similarity of the phase embedding vector pairs, and the parameters of the phase embedding neural network are updated through backpropagation. The process is iterated until the phase self-supervised pre-training loss converges, resulting in a phase embedding neural network that has completed phase self-supervised pre-training.

[0060] During phase self-supervised pre-training, the phase self-supervised pre-training loss is recorded once after each round of parameter update iteration, and the moving average of the phase self-supervised pre-training loss is calculated; when the decrease of the moving average is less than the convergence threshold, the phase self-supervised pre-training loss is determined to have converged.

[0061] Suppose that the phase embedding neural network is used for the first... The phase embedding vector output by the sliding window frame sequence is: The phase embedding vector output for the positive sample sliding window frame sequence is: , for the The output phase embedding vector of the negative sample sliding window frame sequence is: First, normalization is performed; the expression is:

[0062] ;

[0063] in, Represents the phase embedding vector Create a vector with normalized length; Indicates the first Phase embedding vectors output from a sequence of sliding window frames; This represents the index of the sliding window frame sequence.

[0064] Define similarity as the number of clicks on the normalized vector, then the first... The expression for the phase self-supervised pre-training loss of a sliding window frame sequence is:

[0065] ;

[0066] in, Indicates the first Phase self-supervised pre-training loss for a sequence of sliding window frames; This represents a similarity function, which takes two normalized vectors as input and outputs a similarity scalar. This represents the phase embedding vector output from the normalized positive sample sliding window frame sequence. Represents the normalized i-th Output phase embedding vectors from a sequence of negative sample sliding window frames; This represents the negative sample index, with a value ranging from 1 to... ; This represents the total number of negative samples; This represents a temperature coefficient used to scale the numerical scale of similarity before it enters the exponential function, making the distribution of contrastive learning sharper or smoother.

[0067] The expression for calculating the moving average of the phase self-supervised pre-training loss is as follows:

[0068] ;

[0069] in, Represents the moving average; This represents the total number of frames in the sliding window sequence.

[0070] The convergence threshold is taken as the minimum resolvable change of the phase self-supervised pre-training loss moving average within the observation window, and the minimum resolvable change is determined by the maximum fluctuation difference of the phase self-supervised pre-training loss moving average sequence within the observation window.

[0071] The observation window is an interval consisting of several consecutive rounds of parameter update iterations during the phase self-supervised pre-training process. It is used to statistically analyze the changing trend of the phase self-supervised pre-training loss within the interval and to determine whether the phase self-supervised pre-training loss has converged.

[0072] S1.4 Input the sliding window frame sequence into the phase embedding neural network in chronological order. Extract features from the sliding window frame sequence frame by frame using ResNet. The long short-term memory neural network performs recursive calculation and outputs the phase embedding vector corresponding to each sliding window frame sequence, forming a phase embedding vector sequence.

[0073] It should be noted that, within each sliding window frame sequence, the corresponding effective imaging region is extracted frame by frame in order of frame number. The effective imaging regions frame by frame are then fed into the ResNet of the phase embedding neural network. The ResNet performs forward computation on the effective imaging regions frame by frame to obtain the feature vectors frame by frame. The feature vectors frame by frame are arranged into a feature sequence in order of frame number and then fed into the long short-term memory neural network of the phase embedding neural network. The long short-term memory neural network performs recursive computation on the feature sequence step by step and outputs the phase embedding vector of the corresponding sliding window frame sequence at the last time step. The phase embedding vectors obtained from each sliding window frame sequence are arranged in ascending order of the starting frame number of the window to form a phase embedding vector sequence.

[0074] S2. Calculate the inter-frame similarity based on the phase embedding vector sequence, determine the cardiac cycle boundary based on the inter-frame similarity and extract the complete cardiac cycle, calculate the open phase score within the complete cardiac cycle, select the most open phase frame of the valve orifice based on the open phase score, and extract a short time segment centered on the most open phase frame of the valve orifice.

[0075] S2.1 Normalize the phase embedding vector of each frame in the phase embedding vector sequence to obtain the normalized phase embedding vector sequence.

[0076] It should be noted that the vector magnitude of the phase embedding vector for each frame is calculated using the following expression:

[0077] ;

[0078] in, Indicates the first The vector magnitude of each phase embedding vector; This represents the index of the phase embedding vector in the phase embedding vector sequence; The dimension index represents the phase embedding vector, and its value ranges from 1 to... ; This represents the total dimension of the phase embedding vector, i.e., the number of components contained in the phase embedding vector; Indicates the first Phase embedding vectors The Dimensional value.

