An echocardiogram cardiac cycle candidate reordering method

Abstract

The application belongs to the technical field of ultrasonic image processing, and provides an echocardiogram cardiac cycle candidate reordering method, which comprises the following steps: constructing a left ventricular area curve, generating a candidate cycle set and determining initial ED frames and initial ES frames of each candidate cycle; obtaining a candidate reliability score for each candidate cardiac cycle in the candidate cardiac cycle set; obtaining ED and ES key frames by fusing frame-level ED and ES phase probabilities, relative positions and statistical position priors; determining the number of candidates according to the divergence degree of the key frames corresponding to the candidate cycles with high candidate reliability score, and calculating the weighted median of the key frame index according to the candidate reliability weight to obtain consensus key frames; obtaining the integrated prediction error of the candidate cycles, and taking the consensus key frame corresponding to the candidate cycle with the minimum integrated prediction error as the final ED frame and the final ES frame. The application improves the stability and robustness of the alignment of complex echocardiogram sequences.

Description

Technical Field

[0001] This invention relates to the field of ultrasound image processing technology, and more specifically to a method for reordering candidate cardiac cycles in echocardiography. Background Technology

[0002] Automatic estimation of left ventricular ejection fraction (LVEF) in echocardiography typically requires first obtaining consecutive frames of left ventricular segmentation, and then locating the end-diastolic (ED) and end-systolic (ES) frames. The ED frame usually corresponds to the position where the left ventricular area or volume is near its maximum, while the ES frame usually corresponds to the position where the left ventricular area or volume is near its minimum. Therefore, the accurate location of the ED and ES frames directly affects the stability of subsequent LVEF estimation.

[0003] Existing methods typically select local maxima, global maxima, local minima, or global minima directly from the area curve of the entire sequence, or first select a candidate cardiac cycle and then search for area extrema within that candidate cycle. However, in actual echocardiographic videos, there are often multiple similar cycles, local artifacts, segmentation jitter, video boundary truncation, and heart rate changes, making it easy to select incorrect cycles or incorrect keyframes by simply relying on the area curve extrema.

[0004] Furthermore, using only a fixed cycle length or fixed phase position prior may introduce biases in different individuals, with different heart rates, and under different acquisition conditions; using only a single learning model for candidate cycle ranking is also susceptible to the effects of random initialization, limited training samples, and local feature anomalies, leading to unstable candidate ranking results.

[0005] Therefore, there is a need for an echocardiographic cardiac cycle candidate reordering method that can reduce the risk of individual candidate failure. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for reordering candidate echocardiogram cycles, thereby solving the problems of missed candidate cycles, misaligned keyframes, and unstable single candidate decision-making in complex echocardiogram videos.

[0007] This invention provides a method for reordering candidate cardiac cycles in echocardiography, comprising: S1. Obtain the left ventricular probability map sequence or binary mask sequence of continuous echocardiographic video, and construct the left ventricular area curve frame by frame; S2. Generate a set of candidate cycles based on the left ventricular area curve, and determine the initial ED frame and initial ES frame for each candidate cycle; S3. For each candidate cardiac cycle in the candidate cardiac cycle set, obtain the candidate reliability score; S4. Within each candidate period, construct feature vectors to obtain frame-level ED phase probability and ES phase probability, and combine relative position and statistical position priors for fusion scoring to obtain ED and ES keyframes. S5. Based on the degree of disagreement of the keyframes corresponding to the candidate periods with high candidate reliability scores, adaptively determine the number of candidates participating in the consensus calculation, and calculate the weighted median of the keyframe index according to the candidate reliability weights to obtain the ED and ES consensus keyframes. S6. Construct multiple candidate error prediction models with different random initialization parameters to obtain the integrated prediction error of the candidate period. Use the consensus key frame corresponding to the candidate period with the smallest integrated prediction error as the final ED frame and the final ES frame.

[0008] Optionally, the candidate cardiac cycle set is: , in, This represents the set of candidate periods between peaks. This represents a set of candidate periods for a sliding window, constructed starting from a local maximum frame and according to various reference period ratios. This represents the set of candidate boundary completions generated for the truncation period at the beginning or end of the video. The initial ED frame and the initial ES frame are the frames with the largest and smallest areas within the corresponding candidate cardiac cycle, respectively.

[0009] Optionally, the candidate reliability score is: , Supervised ranking score To prefit the obtained candidate cycle supervised ranking model for the first One candidate cardiac cycle output; Annotated prior scores Calculated based on the empirical distribution of candidate number, relative center position, and period length deviation in the fitted data. , , and These represent the probabilities of the pre-sorted sequence number, relative center position, and cycle length deviation of the i-th candidate cardiac cycle in the empirical distribution of the fitted data, respectively. Normative score , Indicates the original score. and They represent the first Initial ED frame index and initial ES frame index for each candidate cardiac cycle. For the first Frame segmentation confidence; This represents a very small positive number used to prevent the denominator from being zero; and These are the proportions of inter-frame differences that satisfy the correct direction of change during the diastolic-to-contraction descent phase and the post-contraction recovery phase, respectively; video boundary risk. ; and They represent the first The start frame index and end frame index of each candidate cardiac cycle; This is the normalized distance between the candidate area curve and the standard periodic template. Curve shape score , Center position score ; For the corresponding weights.

[0010] Alternatively, the candidate periodicity supervised ranking model can be determined using only the fitted data: Candidate periods that simultaneously contain manually labeled ED frames and manually labeled ES frames are used as positive samples, while candidate periods that do not simultaneously contain manually labeled ED frames and manually labeled ES frames are used as negative samples. A supervised ranking model for candidate periods is obtained by performing linear fitting with ridge regularization constraints based on the ranking position, relative position of the center, period length deviation, curve shape, segmentation confidence, and boundary risk.

[0011] Optionally, S4 includes: Construct feature vectors; the components in the feature vectors include the normalized area of ​​the smooth area curve, first-order change and second-order change, stationary point score, peak curvature score and valley curvature score, segmentation confidence, segmentation uncertainty, relative temporal position of the whole video, video edge distance, mean, standard deviation and range of the normalized area within the local window; Based on a pre-trained weighted logistic regression phase model, the frame-level probabilities of ED and ES phases are obtained, and the preset number of video frames with the highest frame-level probabilities are selected as candidate ED and ES frames. For any candidate ED and ES frames, calculate the relative and statistical position priors within the candidate cardiac cycle, respectively. Calculate the fusion score for ED / ES frame pairs.

