A postoperative semen quality dynamic evaluation method based on ultrasonic imageomics

Abstract

The present application relates to the technical field of medical image intelligent analysis, and more particularly to a postoperative semen quality dynamic evaluation method based on ultrasonic imageomics, comprising the following steps: collecting testicular ultrasound images of a target object in two preset time periods before and after operation, segmenting the testicular parenchymal region of each image, respectively extracting imageomics features and deep features, and establishing a feature set corresponding to each time point; calculating the feature difference value at different time points to construct a dynamic difference feature set; inputting the feature set into a trained prediction model to generate a postoperative semen quality recovery evaluation value; calculating the contribution degree of each feature corresponding to the evaluation value, and outputting the evaluation value and contribution degree visualization information. The present application calculates the time series difference value of imageomics and deep features in the two time periods before and after operation, constructs a time series change dynamic difference feature set for model prediction, solves the defect that the existing postoperative evaluation can only detect a single time point and cannot capture tissue dynamic repair and early and long-term recovery prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image analysis technology, and in particular to a method for dynamic assessment of postoperative semen quality based on ultrasound radiomics. Background Technology

[0002] For patients with non-obstructive azoospermia who undergo surgical treatments such as microsurgical sperm retrieval, postoperative follow-up of semen quality recovery is a key clinical focus. Currently, clinical assessment of postoperative semen quality primarily relies on routine semen analysis at fixed outpatient clinic visits after surgery, or on conventional static ultrasound examinations such as two-dimensional and color Doppler to assess testicular volume and blood supply. With the application of computer vision technology in medical imaging, some studies have begun to explore extracting static radiomics texture features from single ultrasound images, or directly constructing convolutional neural network models to assist in assessing testicular spermatogenesis by mapping features from single-period testicular parenchyma images.

[0003] Current postoperative assessments mostly rely on semen analysis or static ultrasound examination at a single postoperative time point. Because they cannot capture the dynamic physiological process of tissue repair over time, they make it difficult to predict the long-term recovery potential in the early postoperative period, resulting in inaccurate predictions. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a method for dynamic assessment of postoperative semen quality based on ultrasound radiomics. It aims to improve the problem that most existing postoperative assessments rely on semen analysis or static ultrasound examination at a single postoperative time point, which can easily lead to inaccurate predictions due to the difficulty in predicting the long-term recovery possibility in the early postoperative period.

[0005] This invention provides the following technical solution: a method for dynamic assessment of postoperative semen quality based on ultrasound radiomics, comprising the following steps: S1. Obtain testicular ultrasound images of the target subject before surgery, during the first preset time period after surgery, and during the second preset time period after surgery; S2. The testicular parenchyma region in each of the testicular ultrasound images is segmented, and radiomics features are extracted and deep features are extracted using a deep learning model to obtain the radiomics feature set and deep feature set corresponding to each time point. S3. By calculating the differences between radiomics features and depth features at different time points, a dynamic difference feature set is generated. The dynamic difference feature set reflects the changes from the first preset time period before surgery to the first preset time period after surgery and the changes from the first preset time period after surgery to the second preset time period after surgery. S4. Input the dynamic difference feature set into the trained prediction model to obtain an assessment value indicating the likelihood of postoperative semen quality recovery. S5. Determine the contribution of each feature in the dynamic difference feature set to the evaluation value, and output the evaluation value and visualization information representing the contribution.

[0006] By employing the above technical solution, the temporal differences between radiomics features and depth features corresponding to the preoperative, postoperative first preset time period, and postoperative second preset time period are calculated. A dynamic difference feature set reflecting the temporal changes of the two stages is generated and input into the model for prediction. This improves upon the current postoperative assessment method, which mostly uses semen analysis or static ultrasound examination at a single postoperative time point. Because it cannot capture the dynamic physiological process of tissue repair over time, it is difficult to predict the long-term recovery possibility in the early postoperative period, resulting in inaccurate predictions.