[0079] Calculate the ratio of the vector magnitude to the components of each dimension of the phase embedding vector, so that the vector magnitude of the phase embedding vector in each frame is uniformly 1, and complete the scaling in one step. Then, recombine the scaled phase embedding vectors in ascending order of the window start frame number of the sliding window frame sequence to obtain the normalized phase embedding vector sequence.

[0080] S2.2 Calculate the inter-frame similarity based on the normalized phase embedding vector sequence to form an inter-frame similarity matrix.

[0081] It should be noted that the number of normalized phase embedding vectors contained in the normalized phase embedding vector sequence is denoted as . , establish OK The two-dimensional matrix structure of columns serves as the inter-frame similarity matrix. The row and column indices of the inter-frame similarity matrix correspond to the indices of the normalized phase embedding vector sequence. All elements of the inter-frame similarity matrix are initially assigned zero values, and the diagonal elements of the inter-frame similarity matrix are assigned 1 values ​​as the self-similarity of the same index vector.

[0082] After completing the size establishment and initial assignment of the inter-frame similarity matrix, the index is... and The normalized phase embedding vector is used to calculate the vector dot product, and the dot product result is assigned to the first element of the inter-frame similarity matrix. Line number Column elements and the first Line number The column elements are used to fill in the inter-frame similarity matrix.

[0083] The expression for calculating the dot product of vectors is:

[0084] ;

[0085] in, Indicates that the index is and The inter-frame similarity values ​​between the normalized phase embedding vectors; and These represent two indices of the normalized phase embedding vector sequence; The index in the normalized phase embedding vector sequence is... The normalized phase embedding vector; The index in the normalized phase embedding vector sequence is... The normalized phase embedding vector; Indicates the transpose symbol; This represents the dot product operation of two vectors.

[0086] S2.3 Determine the cardiac cycle length based on the inter-frame similarity matrix, locate the cardiac cycle boundary based on the cardiac cycle length, and extract the complete cardiac cycle.

[0087] It should be noted that, based on the inter-frame similarity matrix, a similarity sequence parallel to the main diagonal is constructed along the main diagonal with different positive integer index intervals. The average similarity of the similarity sequence under each index interval is calculated and the average similarity of each index interval is compared. The index interval with the largest average similarity is determined as the cardiac cycle length.

[0088] All index positions in the phase embedding vector sequence that satisfy the condition that "the starting index is still within the sequence index range after being offset by one cardiac cycle length along the sequence direction" are selected as starting index candidates. For each starting index candidate, the similarity between the starting index and the offset index position in the inter-frame similarity matrix is ​​calculated as a consistency score. The starting index candidate with the highest consistency score is selected as the starting index of the cardiac cycle boundary, and the ending index of the cardiac cycle boundary is determined as the index position after the starting index of the cardiac cycle boundary is offset by one cardiac cycle length along the sequence direction. Based on the starting index and ending index of the cardiac cycle boundary, the corresponding intervals are truncated in the phase embedding vector sequence to form the phase embedding vector subsequence corresponding to the complete cardiac cycle. The sliding window frame sequence with the same index as the phase embedding vector subsequence is truncated in the sliding window frame sequence to obtain the complete cardiac cycle.

[0089] S2.4 Calculate the average value of the phase embedding vector of the starting frame and the phase embedding vector of the ending frame of the complete cardiac cycle boundary, and normalize the average value to generate a reference phase embedding vector.

[0090] It should be noted that, in the phase embedding vector subsequence corresponding to the complete cardiac cycle, the phase embedding vector of the starting frame of the complete cardiac cycle boundary and the phase embedding vector of the ending frame of the complete cardiac cycle boundary are determined. The phase embedding vector of the starting frame of the complete cardiac cycle boundary and the phase embedding vector of the ending frame of the complete cardiac cycle boundary are summed dimension by dimension and the ratio with 2 is calculated to obtain the average phase embedding vector.

[0091] The vector magnitude of the average phase embedding vector is calculated, and the ratio of the vector magnitude to the average phase embedding vector dimension by dimension is calculated to obtain the normalized average phase embedding vector, which is the reference phase embedding vector. The vector magnitude of the reference phase embedding vector is 1.

[0092] S2.5 Calculate the cosine similarity between the phase embedding vector and the reference phase embedding vector, and invert the cosine similarity to obtain the open phase score. Select the frame corresponding to the maximum open phase score as the most open phase frame of the valve orifice, and extract short time segments from the complete cardiac cycle with the most open phase frame of the valve orifice as the center.