[0012] Optionally, the fusion score is: , in, Let be the frame-level probabilities of the ED phase and ES phase in frame t. These are the bias terms used in pre-training. The weight vector of the feature vector. Let be the normalized feature vector of the t-th frame; Candidate ED and ES frames in candidate cardiac cycle The relative position within, For statistical location priors; Keyframe timing constraints .

[0013] Optionally, ED and ES frames and The results were obtained through weighted logistic regression training on the fitted data, specifically as follows: During training, soft labels are generated using a Gaussian function with a standard deviation of 4 frames, centered on manually annotated ED or ES keyframes, and soft labels with a standard deviation less than 0.05 are set to 0. Optimization is performed using class-balanced weights and a L2 regularization term with a coefficient of 0.01. The format is fixed after training. and .

[0014] Optionally, S5 includes: Candidates are sorted from highest to lowest reliability score, and the degree of divergence is calculated for the highest-ranked preset cardiac cycles. Based on the degree of divergence among the candidates, the candidate heartbeat cycles that actually participate in the consensus calculation are determined; For the identified candidate cardiac cycles, calculate the normalized reliability weights and the weighted median of ED and ES to determine the consensus keyframes.

[0015] Optionally, S6 includes: Multiple candidate error prediction models with different random initialization parameters are constructed, and the input features of each candidate period are input into the model. The average value of the outputs of multiple models is used as the integrated prediction error corresponding to the candidate period. The candidate period corresponding to the minimum integrated prediction error is taken as the final candidate period, and the ED and ES consensus key frames corresponding to the final candidate period are output as the final ED frame and the final ES frame.

[0016] Optionally, the candidate error prediction model uses a multilayer perceptron with two hidden layers. The network structure includes, in sequence: an input layer, a fully connected layer with 64 neurons, a ReLU activation layer, a Dropout layer with a dropout rate of 0.08, a fully connected layer with 32 neurons, a ReLU activation layer, and an output layer with 1 neuron. For any candidate error prediction model, for the first The training objective corresponding to each candidate cardiac cycle is: Each candidate error prediction model is trained by minimizing the mean square error between the model output and the above training objective; and For the first ED and ES consensus keyframes for each candidate heartbeat cycle; and These are real keyframes annotated manually.

[0017] By adopting the above technical solution, this application has the following beneficial effects: This invention improves the recall capability of complete cycles, weak peak cycles, and boundary truncation cycles by generating candidate cycles from multiple sources; comprehensively considers cycle length, curve morphology, phase support, segmentation quality, center position, and boundary risk through candidate reliability scoring; reduces keyframe offset caused by simple reliance on extreme values ​​by fusing frame-level phase probability and statistical phase prior; reduces the impact of single candidate failure by candidate divergence-aware multi-candidate adaptive consensus; and reduces random fluctuations of a single model and improves candidate ranking stability by candidate quality integration and re-ranking, thereby improving the stability and robustness of center-cycle alignment in complex echocardiographic sequences. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0019] Figure 1 This invention provides a flowchart of one of the methods for reordering candidate echocardiographic cardiac cycles according to an embodiment of the present invention. Figure 2 This illustrates a second flowchart of a method for reordering candidate echocardiographic cardiac cycles according to an embodiment of the present invention. Figure 3 This illustration shows a schematic diagram of generating a candidate cycle set based on the left ventricular area curve according to an embodiment of the present invention; Figure 4 This diagram illustrates the integrated reordering of candidate cycle multidimensional reliability scores and candidate quality provided by an embodiment of the present invention. Figure 5 This diagram illustrates the frame-level ED frame and ES frame phase probability and statistical phase prior fusion localization provided by an embodiment of the present invention. Figure 6 This diagram illustrates the adaptive consensus calibration of multiple candidate end-diastolic frames and end-systolic frames based on candidate divergence perception provided in an embodiment of the present invention. Figure 7 A comparison of the rigorous average frame error for different cardiac cycle alignment methods; Figure 8 The frame error distribution of the integrated candidate quality reordering method on the evaluation set. Detailed Implementation

[0020] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.

[0021] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] In one embodiment, such as Figure 1-2 As shown, a method for reordering candidate echocardiographic cardiac cycles is provided, including: S1. Obtain the left ventricular probability map sequence or binary mask sequence of continuous echocardiographic video, and construct the left ventricular area curve frame by frame.

[0023] The left ventricular probability map sequence or binary mask sequence of continuous echocardiographic video can be obtained based on the pre-existing left ventricular segmentation model; and a frame-by-frame left ventricular area curve can be constructed based on the left ventricular probability map sequence or binary mask sequence. Let the first... The left ventricle binary mask is The image width and height are respectively and Then the first Left ventricular area surrogate quantity of the frame Calculate according to formula (1-1): (1-1) in, Indicates the first The number of pixels contained in the left ventricular region of the frame, sequence The original area curve representing the change in left ventricular area with frame index. This represents the total number of video frames. To reduce the impact of frame-by-frame segmentation jitter on local extremum extraction, the original area curve is smoothed using an odd-length moving average. In this embodiment, the window length is set to 5, i.e., the radius... The smooth area curve is calculated according to formula (1-2): (1-2) in, This indicates that frame indices exceeding the beginning and end boundaries of the video will be truncated to 0. Within this range, its effect is equivalent to boundary replication. Subsequent candidate period generation and curve feature calculation are both based on smooth area curves. Finish.

[0024] During the area curve construction process, the system acquires a continuous sequence of left ventricular probability maps or a binary mask sequence. For the probability map sequence, a preset threshold can be used to obtain the binary mask; for the binary mask sequence, the number of foreground pixels in the left ventricular region of each frame can be directly counted to form a frame-by-frame left ventricular area curve. To suppress the impact of frame-by-frame segmentation jitter on the area curve, moving average, low-pass filtering, or other smoothing methods can be used to process the area curve. For probability maps obtained from multiple segmentation models or multiple random inferences, pixel-level variance, mask consistency, or area fluctuation can also be calculated as reference information for segmentation confidence.

[0025] S2. Generate a set of candidate cardiac cycles based on the left ventricular area curve.