[0007] Further, S1 includes the following steps: Preoperative testicular ultrasound images of the target subject were acquired within 48 hours before the operation. The preoperative testicular ultrasound images included three different DICOM (Digital Imaging and Communications in Medicine) format images, and the B-mode gain was set to 50% and the scanning depth was set to 4cm when the preoperative testicular ultrasound images were acquired. The testicular ultrasound images of the target subject were acquired during the first preset time period after surgery, from day 14 to day 21 after surgery, and during the second preset time period after surgery, from day 28 to day 35 after surgery. If the area of ​​hematoma or effusion in the testicular ultrasound image during the first preset time period after surgery or the testicular ultrasound image during the second preset time period after surgery is greater than or equal to 30% of the testicular parenchyma area, the corresponding testicular ultrasound image of the target object will not be included in the subsequent processing steps.

[0008] Further, in S2, the step of segmenting the testicular parenchyma region in each of the said testicular ultrasound images includes: A region of interest is drawn along the edge of the testicular parenchyma in each of the aforementioned testicular ultrasound images, avoiding the tunica albuginea region during the drawing process, to obtain the corresponding testicular parenchyma region of interest at each time point. Randomly select 20% of the testicular ultrasound images of the target subjects, and perform secondary region of interest delineation to obtain the corresponding contrast region of interest; The overlap coefficient between the testicular parenchyma region of interest and the contrasting region of interest is calculated. When the overlap coefficient is greater than 0.85, the testicular parenchyma region of interest is determined to have passed the consistency verification and is allowed to be included in subsequent feature extraction.

[0009] Furthermore, in S2, the step of extracting radiomics features includes: From the testicular parenchyma region of interest that has passed consistency verification, 100 to 150 hand-designed radiomics features, including first-order statistical features, shape features, gray-level co-occurrence matrix features, and gray-level run-length matrix features, are extracted. Calculate the intragroup correlation coefficient of the extracted hand-designed image omics features; The hand-designed imagemic features with an intragroup correlation coefficient greater than 0.8 are retained to form the imagemic feature set corresponding to each time point.

[0010] Furthermore, in S2, the step of extracting deep features using a deep learning model includes: The image blocks of the region of interest of the testicular parenchyma at each corresponding time point are uniformly scaled to a preset image size; The scaled region of interest image block of the testicular parenchyma is input into a pre-trained convolutional neural network model for forward feature mapping; The deep feature vector before the last fully connected layer in the convolutional neural network model is extracted, and after calculating the intra-group correlation coefficient of the deep feature vector, the feature dimensions with an intra-group correlation coefficient greater than 0.8 are retained to obtain the deep feature set corresponding to each time point.

[0011] Furthermore, in S3, the step of calculating the differences between image omics features and the differences between depth features at different time points includes: From the image omics feature set, the entropy features, homogeneity features, and correlation features of the gray-level co-occurrence matrix, as well as the short run emphasis features and long run emphasis features of the gray-level run matrix, are extracted as anchored handcrafted features. The difference between the radiomics feature set corresponding to the first preset time period after surgery and the radiomics feature set corresponding to the preoperative period is calculated to obtain the first phase of manual differential features. The difference between the radiomics feature set corresponding to the second preset time period after surgery and the radiomics feature set corresponding to the first preset time period after surgery is calculated to obtain the second phase of manual differential features. The difference between the depth feature set corresponding to the first preset time period after surgery and the depth feature set corresponding to the preoperative period is calculated to obtain the first-stage depth difference feature. The difference between the depth feature set corresponding to the second preset time period after surgery and the depth feature set corresponding to the first preset time period after surgery is calculated to obtain the second-stage depth difference feature. The first phase manual differential features, the second phase manual differential features, the first phase deep differential features, the second phase deep differential features, and the target subject's baseline clinical history information are spliced ​​together to generate the dynamic differential feature set. The baseline clinical history information includes age, testicular volume, and serum FSH (Follicle-Stimulating Hormone) level.