[0093] It should be noted that within the phase embedding vector subsequence corresponding to the complete cardiac cycle, the reference phase embedding vector is used as the similarity reference. For the phase embedding vector at each index position in the phase embedding vector subsequence, the cosine similarity with the reference phase embedding vector is calculated, and the negative value of the cosine similarity is used as the open phase score.

[0094] The open phase scores at each index position corresponding to the complete cardiac cycle are compared, and the index position corresponding to the maximum open phase score is determined as the index of the most open phase frame of the valve orifice. The sliding window frame sequence and the phase embedding vector subsequence corresponding to the complete cardiac cycle are indexed. Based on the index of the most open phase frame of the valve orifice, the most open phase frame of the valve orifice that is consistent with the index position is determined in the sliding window frame sequence corresponding to the complete cardiac cycle. With the index of the most open phase frame of the valve orifice as the center, a fixed length of continuous frames is extracted from the sliding window frame sequence corresponding to the complete cardiac cycle to form a short time segment.

[0095] The fixed length of the short temporal segment is taken as the length of the valve ROI frame sequence used during the training and inference of the neural implicit boundary network, so that the valve ROI frame sequence obtained by pruning the short temporal segment can directly meet the temporal input requirements of the neural implicit boundary network. The fixed length is taken as an odd number to ensure that the most open phase frame of the valve is located at the center of the short temporal segment. When the most open phase frame of the valve is close to the boundary of the complete cardiac cycle, resulting in insufficient frames on both sides, the missing frames are filled by repeating the boundary frames to keep the short temporal segment length constant.

[0096] It should also be noted that existing technologies often rely on manual frame selection or using simple indicators to directly find the largest opening in the entire sequence, which is easily affected by noise, displacement, and heart rate changes, resulting in unstable cycle boundaries, drift of the largest opening frame, and poor repeatability. This scheme uses a similarity matrix to stably determine the length and boundary of the cardiac cycle, and locks the most open phase frame of the valve orifice through open phase scoring within the complete cycle. Then, it extracts a short time segment of fixed length, thereby improving the consistency and anti-interference ability of the most open phase localization, reducing the risk of human differences and misselection of frames, improving the stability and comparability of area measurement, and reducing the scope of subsequent processing to reduce the amount of computation.

[0097] S3. Perform valve orifice ROI localization and cropping on short time segments to obtain valve orifice ROI frame sequences and back-mapping information. Input the ROI frame sequences into a neural implicit boundary network, output implicit probability functions, and extract the topologically closed mitral valve orifice boundary through isosurface extraction.

[0098] S3.1 Select the center frame of the short time segment, perform contrast enhancement and threshold segmentation, and combine morphological opening and morphological closing operations to obtain a low-echo region cleanup mask.

[0099] It should be noted that, within a short time segment, the center frame located in the middle of the time segment is selected according to the frame number. Gray-level stretching and histogram equalization are performed on the effective imaging area of ​​the center frame to improve gray-level contrast, resulting in an enhanced center frame. Gray-level histograms are calculated on the enhanced center frame, and the Otsu method is used to traverse candidate thresholds. For each candidate threshold, pixels are temporarily divided into low-gray-level and high-gray-level classes based on whether the gray level is less than the candidate threshold or not less than the candidate threshold. The difference in gray-level distribution between the two classes is calculated, and the candidate threshold with the largest difference between the two classes is selected as the segmentation threshold. Pixels are then formally reclassified into low-gray-level and high-gray-level classes according to the segmentation threshold to form a low-echo initial mask.

[0100] On the low-echo initial mask, identify all connected components and count the number of pixels within each component. Represent each connected component as a circular region of equal area, and convert the number of pixels within each component to the equivalent circle diameter. The expression is as follows:

[0101] ;

[0102] in, Indicates the first The equivalent circle diameter of a connected domain; Indicates the first The number of internal pixels of a connected component; This represents the connected component index, used to distinguish different connected components.

[0103] The median of the equivalent circle diameter is taken as the diameter of the circular structural element, and the circular structural element is determined. A morphological opening operation is performed on the low-echo initial mask. The morphological opening operation removes small connected regions and smooths the edges by "erosion first and then dilation". Then, a morphological closing operation is performed on the opening operation result. The morphological closing operation fills the holes and connects the broken regions by "dilation first and then erosion", thus obtaining a low-echo region purification mask.

[0104] S3.2 Select the target connected region of the valve opening ROI from the low echo region purification mask and calculate the bounding rectangle. Generate the valve opening ROI clipping box and record the clipping parameters to generate back-mapping information.