[0026] First, smooth the area curve. Compare adjacent values ​​frame by frame. When the... Frames simultaneously satisfy , At that time, the first The frame is determined to be a local maximum frame; when the first... Frames simultaneously satisfy , At that time, the first The frame is identified as a local minimum frame. This yields a set of local maximum frames sorted in ascending order of frame index: (2-1) And the set of local minimum frames: (2-2) Among them, the set of local maxima frames Used to determine the boundaries of inter-peak candidate periods and sliding window candidate periods; set of local minimum frames. Used to determine the initial ES frame within each candidate period. When one or more local minima exist within a candidate period, the local minima with the smallest area is selected as the initial ES frame; when no local minima exist within a candidate period, the frame with the smallest area within that candidate period is selected as the initial ES frame. The frame with the largest area within a candidate period is selected as the initial ED frame.

[0027] For the Candidate cardiac cycle The initial ED frame and initial ES frame determined based on the extreme values ​​of the area curve are denoted as follows: and The initial ED / ES frame is used to calculate the area reduction ratio, descent segment change ratio, rise segment change ratio, ED / ES segmentation confidence, and original candidate period score for the candidate period. It also participates in candidate period deduplication, pre-sorting, and reliability scoring. The initial ED / ES frame is the initial localization result during the candidate period generation stage, not the final output ED / ES frame for that candidate period.

[0028] In determining the first After generating the initial ED and initial ES frames for each candidate cardiac cycle, the system calculates the original score of the candidate cycle according to preset weights based on the normalized area difference between the two frames, the percentage change in the descent phase, the percentage change in the ascent phase, the segmentation confidence, the segmentation uncertainty, the degree of area abrupt change, and the ED / ES frame interval deviation. And compare the score with the candidate cardiac cycle The data is saved in conjunction with the data for step S3 to perform candidate reliability scoring.

[0029] The normalized area difference refers to the difference between the smoothed areas corresponding to the initial ED frame and the initial ES frame, divided by the difference between the maximum and minimum values ​​of the smoothed area curve for the entire video. The falling segment refers to the continuous frame interval from the initial ED frame to the initial ES frame; the percentage change in the falling segment is the proportion of adjacent frame pairs whose area is smaller than the previous frame out of all adjacent frame pairs within this interval. The rising segment refers to the continuous frame interval from the initial ES frame to the end frame of the candidate period; the percentage change in the rising segment is the proportion of adjacent frame pairs whose area is larger than the previous frame out of all adjacent frame pairs within this interval. The segmentation confidence and segmentation uncertainty are obtained statistically from the segmentation probability map of the corresponding frames. The degree of area abrupt change is obtained statistically from the area difference between adjacent frames within the candidate period. The ED / ES frame interval deviation is calculated based on the deviation of the frame interval between the initial ED frame and the initial ES frame relative to the reference period length L. All the above parameters are directly calculated from the area curve, segmentation probability map, and candidate period boundary obtained in step S1.

[0030] set The median of the effective intervals between all adjacent local maxima frames is used to obtain the video reference period length. : (2-3) in, and These represent the minimum and maximum number of frames allowed in the candidate period, respectively. In this embodiment, they are set to... =8, =90.

[0031] When there is no adjacent local maximum interval satisfying the above length range, the reference period length is taken as... .

[0032] The candidate cardiac cycle set is as follows: (2-4) in, This represents the set of inter-peak candidate periods composed of adjacent or nearest-neighbor area peaks. This represents a set of candidate periods for a sliding window, constructed starting from a local maximum frame and according to various reference period ratios. This represents the set of candidate boundary completions generated for the truncation period at the beginning or end of the video.

[0033] (2-5) and Represent the sets of local maxima frames respectively The Middle The and the first The index of a local maximum frame; Indicates from the first Frame to the A continuous frame interval, including the two endpoints and all frames in between; These represent the minimum and maximum number of frames allowed in the candidate period, respectively; (2-6) Here and middle The same, both represent the indices of the j-th and k-th local maxima frames in set P. To preset the reference period ratio, Indicates the last frame of the video. This is used to ensure that the end frame of the candidate period does not exceed the last frame of the video. .

[0034] (2-7)

[0035] gather The elements in the text are consecutive frame intervals, not... , and Three values. This indicates the index of the end frame corresponding to a preset length extending from the first frame of the video. This indicates the index of the starting frame corresponding to a preset length extending forward from the last frame of the video. The preset length is... , and Frame. When the candidate interval exceeds the video range, it is truncated to frame 0 to frame 1. Within the frame range.

[0036] The length of each candidate period is not a fixed value, but is determined by the reference period length L, the area peak position, and the video boundary. In this embodiment, the candidate period length is limited to the range of 8 to 90 frames. Deduplication is performed on candidate periods that are the same in the start frame, end frame, initial ED frame, and initial ES frame, and a maximum of 80 candidate periods are retained after pre-sorting. Figure 2 The different candidate intervals in the data correspond to the inter-peak candidate period, the sliding window candidate period, and the boundary completion candidate period, respectively.

[0037] like Figure 3 As shown, during the candidate cardiac cycle generation process, the system generates multi-source candidate cycles based on the smoothed left ventricular area curve. Peak-valley candidates are determined by adjacent local area peaks or peak-valley structures in the area curve and are used to capture cardiac cycles with obvious periodic changes; sliding window candidates are obtained by sliding across the entire sequence using a preset or estimated cycle length and are used to supplement candidate cycles that may be missed by peak-valley detection; boundary completion candidates are used to cover truncated cycles that may exist at the beginning or end of the video sequence. Figure 2 The text simultaneously identifies peak and valley candidates, sliding window candidates, and boundary completion candidates to illustrate the coverage effect of different candidate sources on different types of cycles.

[0038] S3. For each candidate cardiac cycle in the candidate cardiac cycle set, obtain the candidate reliability score.

[0039] The candidate cardiac cycles set C obtained in step S2 does not inherently have a specific order. The system first deduplicates candidate cycles where the start frame, end frame, initial ED frame, and initial ES frame are identical. Then, it pre-sorts the candidate cycles based on factors such as the area variation, the similarity between the candidate cycle length and the reference cycle length L, the temporal relationship between the ED and ES frames, and the video boundary position of the candidate cycle, retaining a maximum of the first 80 candidate cycles. After pre-sorting, the candidate cycles are numbered sequentially according to the sorting results. This represents the candidate cardiac cycle located at position i after pre-sorting.

[0040] For the i-th candidate cardiac cycle after deduplication and presorting The supervised ranking score, the labeled statistical prior score, the standardization score, the curve shape score, and the center position score are calculated respectively, and the candidate reliability score is calculated according to formula (3-1).