[0012] Further, in S4, the step of inputting the dynamic difference feature set into the trained prediction model includes: The dynamic difference feature set is input into a pre-trained gradient boosting tree classifier, which divides the sample data into a training set, a validation set, and an independent test set in a 7:1:2 ratio during the training phase. During the training phase, the gradient boosting tree classifier uses an oversampling algorithm to oversample only the samples in the training set that have a positive semen recovery outcome, while maintaining the original sample ratio between the validation set and the independent test set. During the training phase, the gradient boosting tree classifier performs grid search on the validation set to optimize hyperparameters using the area under the precision-recall curve as the metric for hyperparameter tuning, and establishes a probability decision threshold under the constraint that the negative prediction value is greater than 95%. The gradient boosting tree classifier with optimized hyperparameters is used to perform classification inference on the input dynamic differential feature set, and the evaluation value is calculated and output.

[0013] Furthermore, in S5, the step of outputting the evaluation value and the visualization information representing the contribution includes: The marginal contribution of each feature in the dynamic difference feature set to the evaluation value is calculated using a game theory-based attribution algorithm. The calculated marginal contribution values ​​are then sorted in descending order to select the top three core difference features. When the evaluation value is greater than or equal to a preset probability decision threshold, text information is generated; The absolute scores corresponding to the core difference features are mapped into bar charts that show the direction and intensity of the feature's influence. The spatial texture sub-regions corresponding to the core difference features on the testicular ultrasound images acquired during the second preset time period after the operation are determined, and a heat map based on the marginal contribution is superimposed on the testicular ultrasound images acquired during the second preset time period after the operation. The evaluation value, the text information, the bar chart, and the heat map are output.

[0014] The present invention has the following beneficial effects: 1. In this invention, by calculating the temporal differences between radiomics features and depth features corresponding to the first preset time period before surgery and the second preset time period after surgery, a dynamic difference feature set reflecting the temporal changes of the two stages is generated and input into the model for prediction. This improves the problem that most existing postoperative assessments rely on semen analysis or static ultrasound examination at a single postoperative time point. Since these methods cannot capture the dynamic physiological process of tissue repair over time, it is difficult to predict the long-term recovery possibility in the early postoperative period, resulting in inaccurate predictions.

[0015] 2. In this invention, the manually designed radiomics features of the region of interest are extracted and the deep features mapped by the convolutional neural network are obtained. The stability of the two types of features is filtered by the intragroup correlation coefficient. Then, the feature difference between the two stages is calculated and spliced ​​with clinical history information. This improves the problem that most existing feature extraction methods use full-image features or unfiltered single-modal features. These features are easily affected by operation methods or equipment noise and ignore the influence of clinical physiological baseline, resulting in insufficient feature robustness and incomplete data links.

[0016] 3. In this invention, by using an attribution algorithm to calculate the marginal contribution of features to the evaluation value, the absolute scores corresponding to the core difference features are mapped into a bar chart. At the same time, the bar chart is mapped inversely and superimposed onto the ultrasound images of the second preset time period after surgery to display a heat map based on the marginal contribution. This improves the problem that existing models mostly predict outputs by directly giving probability scores or labels. Due to the lack of interpretable means for the basis of model decision-making, clinicians cannot intuitively know the core ultrasound indicators that lead to poor recovery. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of a postoperative semen quality dynamic assessment method based on ultrasound radiomics according to the present invention. Figure 2 This is a schematic diagram of the process for acquiring multi-time-point testicular ultrasound images and segmenting the testicular parenchyma region in this invention. Figure 3 This is a schematic diagram of the process for extracting radiomics features and depth features and screening stability in this invention; Figure 4 This is a schematic diagram of the dynamic difference feature set generation process in this invention; Figure 5 This is a schematic diagram of the postoperative semen quality recovery probability prediction and visualization attribution process based on gradient boosting tree classifier in this invention; Figure 6 This is a schematic diagram of the module architecture of a postoperative semen quality dynamic assessment system based on ultrasound radiomics in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: In the first embodiment of the present invention, a method for dynamic assessment of postoperative semen quality based on ultrasound radiomics is provided, such as... Figure 1 As shown, the procedure includes the following steps: S1, acquiring testicular ultrasound images of the target subject before surgery, during the first preset time period after surgery, and during the second preset time period after surgery; like Figure 2 As shown, S1 further includes the following steps: Preoperative testicular ultrasound images of the target subject were acquired within 48 hours before the operation. The preoperative testicular ultrasound images included DICOM format images of three different sections. The B-mode gain was set to 50% and the scanning depth was set to 4cm during the acquisition of the preoperative testicular ultrasound images. Acquire testicular ultrasound images of the target subject during the first pre-defined postoperative time period from day 14 to day 21 after surgery, and testicular ultrasound images during the second pre-defined postoperative time period from day 28 to day 35 after surgery. If the area of ​​hematoma or effusion in the testicular ultrasound image at the first or second pre-specified time interval after surgery is greater than or equal to 30% of the testicular parenchyma area, the corresponding testicular ultrasound image of the target subject will not be included in the subsequent processing steps.