[0105] It should be noted that, in the low-echo area cleanup mask, all connected components are identified and the number of internal pixels is calculated for each connected component. The connected component with the largest number of internal pixels is determined as the target connected component of the ROI. The coordinates of all foreground pixels are extracted from the target connected component of the ROI, and the minimum, maximum, minimum, and maximum values ​​of the horizontal, vertical, and lateral coordinates are calculated respectively. The minimum, maximum, minimum, and maximum values ​​of the horizontal, vertical, and lateral coordinates are used to determine the bounding rectangle that can completely enclose the target connected component of the ROI. The coordinates of the upper left and lower right corners of the bounding rectangle, as well as the width and height of the bounding rectangle, are determined as the clipping box of the ROI. Back-mapping information is generated using the coordinates of the upper left and lower right corners, the width, and the height of the clipping box of the ROI, so that the back-mapping information corresponds to the clipping box of the ROI.

[0106] S3.3. Use the ROI clipping box to clip short time segments frame by frame to form a ROI frame sequence. Perform size unification processing on each frame of the ROI frame sequence and write the clipping and scaling parameters into the back-mapping information.

[0107] It should be noted that, based on the valve ROI cropping box, in each frame of a short temporal segment, a sub-region of the image is cropped by the rectangular area defined by the upper left and lower right corner coordinates of the valve ROI cropping box. These sub-regions are arranged sequentially according to the frame number of the short temporal segment to form a valve ROI frame sequence. For each frame of the valve ROI frame sequence, interpolation scaling is used to transform the image size to the input image size required by the neural implicit boundary network to complete the size unification process. The back-mapping information is then updated so that the back-mapping information includes the cropping offset corresponding to the valve ROI cropping box and the scaling ratio parameter corresponding to the size unification process. This ensures that the back-mapping information simultaneously includes the cropping offset and the scaling ratio parameter and corresponds to the valve ROI frame sequence.

[0108] S3.4. A neural implicit boundary network is constructed by combining an image feature extraction network and an implicit probability function output network. The valve ROI frame sequence is input into the neural implicit boundary network in chronological order. The image feature extraction network extracts frame-by-frame features from the valve ROI frame sequence, and the implicit probability function output network outputs implicit probability functions from the frame-by-frame features.

[0109] It should be noted that the image feature extraction network and the implicit probability function output network are cascaded together to form a neural implicit boundary network. The image feature extraction network is used to receive a single-frame ROI image of the mitral valve orifice and output the corresponding feature tensor. The image feature extraction network can use a ResNet backbone structure to implement feature extraction. The implicit probability function output network is used to receive the feature tensor and generate probability values ​​on the pixel coordinate grid. The implicit probability function output network can be implemented using a multilayer perceptron. Specifically, a pixel coordinate grid with the same size as the image is established for the single-frame ROI image. After normalizing each pixel coordinate, the corresponding position feature vector is concatenated with the feature tensor after interpolation and scaling to the same size as the image. The implicit probability function output network outputs a scalar probability value between 0 and 1 for each pixel coordinate position. The scalar probability value represents the probability that the pixel belongs to the inside of the mitral valve orifice. The scalar probability values ​​of all pixels together constitute the implicit probability function corresponding to the single-frame ROI image of the mitral valve orifice.

[0110] The acquired and manually annotated mitral valve ultrasound continuous frame sequences were processed through standardized preprocessing, sliding window segmentation, phase embedding vector generation, cardiac cycle boundary determination, short time segment extraction, and valve orifice ROI localization and cropping to obtain valve orifice ROI frame sequences for parameter optimization. For each frame of the valve orifice ROI frame sequences used for parameter optimization, the mitral valve orifice boundaries were manually delineated, and corresponding binary annotation masks were generated. The valve orifice ROI frame sequences used for parameter optimization were paired with the binary annotation masks to form a training sample set, which was then divided into a training set and a validation set. During parameter optimization, the valve orifice ROI frame sequences in the training set were processed by a neural implicit boundary network to obtain frame-by-frame implicit probability functions. The frame-by-frame implicit probability functions and the corresponding binary annotation masks in the training set were used to calculate the supervision error. The supervision error was used as the training loss to iteratively update the image feature extraction network and the implicit probability function output network parameters.

[0111] Let the number of frames contained in a short temporal segment be The pixel set of a single-frame lobe-mouth ROI image is pixel coordinates are The implicit probability function in the th Frame pixels The output probability at is The binary annotation mask in the first... Frame pixels The label at the location is Then the supervision error can be taken as the pixel-wise binary cross-entropy, expressed as follows:

[0112] ;

[0113] in, Indicates the monitoring error; Indicates the number of elements in the pixel set; This represents the frame number index within a short time segment, with values ​​ranging from 1 to... ; Represents pixel coordinates.