[0041] (3-1) Supervised ranking score To prefit the obtained candidate cycle supervised ranking model for the first Output of candidate cardiac cycles.

[0042] In this embodiment, the candidate periodicity supervised ranking model is determined solely using fitted data: candidate periods containing both manually labeled ED frames and manually labeled ES frames are considered positive samples, while candidate periods not containing both are considered negative samples; a ridge-regularized linear fit is performed based on the candidate period's ranking position, relative center position, periodicity length deviation, curve shape, segmentation confidence, and boundary risk. After fitting, the relevant information of the i-th candidate period is input into the supervised ranking model to obtain... The evaluation data is not used in determining the supervised ranking model.

[0043] Annotated prior scores Calculated based on the empirical distribution of candidate number, relative center position, and period length deviation in the fitted data. (3-2) `index` represents the sequence number of the candidate cardiac cycle in the pre-sorted results, `pos` represents the relative position of the candidate cardiac cycle center in the video, and `len` represents the deviation of the candidate cardiac cycle length from the reference cycle length `L`; `u` takes values ​​from `index`, `pos`, and `len` in sequence. , and Let these represent the probabilities of the pre-sorted index, relative center position, and cycle length deviation of the i-th candidate cardiac cycle in the empirical distribution of the fitted data, respectively: (3-3) pos and len are each divided into 10 equally spaced intervals; candidate cardiac cycles that simultaneously include manually labeled ED frames and ES frames are considered positive candidates. The addition of 1 and 2 in equation (3-3) is used to avoid... The value is either 0 or 1. The above empirical distribution is obtained only from the statistics of the fitted data and remains unchanged during the evaluation stage.

[0044] Normative score (3-4) In equation (3-4), This represents the original score of the candidate cycle obtained in step S2 when generating the i-th candidate cardiac cycle, which is used to characterize the area change, temporal trend, segmentation quality, and rationality of the initial ED / ES frame interval of the candidate cycle. The first expression defined by (4-8) Frame segmentation confidence; and They represent the first Initial ED frame index and initial ES frame index for each candidate cardiac cycle; Indicates the first The smoothed left ventricular area corresponding to the frame; and They represent the first The start and end frame indices of each candidate cardiac cycle; L represents the video reference cycle length determined based on the interval between adjacent local maxima frames; This represents a very small positive number used to prevent the denominator from being zero; Indicates the first Video boundary risks for each candidate heartbeat cycle; Indicates the first The distance between the normalized area curve of each candidate cardiac cycle and the standard cycle template.

[0045] Indicates the first In the descending segment of a candidate cardiac cycle from the initial ED frame to the initial ES frame, the proportion of adjacent frame pairs whose smooth area of ​​the next frame is no greater than that of the previous frame to the total number of adjacent frame pairs in the descending segment. This represents the proportion of adjacent frame pairs in the rise segment from the initial ES frame to the candidate period end frame, where the smoothed area of ​​the next frame is not less than that of the previous frame, out of the total number of adjacent frame pairs in that rise segment.

[0046] Video boundary risk is calculated using candidate edge distance: (3-5) for The video boundary terms in the text. Therefore, Directly used for normative scoring The calculation. The closer the candidate cardiac cycle is to the beginning or end of the video, the better. The larger the value, the smaller the video boundary term, thus reducing the normalization score for that candidate period.

[0047] Curve shape score , This is the normalized distance between the candidate area curve and the standard periodic template. and These represent the proportions of inter-frame differences that satisfy the correct direction of change during the diastolic-to-contraction descent phase and the post-contraction recovery phase, respectively.

[0048] (3-6) (3-7) Where mean represents the average of the corresponding sampling positions. This means resampling the candidate area curve to 100 equally spaced points. Normalize means normalizing the candidate area curve to 0 to 1. MAE means average absolute distance. The smaller the value, the higher the value. The closer a candidate heart cycle is to your typical heart cycle pattern in the data, the better.

[0049] Center position score: (3-8) , , , , These are the corresponding weights. Each scoring weight is selected using only the fitted data through a grid search, with the minimum strict average frame error as the selection criterion, and remains unchanged on the evaluation data after determination. The basic combined weights used in this embodiment are... =2.0、 =1.0、 =0.5、 =0 and =0.5. In this embodiment =0, therefore the independent curve shape score is 0. It does not directly participate in the calculation of the candidate reliability score Qi; the template distance, the percentage change in the descent phase, and the percentage change in the rise phase are still calculated through... Participate in the calculation. Because =0, independent curve shape score Not directly involved in this basic combination The calculations are as follows. However, the curve template distance, the percentage change in the area decreasing segment, and the percentage change in the area increasing segment are still included. Therefore, the information about the curve shape is still obtained through... right It has an impact.

[0050] like Figure 4 As shown, during the candidate reliability scoring process, the system quantitatively evaluates the cycle length, curve shape, phase support, segmentation confidence, center position, and boundary risk of each candidate cycle. Specifically, cycle length determines whether the candidate cycle falls within a reasonable cardiac cycle range; curve shape determines whether the area change within the candidate cycle conforms to the basic changes in cardiac diastole and systole; phase support measures the degree of phase localization support between the end-diastolic and end-systolic frames within the candidate cycle; segmentation confidence reflects the stability of the left ventricle segmentation results within the candidate cycle; center position determines whether the candidate cycle's location in the video sequence is reasonable; and boundary risk reduces the candidate reliability score when there is truncation at the beginning or end, keyframes close to the candidate boundary, or abnormal candidate structure. The candidate reliability score results are used for subsequent high-reliability candidate selection, candidate consensus calculation, and candidate quality re-ranking.

[0051] S4. Within each candidate period, construct a feature vector to obtain the frame-level ED phase probability and ES phase probability, and combine the relative position and statistical position prior to perform fusion scoring to obtain ED and ES keyframes.

[0052] First, the smoothed left ventricular area curve is normalized. Let... Indicates the current frame index. This represents the frame index for traversing the entire video sequence, then the first... Normalized area value of the frame Calculate according to formula (4-1): (4-1) In equation (4-1), Indicates the first The smoothed area value corresponding to the frame. and These represent the minimum and maximum values ​​of the smoothed area curve for the entire video sequence, respectively. This represents a very small positive number used to prevent the denominator from being zero.