[0020] Specifically, preoperative testicular ultrasound images of the target subject were acquired within 48 hours prior to surgery. These images included DICOM format images with three different cross-sections, and the B-mode gain was set to 50% and the scanning depth to 4 cm during acquisition. Postoperative testicular ultrasound images were also acquired for the first pre-defined time period (days 14-21) and the second pre-defined time period (days 28-35). If the area of ​​hematoma or effusion in the first or second pre-defined time period postoperative ultrasound image was greater than or equal to 30% of the testicular parenchyma area in that image, the corresponding testicular ultrasound image of the target subject was not included in subsequent processing steps. Baseline clinical history information of the target subject, including age, testicular volume, and serum FSH level, was recorded and associated with the corresponding testicular ultrasound images at each time point.

[0021] S2. Segment the testicular parenchyma region in each testicular ultrasound image, extract radiomics features and use a deep learning model to extract deep features, and obtain the radiomics feature set and deep feature set corresponding to each time point. Furthermore, in S2, the step of segmenting the testicular parenchyma region in each testicular ultrasound image includes: The region of interest was delineated along the edge of the testicular parenchyma in each testicular ultrasound image, avoiding the tunica albuginea region during the delineation, to obtain the corresponding testicular parenchyma region of interest at each time point. Randomly select 20% of the target subjects' testicular ultrasound images, perform secondary region of interest delineation, and obtain the corresponding contrast region of interest; The overlap coefficient between the testicular parenchyma region of interest and the contrast region of interest is calculated. When the overlap coefficient is greater than 0.85, the testicular parenchyma region of interest is determined to have passed the consistency verification and is allowed to be included in subsequent feature extraction.

[0022] Specifically, a region of interest (ROI) is drawn along the edge of the testicular parenchyma in each testicular ultrasound image, avoiding the tunica albuginea region, to obtain the corresponding ROI for each time point. Testicular ultrasound images of 20% of the target subjects are randomly selected, and another operator independently draws a second ROI to obtain the corresponding contrast ROI. The overlap coefficient between the ROI and the contrast ROI is calculated. In this embodiment, the overlap coefficient uses the Dice coefficient, and its calculation formula is as follows: Where A represents the set of pixels in the region of interest of the testicular parenchyma, B represents the set of pixels in the region of interest for comparison, and |A| and |B| represent the total number of pixels in the corresponding sets, respectively; when the calculated D>0.85, the region of interest of the testicular parenchyma is determined to have passed the consistency verification and is allowed to be included in subsequent feature extraction.

[0023] like Figure 3 As shown, further, in S2, the steps for extracting radiomics features include: From the region of interest of testicular parenchyma that has passed consistency verification, 100 to 150 hand-designed radiomics features, including first-order statistical features, shape features, gray-level co-occurrence matrix features, and gray-level run-length matrix features, are extracted. Calculate the intragroup correlation coefficients of the extracted hand-designed image omics features; Manually designed imagemic features with an intragroup correlation coefficient greater than 0.8 were retained to form imagemic feature sets corresponding to each time point.