[0114] After each round of parameter updates, the validation loss and validation set segmentation index are calculated using the ROI frame sequence of the validation set and the corresponding binary labeled mask of the validation set. The tolerance threshold is taken as the maximum fluctuation difference of the moving average of the validation loss within the observation window, and the segmentation index improvement threshold is taken as the maximum fluctuation difference of the moving average of the validation set segmentation index within the observation window. When the decrease of the validation loss is less than the tolerance threshold and the increase of the validation set segmentation index is less than the segmentation index improvement threshold in several consecutive iterations, the parameter updates are stopped and the parameters are determined.

[0115] The expression for calculating the validation set splitting metric is as follows:

[0116] ;

[0117] in, This represents the validation set segmentation metric; a larger value indicates a better segmentation.

[0118] After the parameters are determined, the ROI frame sequence is input into the image feature extraction network in chronological order to obtain frame-by-frame feature tensors. The implicit probability function output network outputs the implicit probability function corresponding to the frame-by-frame ROI frame sequence to be detected on the pixel coordinate grid.

[0119] The expression for the implicit probability function is:

[0120] ;

[0121] ;

[0122] in, Indicates the first Frame pixels The implicit probability function at that location; Represents the Sigmoid function; This represents the forward mapping of the implicit probability function output network; Indicates the first Frame pixels The eigenvector at that location; Represents pixels The normalized coordinate vector at that location; Indicates the first Frame pixels The score at a given point indicates the tendency of the current pixel to belong to the inside of the lobe opening.

[0123] S3.5. Perform isosurface extraction on the implicit probability function frame by frame to obtain the topologically closed mitral valve orifice boundary, and map the mitral valve orifice boundary back to the short time segment based on the back mapping information.

[0124] It should be noted that, for the implicit probability function corresponding to each frame in the ROI frame sequence, the pixel-by-pixel output probability values ​​of the implicit probability function on the pixel coordinate grid are arranged into a two-dimensional matrix according to the pixel coordinate order, resulting in a probability map with the same size as the ROI frame sequence image. During the parameter determination stage, a set of candidate thresholds is generated at equal intervals between zero and one, and substituted one by one. The probability map is then divided into binary masks, and the validation set segmentation index is calculated with the corresponding binary labeled mask of the validation set. The candidate threshold that maximizes the validation set segmentation index is selected as the equivalent threshold. The probability threshold is determined and fixed along with the determined parameters. The probability map is used to extract the corresponding contour lines according to the equivalent probability threshold using the moving block algorithm to obtain a set of boundary points connected in sequence. The boundary point set is connected end to end to form the topologically closed mitral valve orifice boundary. Based on the scaling parameter and clipping offset of the corresponding frame in the back-mapping information, the coordinates of each boundary point of the mitral valve orifice boundary are first inversely scaled according to the scaling parameter, and then inversely translated according to the clipping offset, so as to obtain the topologically closed mitral valve orifice boundary corresponding to each frame of the short time segment.

[0125] It should also be noted that existing technologies mostly calculate the area by directly segmenting frame by frame or manually delineating regions on mitral valve ultrasound images. Due to the influence of speckle noise, echo heterogeneity, and blurred boundaries, the boundaries are prone to non-closure or positional drift, resulting in large fluctuations in the area results and insufficient repeatability and consistency. This solution locates and trims the valve orifice ROI within short time segments to form a frame sequence with consistent positions that can be returned to the original coordinates. Then, it provides the probability of the valve orifice inside each frame and extracts the closed contour to obtain the topologically closed mitral valve orifice boundary and maps it back to the short time segment. This improves the integrity and stability of the boundary, reduces the area error caused by boundary breakage and inter-frame jitter, improves the consistency of results under different operators and different acquisition conditions, and provides a more intuitive and verifiable boundary basis for subsequent area calculation and boundary overlay map.

[0126] S4. Calculate the mitral valve area of ​​each frame in the short time segment based on the topologically closed mitral valve valve boundary, and obtain the mitral valve valve area value based on the valve area of ​​each frame. Superimpose the mitral valve valve area value with the closed boundary of the corresponding phase frame according to the back-mapping information to form a boundary superposition map.

[0127] S4.1 Obtain the corresponding ultrasound image calibration information from the short time sequence segment, and calculate the conversion factor of the actual area of ​​a single pixel based on the ultrasound image calibration information. Fill the region of the mitral valve orifice boundary of the topological closure of the short time sequence segment frame by frame and count the number of internal pixels. Calculate the valve orifice area value of each frame in combination with the conversion factor of the actual area of ​​a single pixel.