[0053] Then, the first and second order changes of the normalized area curve are calculated using discrete gradients: , (4-2) No. The feature vector of a frame consists of the following thirteen components: (4-3) Wherein, the normalized area, the first-order change, and the second-order change are respectively: , , (4-4) The station score is (4-5) Peak curvature score and valley curvature score are respectively: (4-6) (4-7) In the formula, , This represents the 90th percentile of the absolute value of the change in the entire video sequence. To prevent extremely small positive numbers with a denominator of zero, This means restricting x to the interval [0,1]. set up Let be the predicted probability of the left ventricle at pixel position (x, y) in frame t. M tGiven the corresponding left ventricular binary mask, the segmentation confidence level is: (4-8) When multiple probability maps are obtained using R random inferences, the segmentation uncertainty is: (4-9) Without performing multiple random inferences, let .

[0054] The relative temporal position and video edge distance of the entire video are as follows: (4-10) (4-11) Centered on frame t, three frames are extracted within the video boundary, forming a local window containing a maximum of seven frames. The mean, standard deviation, and range of the normalized area within this local window are as follows: (4-12) (4-13) (4-14) Finally, component-by-component standardization is performed using the mean and standard deviation of each feature component in the fitted data: (4-15) In the formula, and Let represent the mean and standard deviation of the j-th feature component in the fitted data, respectively.

[0055] In this embodiment, two weighted logistic regression phase models, ED and ES, are trained respectively, and their frame-level probabilities are calculated according to equation (4-16): (4-16) in, and Let represent the probabilities that the t-th frame belongs to the ED phase and the ES phase, respectively; This is the bias term for the phase model; This is the weight vector corresponding to the thirteen-dimensional standardized frame features; Let be the normalized feature vector of the t-th frame.

[0056] When calculating the ED phase probability, the bias term and weight vector corresponding to the ED phase model are used, and when calculating the ES phase probability, the bias term and weight vector corresponding to the ES phase model are used; both models are applied to the same standardized feature vector in the t-th frame.

[0057] and Both models were trained using weighted logistic regression on the fitted data. During training, soft labels were generated using a Gaussian function with a standard deviation of 4 frames, centered on manually labeled ED or ES keyframes, with labels less than 0.05 set to 0. Optimization was performed using class-balanced weights and a L2 regularization term with a coefficient of 0.01. This embodiment was trained for 800 epochs with an initial learning rate of 0.1, which was halved at 40% and 75% of the training epochs, respectively. Both models were trained using only the fitted data, and the training time was fixed after completion. and The evaluation data will no longer be updated.

[0058] For each candidate cardiac cycle generated in step S2 Maintain the starting frame of this candidate period and end frame Unchanged, within the interval From all video frames, candidate ED frames and candidate ES frames are re-selected based on frame-level phase probability.

[0059] Specifically, select The 12 largest video frames were selected as candidate ED frames. The 12 largest video frames are selected as candidate ES frames; if a candidate period contains fewer than 12 frames, all frames within that candidate period are retained. Candidate ED / ES frames and the initial ED / ES frames determined in step S2 are located in the same candidate period. Within the scope of the candidate ED / ES frame, it is not required to be centered on the initial ED / ES frame, nor is it required to include the initial ED / ES frame. Subsequently, the candidate ED / ES frames are combined and scored using Equation (4-19). The frame pair with the highest score is used as the corrected ED / ES frame for that candidate period and replaces the initial ED / ES frame in subsequent multi-period consensus and candidate quality reordering.

[0060] For any candidate ED frame and candidate ES frames In the candidate cardiac cycle The relative positions within are as follows: , (4-17) The statistical location priors for ED and ES are calculated from candidate cardiac cycles in the fitted data that simultaneously contain manually labeled ED and ES keyframes. Let the start and end frames of the nth candidate cardiac cycle be... and The manually annotated ED and ES keyframes are as follows: and ,but: , (4-18) The fitting obtained in this embodiment =0.1111、 =0.6053. The fusion score of the ED / ES frame pair is calculated according to equation (4-19): (4-19) Keyframe timing constraints A positive reward of 0.10 is given when the candidate ES frame is located after the candidate ED frame; a negative penalty of 0.30 is imposed when the candidate ES frame is not located after the candidate ED frame.

[0061] No. The ED and ES keyframes corresponding to each candidate cardiac cycle are determined according to equation (4-20): (4-20) like Figure 5 As shown, during keyframe phase localization, the aforementioned two weighted logistic regression phase models, ED and ES, predict the probability of each frame being an end-diastolic or end-systolic frame based on the area sequence and its local variation characteristics within the candidate period. The local variation characteristics may include at least one of the following: area value, first-order change of the area curve, second-order change of the area curve, relative position within the candidate period, and segmentation confidence. Subsequently, the system fuses the frame-level end-diastolic and end-systolic phase probabilities with the relative position prior obtained from historical labeled data to determine the ED and ES keyframes within the candidate period. By fusing phase probabilities with statistical phase priors, keyframe shifts caused by local artifacts, segmentation jitter, or heart rate changes when relying solely on the maximum or minimum area values ​​can be reduced.

[0062] S5. Based on the degree of disagreement of the keyframes corresponding to the candidate periods with high candidate reliability scores, adaptively determine the number of candidates participating in the consensus calculation, and calculate the weighted median of the keyframe index according to the candidate reliability weights to obtain the ED and ES consensus keyframes.

[0063] All candidate cardiac cycles from the same echocardiogram video are evaluated according to the candidate reliability score obtained in step S3. Sort from highest to lowest. For the top eight candidate cardiac cycles, calculate the candidate divergence level using the range of their ED and ES keyframe indices: (5-1) Where i=1,2,…,8 represents the top eight candidate cardiac cycles with the highest reliability scores, rather than the original sequence numbers when the candidate cardiac cycles were generated. The smaller the candidate divergence degree D, the more consistent the ED and ES keyframe positions given by multiple high-reliability candidate cardiac cycles.

[0064] Number of candidates involved in the formula calculation Determine according to formula (5-2): (5-2) in, This represents the actual number of candidate cardiac cycles participating in the consensus calculation, not the total number of candidate cardiac cycles in the candidate pool. When candidates are aligned, the top 8 are retained to utilize more cycle information; when there is significant divergence among candidates, only the top 3 highly reliable candidates are retained to reduce the impact of abnormal candidates. Threshold The threshold is determined solely through fitted data. Specifically, the j / 50th percentile of the candidate divergence level D in the fitted data, i.e., the 2nd, 4th, ..., 98th percentiles, are calculated as 49 candidate thresholds, and the strict average frame error corresponding to each candidate threshold is calculated. The divergence level value corresponding to the minimum strict average frame error is selected as the final threshold. In this embodiment, the following is obtained from the fitted data: =75.5. Once this threshold is determined, it remains unchanged in the evaluation data and is not reselected or adjusted using the evaluation data.