[0024] Specifically, 100 to 150 hand-designed radiomics features, including first-order statistical features, shape features, gray-level co-occurrence matrix features, and gray-level run-length matrix features, are extracted from the regions of interest (ROIs) of the testicular parenchyma that have passed consistency verification. Features are extracted separately using two ROI delineations that have passed consistency verification, and the intragroup correlation coefficient of the same feature between the two extractions is calculated using the following formula: ;in Indicates the mean square among target objects. In this scheme, the mean square within the target object is represented. ;reserve The hand-designed image omics features constitute the image omics feature sets corresponding to each time point.

[0025] Furthermore, in S2, the steps for extracting deep features using a deep learning model include: The image blocks of the region of interest of the testicular parenchyma at each time point are uniformly scaled to the preset image size; The scaled region of interest image of the testicular parenchyma is input into a pre-trained convolutional neural network model for forward feature mapping; The deep feature vectors before the last fully connected layer in the convolutional neural network model are extracted. After calculating the intra-group correlation coefficient of the deep feature vectors, the feature dimensions with an intra-group correlation coefficient greater than 0.8 are retained to obtain the deep feature sets corresponding to each time point.

[0026] Specifically, the region of interest (ROI) image blocks of the testicular parenchyma at each time point are uniformly scaled to a preset image size, preferably 224×224 pixels in this embodiment. The scaled ROI image blocks of the testicular parenchyma are then input into a pre-trained convolutional neural network (CNN) model for forward feature mapping, using the ResNet50 architecture in this embodiment. The depth feature vector before the last fully connected layer in the CNN model is extracted, preferably 1024-dimensional in this embodiment. The depth feature vector is extracted using two ROI delineations that have passed consistency verification, and the intra-group correlation coefficient of each feature dimension between the two extractions is calculated using the aforementioned intra-group correlation coefficient calculation formula. Feature dimensions with ICC > 0.8 are retained to obtain the depth feature set corresponding to each time point.

[0027] S3. By calculating the differences between radiomics features and depth features at different time points, a dynamic difference feature set is generated. The dynamic difference feature set reflects the changes from the first preset time period before surgery to the first preset time period after surgery and the changes from the first preset time period after surgery to the second preset time period after surgery. like Figure 4 As shown, further, in S3, the steps for calculating the differences between image omics features and the differences between depth features at different time points include: From the image omics feature set, we extract the entropy features, homogeneity features, and correlation features of the gray-level co-occurrence matrix, as well as the short run emphasis features and long run emphasis features of the gray-level run matrix, as anchor handcrafted features. The difference between the radiomics feature set corresponding to the first preset time period after surgery and the radiomics feature set corresponding to the preoperative period is calculated to obtain the first phase of manual differential features. The difference between the radiomics feature set corresponding to the second preset time period after surgery and the radiomics feature set corresponding to the first preset time period after surgery is calculated to obtain the second phase of manual differential features. The difference between the depth feature set corresponding to the first preset time period after surgery and the corresponding depth feature set before surgery is calculated to obtain the first-stage depth difference feature. The difference between the depth feature set corresponding to the second preset time period after surgery and the depth feature set corresponding to the first preset time period after surgery is calculated to obtain the second-stage depth difference feature. The first phase of manual differential features, the second phase of manual differential features, the first phase of in-depth differential features, the second phase of in-depth differential features, and the baseline clinical history information of the target subjects are spliced ​​together to generate a dynamic differential feature set. The baseline clinical history information includes age, testicular volume, and serum FSH level.