[0128] It should be noted that the corresponding ultrasound image calibration information is obtained from the short time-series segments, and the conversion factor of the actual area of ​​a single pixel is calculated based on the ultrasound image calibration information. The mitral valve orifice boundary of the topological closure of the short time-series segment is filled frame by frame, and the number of internal pixels is counted. The valve orifice area value of each frame is calculated in combination with the conversion factor of the actual area of ​​a single pixel. The horizontal pixel spacing and vertical pixel spacing are read and parsed from the spatial resolution field in the ultrasound image file header corresponding to each frame of the short time-series segment or from the calibration parameters output by the ultrasound equipment. The conversion factor of the actual area of ​​a single pixel in the current frame is calculated based on the product of the horizontal pixel spacing and the vertical pixel spacing. When the pixel spacing of each frame in the short time-series segment is consistent, a unified horizontal pixel spacing and vertical pixel spacing can be used to calculate a unified conversion factor of the actual area of ​​a single pixel. The horizontal pixel spacing, vertical pixel spacing, and conversion factor of the actual area of ​​a single pixel are then correlated with the frame number to form a conversion factor index for subsequent frame-by-frame retrieval.

[0129] Using the topologically closed mitral valve orifice boundary of the current frame as input, the topologically closed mitral valve orifice boundary is rasterized into a closed contour on a blank binary mask of the same size as the current frame image. Region filling is then performed inside the closed contour to obtain the binary filling mask of the valve orifice region of the current frame. The binary filling mask is used as the pixel statistics input to count the number of foreground pixels to obtain the number of pixels inside the valve orifice of the current frame. The conversion coefficient of the actual area of ​​a single pixel corresponding to the current frame is retrieved from the conversion coefficient index according to the frame number. The product of the number of pixels inside the valve orifice of the current frame and the conversion coefficient of the actual area of ​​a single pixel is used as the valve orifice area value of the current frame. The valve orifice area value corresponding to the frame number is output for all frames of the short time segment.

[0130] S4.2 Select the largest mitral valve area value from the valve area values ​​of each frame in the short time segment as the mitral valve valve area value, and determine the mitral valve valve boundary of the corresponding phase frame and the topological closure of the corresponding phase frame.

[0131] It should be noted that, among the mitral valve area values ​​in each frame of the short time segment, the largest mitral valve area value is selected as the mitral valve area value, and the corresponding phase frame and the topologically closed mitral valve valve boundary of the corresponding phase frame are determined: using the mitral valve area values ​​in each frame of the short time segment and the topologically closed mitral valve valve boundary corresponding to each frame as input, the valve area values ​​are fused. The fusion process adopts a maximum value strategy, that is, the valve area values ​​of each frame are traversed and compared frame by frame to obtain the maximum valve area value, and the frame number of the frame that obtains the maximum valve area value is recorded simultaneously as the maximum value index; when there are multiple frames with valve area values ​​that are tied for the maximum value, the frame number with the earliest time sequence among the frames with the tied maximum values ​​is selected as the maximum value index to ensure that the phase frame determination result is unique.

[0132] Using the maximum value index as the frame-level retrieval key, frames with the same frame number as the maximum value index are identified in short time segments as the phase frames corresponding to the mitral valve orifice area values. Based on the maximum value index, the topologically closed mitral valve orifice boundaries corresponding to the frame number of the maximum value index are extracted from the topologically closed mitral valve orifice boundaries of each frame as the closure boundaries of the corresponding phase frames. The maximum orifice area value, the corresponding phase frame number, and the topologically closed mitral valve orifice boundaries of the corresponding phase frames are output and stored in association.

[0133] S4.3 Based on the back-mapping information, map the topologically closed mitral valve orifice boundary of the corresponding phase frame to the original coordinates of the corresponding phase frame, overlay and draw the mitral valve orifice boundary on the corresponding phase frame image and label the mitral valve orifice area value to generate a boundary overlay map.