[0065] According to The first row arranged from highest to lowest For each candidate cardiac cycle, the normalized reliability weight and the weighted median of ED and ES are calculated according to equations (5-3) to (5-5): (5-3) (5-4) (5-5) in, Indicates candidate reliability score The highest front One candidate heartbeat cycle; Indicates the reliability score of the candidates Limited to the interval [ Within 20,20], to prevent numerical overflow during exponent calculation; Let represent the normalized reliability weight of the i-th candidate cardiac cycle, and satisfy . ; This represents the weighted median.

[0066] When calculating the weighted median of ED, the first One candidate heartbeat cycle Arrange the frames in ascending order of their frame indices, and then arrange the corresponding weights in the same order. Then, starting from the smallest frame index, the weights are added sequentially until the cumulative weight is not less than the minimum weight. The ED frame index is The ES weighted median is calculated in the same way, resulting in... .

[0067] like Figure 6 As shown, in the multi-candidate adaptive consensus process, the system obtains multiple sets of candidate end-diastolic frame indices and candidate end-contraction frame indices from the candidate periods ranked high in candidate reliability scores, and calculates the degree of keyframe divergence among highly reliable candidates. When the degree of candidate divergence is small, it indicates that the judgments on the keyframe positions of multiple candidate periods are relatively consistent. In this case, more candidates are used to participate in the consensus calculation to make full use of multi-candidate information. When the degree of candidate divergence is large, it indicates that there is significant inconsistency among candidates. In this case, fewer candidates are used to participate in the consensus calculation to reduce the impact of abnormal candidates on the final result. Subsequently, the system determines the candidate weights based on the candidate reliability scores, and calculates the weighted median for the end-diastolic frame index and the end-contraction frame index respectively to obtain the consensus end-diastolic frame and the consensus end-contraction frame.

[0068] S6. Construct multiple candidate error prediction models with different random initialization parameters to obtain the integrated prediction error of the candidate period. Use the consensus key frame corresponding to the candidate period with the smallest integrated prediction error as the final ED frame and the final ES frame.

[0069] S61. Construct multiple candidate error prediction models with different random initialization parameters, and input the input features of each candidate period into the model. The average value of the outputs of multiple models is used as the integrated prediction error corresponding to the candidate period.

[0070] This embodiment constructs seven candidate error prediction models with different random initialization parameters and performs integrated quality re-ranking on candidate cardiac cycles. Each candidate error prediction model is trained using only fitted data, without using evaluation data for parameter updates or model selection. Each candidate error prediction model employs a multilayer perceptron with two hidden layers. Its network structure is as follows: input layer, fully connected layer with 64 neurons, ReLU activation layer, Dropout layer with a dropout rate of 0.08, fully connected layer with 32 neurons, ReLU activation layer, and output layer with 1 neuron.

[0071] No. Input features of candidate cardiac cycles This includes candidate period length, area change magnitude, curve shape, ED and ES phase probabilities, relative positions of ED and ES, candidate period boundary distance, segmentation confidence, statistical prior bias, and candidate risk features. Manually annotated real ED and ES keyframe indices and the error fields calculated from them are used only to construct training objectives and are not used as model input features.

[0072] No. The training objective for each candidate cardiac cycle is: (5-6) in, and For the first The ED and ES keyframes of each candidate cardiac cycle are determined according to equation (4-20); and These are manually annotated real keyframes. Each candidate error prediction model is trained by minimizing the mean squared error between the model output and the training objective.

[0073] In this embodiment, the seven random initialization seeds used are 1, 2, 3, 5, 7, 11, and 13. Let the first seed be... The candidate error prediction models are: Then the first The ensemble prediction error for each candidate cardiac cycle is calculated according to equation (5-7): (5-7) in, Indicates the first The candidate error prediction model for the first... The prediction error of the output of each candidate cardiac cycle; This indicates the number of candidate error prediction models participating in the integration. The seven models use the same network structure and fitted data, but with different random initialization parameters; their outputs are taken as the arithmetic mean to reduce prediction fluctuations caused by differences in initialization of individual models.

[0074] S62. Take the candidate period corresponding to the minimum integrated prediction error as the final candidate period, and output the ED and ES consensus key frames corresponding to the final candidate period as the final ED frame and the final ES frame.

[0075] The final candidate period is determined by the following formula: (5-8) And output the consensus keyframe corresponding to the candidate period: , (5-9) In one optional implementation, the system uses equation (5-3) to perform weighted median consensus on the ED and ES frame indices of multiple highly reliable candidates to obtain consensus keyframes. In the candidate quality integration reordering method adopted in this embodiment, the system further uses equation (5-7) to calculate the integration prediction error of each candidate period and selects the candidate period with the smallest integration prediction error as the final target period. The two methods constitute the multi-candidate consensus calibration method and the candidate quality reordering method, respectively. In specific implementation, the output method used is predetermined, rather than being randomly selected for the same video.

[0076] After determining the final end-diastolic frame and the final end-systolic frame, the area-substitute left ventricular ejection fraction is calculated based on the corresponding left ventricular areas of the two frames, as expressed by equation (5-10): (5-10) in, This indicates the left ventricular area surrogate quantity corresponding to the final ED frame. This represents the left ventricular area surrogate quantity corresponding to the final ES frame. The area surrogate left ventricular ejection fraction is used for algorithm evaluation and relative comparison, and is not directly equivalent to the clinical volume measurement results obtained based on the actual three-dimensional volume.

[0077] like Figure 4 As shown, during the candidate quality integration and re-ranking process, the system uses multiple candidate error prediction models to predict candidate positioning errors separately, and integrates the prediction results of multiple models to obtain the integrated prediction error of the candidate period. The candidate error prediction models can be trained using different random initialization methods. The model input consists of candidate features available during the inference phase, including at least one of the following: candidate period length, area change amplitude, curve morphology features, phase probability features, keyframe relative position, boundary distance, segmentation confidence, and risk features. The model input does not include the true end-diastolic frame index, the true end-contraction frame index, or error information directly calculated using the true index, thus avoiding label leakage. The system re-ranks the candidate periods based on the integrated prediction error and determines the final ED frame and the final ES frame.

[0078] During the result output and verification process, the system calculates the area-proximity left ventricular ejection fraction based on the left ventricular area corresponding to the final end-diastolic frame and the final end-systolic frame, and outputs the candidate source, candidate reliability score, phase position reliability, and manual verification prompts. For candidates with a large degree of divergence D and ensemble prediction error... High, segmentation confidence Lower or boundary risk For higher-resolution video sequences, the system generates a review prompt for further manual verification.

[0079] The training data for the candidate quality model consists of candidate periodic features and their localization errors in the training partitions. The evaluation partitions do not participate in model fitting or strategy selection. During the inference phase, only candidate features are input, and manually labeled end-diastolic and end-systolic frames are not read, thus ensuring the independence of the inference process.

[0080] To evaluate the case of whole-cycle misalignment, the period-aware equivalent frame error can also be calculated. Let the predicted frame be... The labeled frame is t, and the estimated period length is . The periodic sensing equivalent frame error is expressed by equation (5-11): (5-11) Where k is an integer. Strict frame error is used to evaluate the direct difference between the predicted frame and the original labeled frame, while period-aware equivalent frame error is used to identify the error situation when the prediction result falls within adjacent equivalent cardiac cycles. The above evaluation metrics are only used for offline evaluation and do not change the actual inference process of the system. During calculation, k is successively taken as an integer that ensures the frame position after a full-cycle shift remains within the video frame range. The absolute difference between the predicted frame after k estimated cycles of shift and the manually labeled frame is calculated, and the minimum value is taken as the period-aware equivalent frame error. When k=0, the calculation result corresponds to the strict frame error between the predicted frame and the original labeled frame; when the prediction result is located within adjacent equivalent cardiac cycles, a positive or negative integer k is used to eliminate the full-cycle shift.

[0081] Strict frame error refers to the absolute difference in frame index between a single ED predicted frame or ES predicted frame and its corresponding manually annotated frame; strict average frame error refers to the arithmetic mean of the ED strict frame error and the ES strict frame error of the same video.

[0082] The following experiments verify the effectiveness of the method provided in the embodiments of the present invention. The experiments of the present invention use the EchoNet-Dynamic public dataset released by Stanford University, which contains 10,030 112×112 pixel apical four-chamber echocardiography videos and provides manual recordings of left ventricular ejection fraction and ED / ES keyframes, which can be used for left ventricular segmentation and cardiac function estimation studies and verification.

[0083] The experiments utilized echocardiographic left ventricular segmentation sequences, candidate cycle data, and manually labeled information, with a fixed evaluation set of 629 videos. To ensure comparability between different methods, strict frame error, ED error, ES error, and cycle-aware error were all uniformly re-aggregated from video-by-video output; the median peak-to-peak candidate cycle length was used as the equivalent cycle for cycle-aware error. Comparison methods were named according to their actual technical modules used, rather than internal version numbers from the development process.

[0084] Table 1

[0085] Note: All frame localization metrics and area-proximity LVEF metrics were re-aggregated on the same 629 evaluation videos. Periodicity-aware error used a uniform median peak-to-peak candidate period length as the equivalent period. The area-proximity LVEF for each method was calculated independently on the same original left ventricular area curve based on the predicted ED / ES frames of that method, and a uniform linear calibration coefficient fitted only from the training set was applied.

[0086] Table 2

[0087] Table 2 provides supplementary evaluations of the area-surrogate LVEF results of the integrated candidate quality re-ranking method from four aspects: average error, error dispersion, correlation, and consistency. These results are calculated independently based on the original left ventricular area corresponding to the predicted ED / ES frames, using the same training set calibration coefficients and evaluation video set as other methods.

[0088] like Figure 7 As shown in Table 1, with the gradual introduction of phase probability, temporal confidence, multi-candidate consensus, divergence-aware adaptive strategy, and candidate quality re-ranking, the strict average frame error (the arithmetic mean of the strict frame errors of ED and ES) decreased from 33.2043 frames to 29.2806 frames. Specifically, the strict frame error of ED decreased from 34.2035 frames to 28.9348 frames, and the strict frame error of ES decreased from 32.2051 frames to 29.6264 frames. The area proxy LVEF MAE recalculated independently by each method did not decrease monotonically in sync with the strict frame error. The area proxy LVEF MAE of the divergence-aware adaptive consensus method was the lowest, at 7.1310 percentage points.

[0089] The integrated candidate quality reordering method, applied across 629 evaluation videos, achieved a strict mean frame error of 29.2806 frames, an end-diastolic frame error of 28.9348 frames, an end-contraction frame error of 29.6264 frames, a uniform caliber period-aware error of 5.3609 frames, and an area-proxy LVEF MAE of 7.9915 percentage points, all of which are average across all evaluation videos. Its strict mean frame error is significantly higher than the recall candidate and prior calibration methods, which reduce the error by 3.9237 frames (11.82%). However, its area-proxy LVEF MAE is higher than the divergence-aware adaptive consensus method, indicating that the absolute frame index error and the area-proxy LVEF error are not perfectly synchronized.

[0090] As shown in Table 2, the integrated candidate quality re-ranking method, based on the area proxy LVEF calculated independently for its predicted ED / ES frames, achieved the following results on the same evaluation set: MAE 7.9915 percentage points, RMSE 10.6298 percentage points, median absolute error 6.3661 percentage points, Pearson correlation coefficient 0.6546, coefficient of determination 0.2702, Bland-Altman bias -1.3159 percentage points, and 95% consensus limits [-22.0065, 19.3746] percentage points. These results indicate that different periodic alignment methods should be evaluated separately from two dimensions: keyframe localization error and area proxy LVEF error.

[0091] like Figure 8As shown, the error distribution still contains a large number of long-tailed samples, indicating that this system is suitable for reducing overall localization error and providing risk warnings, but candidate quality scores should not be interpreted as clinical diagnostic conclusions. For samples with high discrepancies, high prediction errors, or low segmentation confidence, the system outputs a manual review prompt.

[0092] The above embodiments are only used to provide a detailed description of the technical solutions of this application. However, the descriptions of the above embodiments are only for helping to understand the methods of the embodiments of the present invention and should not be construed as limiting the embodiments of the present invention. Variations or substitutions that can be easily conceived by those skilled in the art should be covered within the protection scope of the embodiments of the present invention.