[0134] It should be noted that, based on the backmapping information, the topologically closed mitral valve orifice boundary of the corresponding phase frame is mapped to the original coordinates of the corresponding phase frame. The topologically closed mitral valve orifice boundary is then overlaid on the corresponding phase frame image, and the mitral valve orifice area value is labeled to generate a boundary overlay map. Specifically, the mitral valve orifice area value, the corresponding phase frame number, and the topologically closed mitral valve orifice boundary of the corresponding phase frame are used as input. The corresponding backmapping information is retrieved according to the corresponding phase frame number. Based on the backmapping information, coordinate backmapping processing is performed on the coordinates of each boundary point in the closed boundary of the corresponding phase frame. Specifically, the boundary point coordinates are inversely scaled based on the scaling ratio parameter to restore the scale before clipping, and the inversely scaled boundary points are then resized based on the clipping offset parameter. The coordinates are reverse-translated to restore the original coordinate system of the corresponding phase frame, obtaining the topological closure boundary of the mitral valve orifice under the original coordinates of the corresponding phase frame. Optionally, the coordinates of the back-mapped boundary points are rounded and the boundary is clipped to adapt to pixel-level rendering. The original image of the corresponding phase frame is read from the short time segment with the corresponding phase frame number as the base map. The topological closure boundary after back-mapping is drawn as a superimposed contour by connecting the boundary points end to end in order and superimposed on the base map. At the same time, the area text label is drawn at the centroid position of the closure boundary with the mitral valve orifice area value as the label content, so that the superimposed contour and the area value are visible on the same frame image. The boundary superimposed map containing the original phase frame image, the back-mapped closure boundary and the area label is output and saved.

[0135] In summary, this invention achieves consistent selection of the most open phase frame of the mitral valve orifice under different individuals and imaging quality conditions by automatically and stably locating key phases of the cardiac cycle, thereby reducing the differences in manual frame selection and fluctuations in repeated measurements, and improving the stability and repeatability of mitral valve orifice area assessment; by reliably closing the valve orifice boundary to obtain and output an overlayable and verifiable result image, it achieves the synchronous presentation of area calculation results and image evidence, thereby reducing the risk of misjudgment caused by incomplete boundaries and misaligned annotations, and improving measurement accuracy, interpretability, and clinical efficiency.

[0136] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting the mitral valve orifice area based on artificial intelligence, characterized in that: include, A continuous frame sequence of mitral valve ultrasound was acquired and standardized preprocessed. A sliding window frame sequence was formed with a fixed length and step size. A phase embedding neural network was constructed using ResNet and Long Short-Term Memory neural network. The phase embedding neural network was pre-trained with phase self-supervision. The sliding window frame sequence was input into the phase embedding neural network to obtain a phase embedding vector sequence. Inter-frame similarity is calculated based on phase embedding vector sequences. Cardiac cycle boundaries are determined based on this similarity, and complete cardiac cycles are extracted. Open-phase scoring is then calculated within the complete cardiac cycle. The specific steps are as follows: The average value of the phase embedding vector of the starting frame and the phase embedding vector of the ending frame of the complete cardiac cycle boundary is calculated and normalized to generate a reference phase embedding vector. Calculate the cosine similarity between the phase embedding vector and the reference phase embedding vector, and invert the cosine similarity to obtain the open phase score; The most open phase frame of the valve orifice is selected based on the open phase score, and a short time segment centered on the most open phase frame of the valve orifice is extracted. Short time segments are used to locate and crop the ROI of the mitral valve orifice to obtain the ROI frame sequence and back-mapping information. A neural implicit boundary network is constructed by an image feature extraction network and an implicit probability function output network. The ROI frame sequence is input into the neural implicit boundary network, and the implicit probability function is output. The topologically closed mitral valve orifice boundary is obtained by isosurface extraction. The mitral valve orifice area is calculated for each frame of a short time segment based on the topologically closed mitral valve orifice boundary, and the mitral valve orifice area value is obtained from the orifice area of ​​each frame. The specific steps are as follows. Obtain the corresponding ultrasound image calibration information from short time segments, and calculate the conversion factor of the actual area of ​​a single pixel based on the ultrasound image calibration information; The region of the mitral valve orifice boundary of the short temporal segment topological closure is filled frame by frame, and the number of internal pixels is counted. The orifice area value of each frame is calculated by combining the actual area conversion factor of a single pixel. The largest mitral valve area value is selected from the valve area values ​​of each frame in the short time segment as the mitral valve area value, and the phase frame corresponding to the mitral valve area value and the topologically closed mitral valve valve boundary of the corresponding phase frame are determined. The mitral valve orifice area value is superimposed with the closed boundary of the corresponding phase frame according to the back-mapping information to form a boundary superposition map.