Claims

1. A method for reordering candidate cardiac cycles in echocardiography, characterized in that, include: S1. Obtain the left ventricular probability map sequence or binary mask sequence of continuous echocardiographic video, and construct the left ventricular area curve frame by frame; S2. Generate a set of candidate cycles based on the left ventricular area curve, and determine the initial ED frame and initial ES frame for each candidate cycle; S3. For each candidate cardiac cycle in the candidate cardiac cycle set, obtain the candidate reliability score; S4. Within each candidate period, construct feature vectors to obtain frame-level ED phase probability and ES phase probability, and combine relative position and statistical position priors for fusion scoring to obtain ED and ES keyframes. S5. Based on the degree of disagreement of the keyframes corresponding to the candidate periods with high candidate reliability scores, adaptively determine the number of candidates participating in the consensus calculation, and calculate the weighted median of the keyframe index according to the candidate reliability weights to obtain the ED and ES consensus keyframes. S6. Construct multiple candidate error prediction models with different random initialization parameters to obtain the integrated prediction error of the candidate period. Use the consensus key frame corresponding to the candidate period with the smallest integrated prediction error as the final ED frame and the final ES frame.

2. The method according to claim 1, characterized in that, The candidate cardiac cycle set is as follows: ， in, This represents the set of candidate periods between peaks. This represents a set of candidate periods for a sliding window, constructed starting from a local maximum frame and according to various reference period ratios. This represents the set of candidate boundary completions generated for the truncation period at the beginning or end of the video. The initial ED frame and the initial ES frame are the frames with the largest and smallest areas within the corresponding candidate cardiac cycle, respectively.

3. The method according to claim 2, characterized in that, The candidate reliability score is: ， Supervised ranking score To prefit the obtained candidate cycle supervised ranking model for the first One candidate cardiac cycle output; Annotated prior scores Calculated based on the empirical distribution of candidate number, relative center position, and period length deviation in the fitted data. , , and These represent the probabilities of the pre-sorted sequence number, relative center position, and cycle length deviation of the i-th candidate cardiac cycle in the empirical distribution of the fitted data, respectively. Normative score ， Indicates the original score. and They represent the first Initial ED frame index and initial ES frame index for each candidate cardiac cycle. For the first Frame segmentation confidence; This represents a very small positive number used to prevent the denominator from being zero; and These are the proportions of inter-frame differences that satisfy the correct direction of change during the diastolic-to-contraction descent phase and the post-contraction recovery phase, respectively; video boundary risk. ; and They represent the first The start frame index and end frame index of each candidate cardiac cycle; This is the normalized distance between the candidate area curve and the standard periodic template. Curve shape score , Center position score ; For the corresponding weights.

4. The method according to claim 3, characterized in that, The candidate periodicity supervised ranking model is determined using only fitted data: Candidate periods that simultaneously contain manually labeled ED frames and manually labeled ES frames are used as positive samples, while candidate periods that do not simultaneously contain manually labeled ED frames and manually labeled ES frames are used as negative samples. A supervised ranking model for candidate periods is obtained by performing linear fitting with ridge regularization constraints based on the ranking position, relative position of the center, period length deviation, curve shape, segmentation confidence, and boundary risk.

5. The method according to claim 3, characterized in that, S4 include: Construct feature vectors; The components in the feature vector include the normalized area of ​​the smooth area curve, the first-order change and the second-order change, the stationary point score, the peak curvature score and the valley curvature score, the segmentation confidence, the segmentation uncertainty, the relative temporal position of the whole video, the video edge distance, and the mean, standard deviation and range of the normalized area within the local window. Based on a pre-trained weighted logistic regression phase model, the frame-level probabilities of ED and ES phases are obtained, and the preset number of video frames with the highest frame-level probabilities are selected as candidate ED and ES frames. For any candidate ED and ES frames, calculate the relative and statistical position priors within the candidate cardiac cycle, respectively. Calculate the fusion score for ED / ES frame pairs.

6. The method according to claim 5, characterized in that, The fusion score is: ， in, Let be the frame-level probabilities of the ED phase and ES phase in frame t. These are the bias terms used in pre-training. The weight vector of the feature vector. Let be the normalized feature vector of the t-th frame; Candidate ED and ES frames in candidate cardiac cycle The relative position within, For statistical location priors; Keyframe timing constraints .

7. The method according to claim 6, characterized in that, ED and ES frames and The results were obtained by weighted logistic regression training using the fitted data, specifically as follows: During training, soft labels are generated using a Gaussian function with a standard deviation of 4 frames, centered on manually annotated ED or ES keyframes, and soft labels with a standard deviation less than 0.05 are set to 0. Optimization is performed using class-balanced weights and a L2 regularization term with a coefficient of 0.

01. The format is fixed after training. and .

8. The method according to claim 3, characterized in that, S5 include: Candidates are sorted from highest to lowest reliability score, and the degree of divergence is calculated for the highest-ranked preset cardiac cycles. Based on the degree of divergence among the candidates, the candidate heartbeat cycles that actually participate in the consensus calculation are determined; For the identified candidate cardiac cycles, calculate the normalized reliability weights and the weighted median of ED and ES to determine the consensus keyframes.

9. The method according to claim 8, characterized in that, S6 include: Multiple candidate error prediction models with different random initialization parameters are constructed, and the input features of each candidate period are input into the model. The average value of the outputs of multiple models is used as the integrated prediction error corresponding to the candidate period. The candidate period corresponding to the minimum integrated prediction error is taken as the final candidate period, and the ED and ES consensus key frames corresponding to the final candidate period are output as the final ED frame and the final ES frame.

10. The method according to claim 9, characterized in that, The candidate error prediction model uses a multilayer perceptron with two hidden layers. The network structure includes, in sequence: an input layer, a fully connected layer with 64 neurons, a ReLU activation layer, a Dropout layer with a dropout rate of 0.08, a fully connected layer with 32 neurons, a ReLU activation layer, and an output layer with 1 neuron. For any candidate error prediction model, for the first The training objective corresponding to each candidate cardiac cycle is: Each candidate error prediction model is trained by minimizing the mean square error between the model output and the above training objective; and For the first ED and ES consensus keyframes for each candidate heartbeat cycle; and These are real keyframes annotated manually.

Images