2. The mitral valve orifice area detection method based on artificial intelligence as described in claim 1, characterized in that: The acquired mitral valve ultrasound continuous frame sequence is standardized and preprocessed, and then formed into a sliding window frame sequence with a fixed length and step size. The specific steps are as follows. Acquire a continuous frame sequence of mitral valve ultrasound data, and number and sort the frames according to the acquisition order; The mitral valve ultrasound continuous frame sequence was subjected to frame-by-frame normalization preprocessing to obtain the normalized mitral valve ultrasound continuous frame sequence. The standardized mitral valve ultrasound continuous frame sequence is divided into sliding window frame sequences by using a fixed window length and step size.

3. The mitral valve orifice area detection method based on artificial intelligence as described in claim 2, characterized in that: The specific steps for inputting the sliding window frame sequence into the phase embedding neural network to obtain the phase embedding vector sequence are as follows. The sliding window frame sequence is input into the phase embedding neural network in chronological order. ResNet extracts features from the sliding window frame sequence frame by frame. The long short-term memory neural network outputs the phase embedding vector at each time step, forming a phase embedding vector sequence that corresponds to each frame of the sliding window frame sequence.

4. The mitral valve orifice area detection method based on artificial intelligence as described in claim 3, characterized in that: The process of calculating inter-frame similarity based on phase embedding vector sequences, determining cardiac cycle boundaries based on inter-frame similarity, and extracting complete cardiac cycles involves the following specific steps. The phase embedding vector of each frame in the phase embedding vector sequence is normalized to obtain the normalized phase embedding vector sequence. Inter-frame similarity is calculated based on the normalized phase embedding vector sequence to form an inter-frame similarity matrix; The cardiac cycle length is determined based on the inter-frame similarity matrix, and the cardiac cycle boundary is located based on the cardiac cycle length to extract the complete cardiac cycle.

5. The mitral valve orifice area detection method based on artificial intelligence as described in claim 4, characterized in that: The steps for calculating the open phase score within a complete cardiac cycle, selecting the most open valve phase frame based on the open phase score, and extracting a short time segment centered on the most open valve phase frame are as follows. The frame corresponding to the maximum open phase score is selected as the most open phase frame of the valve orifice, and a short time segment is extracted from the complete cardiac cycle with the most open phase frame of the valve orifice as the center.

6. The mitral valve orifice area detection method based on artificial intelligence as described in claim 5, characterized in that: The specific steps for locating and cropping the ROI (Region of Interest) of short temporal segments to obtain the ROI frame sequence and back-mapping information are as follows. Select the center frame of a short time segment, perform contrast enhancement and threshold segmentation, and combine morphological opening and closing operations to obtain a low-echo region cleanup mask; Select the target connected region of the valve orifice ROI from the low echo region purification mask and calculate the bounding rectangle. Generate the valve orifice ROI clipping box and record the clipping parameters to generate back-mapping information. The short temporal segment is cropped frame by frame using the ROI clipping box to form the ROI frame sequence. The size of the ROI frame sequence is uniformly processed frame by frame, and the clipping and scaling parameters are written back to the mapping information.

7. The mitral valve orifice area detection method based on artificial intelligence as described in claim 6, characterized in that: The specific steps are as follows: inputting the ROI frame sequence into the neural implicit boundary network, outputting the implicit probability function, and extracting the topologically closed mitral valve orifice boundary through isosurface extraction. The valve orifice ROI frame sequence is input into the neural implicit boundary network in chronological order. The image feature extraction network extracts frame-by-frame features from the valve orifice ROI frame sequence, and the implicit probability function output network outputs implicit probability functions from the frame-by-frame features. The mitral valve orifice boundary is obtained by performing isosurface extraction on the implicit probability function frame by frame, and the mitral valve orifice boundary is mapped back to a short time segment based on the back mapping information.

8. The mitral valve orifice area detection method based on artificial intelligence as described in claim 7, characterized in that: The specific steps for obtaining the topologically closed mitral valve orifice boundary by performing isosurface extraction on the implicit probability function frame by frame are as follows. When multiple closed boundaries are obtained by isosurface extraction, region filling is performed on each of the multiple closed boundaries and the number of pixels inside each closed boundary is calculated. The closed boundary with the largest number of internal pixels is selected as the mitral valve orifice boundary of topological closure.

9. The mitral valve orifice area detection method based on artificial intelligence as described in claim 1, characterized in that: The specific steps for superimposing the mitral valve orifice area value with the corresponding phase frame's closed boundary using back-mapping information to form a boundary overlay map are as follows: Based on the back-mapping information, the mitral valve orifice boundary of the topological closure of the corresponding phase frame is mapped to the original coordinates of the corresponding phase frame; The boundary of the mitral valve orifice is overlaid on the corresponding phase frame image and the area value of the mitral valve orifice is marked to generate a boundary overlay map.

Images