[0028] Specifically, from the radiomics feature set, entropy, homogeneity, and correlation features of the gray-level co-occurrence matrix, as well as short-run and long-run emphasis features of the gray-level run matrix, are extracted as anchored manual features. Z-score standardization is applied to each feature dimension in both the radiomics feature set and the deep feature set to ensure zero mean and unit variance for each dimension. The difference between the radiomics feature set corresponding to the first pre-set postoperative time period and the corresponding preoperative radiomics feature set is calculated to obtain the first-stage manual difference feature. Similarly, the difference between the radiomics feature set corresponding to the second pre-set postoperative time period and the corresponding first pre-set postoperative time period is calculated to obtain the second-stage manual difference feature. Features: The difference between the depth feature set corresponding to the first preset time period after surgery and the corresponding depth feature set before surgery is calculated to obtain the first-stage depth difference feature. The difference between the depth feature set corresponding to the second preset time period after surgery and the depth feature set corresponding to the first preset time period after surgery is calculated to obtain the second-stage depth difference feature. The difference is calculated by subtracting the standardized feature vectors at the corresponding time points element by element. After Z-score standardization of the target subject's age, testicular volume, and serum FSH level, the features are horizontally concatenated with the first-stage manual difference feature, the second-stage manual difference feature, the first-stage depth difference feature, and the second-stage depth difference feature according to the feature dimension to generate a dynamic difference feature set.

[0029] like Figure 5As shown in Figure S4, input the dynamic difference feature set into the trained prediction model to obtain an assessment value indicating the likelihood of postoperative semen quality recovery. Furthermore, in S4, the step of inputting the dynamic difference feature set into the trained prediction model includes: The dynamic difference feature set is input into a pre-trained gradient boosting tree classifier. During the training phase, the gradient boosting tree classifier divides the sample data into training set, validation set and independent test set in a 7:1:2 ratio. During the training phase, the gradient boosting tree classifier uses an oversampling algorithm to oversample only the samples with positive semen recovery outcomes in the training set, while maintaining the original sample ratio between the validation set and the independent test set. During the training phase, the gradient boosting tree classifier performs grid search on the validation set using the area under the precision-recall curve as the metric for hyperparameter tuning to optimize hyperparameters, and establishes a probability decision threshold under the constraint that the negative prediction value is greater than 95%. The gradient boosting tree classifier with optimized hyperparameters is used to perform classification inference on the input dynamic differential feature set and calculate the output evaluation value.

[0030] Specifically, the gold standard for semen recovery outcome labeling is based on two microscopic examinations of semen sediment at 6 months post-surgery. The presence of sperm in either examination is considered positive, while the absence of sperm in both examinations is considered negative. The dynamic difference feature set is input into a pre-trained gradient boosting tree classifier. In this embodiment, the XGBoost algorithm is used. During the training phase, the gradient boosting tree classifier hierarchically divides the sample data into training, validation, and independent test sets in a 7:1:2 ratio, based on the semen recovery outcome label. During training, an oversampling algorithm is used only for samples with a positive semen recovery outcome in the training set. In this embodiment, the SMOTE algorithm is used, which selects the nearest neighbor samples of the positive samples in the feature space and generates new samples through linear interpolation. The interpolation formula is: ;in This is the feature vector of the original positive sample. for The feature vector of random nearest neighbor samples, for Random numbers within the interval; the original sample ratio of the validation set and independent test set remains unchanged during the training phase; during the training phase, the area under the precision-recall curve is used as the indicator for hyperparameter tuning on the validation set, and the hyperparameters are optimized through grid search, and a probability decision threshold is established under the constraint that the negative prediction value is greater than 95%; the gradient boosting tree classifier after optimizing the hyperparameters is used to perform classification inference on the input dynamic differential feature set, and the output predicted probability value is used as the evaluation value.

[0031] S5. Determine the contribution of each feature in the dynamic difference feature set to the evaluation value, and output the evaluation value and the visualization information of the contribution. Furthermore, in S5, the steps for outputting evaluation values ​​and visualizations representing contributions include: The marginal contribution of each feature in the dynamic difference feature set to the evaluation value is calculated using a game theory-based attribution algorithm. The calculated marginal contributions are then sorted in descending order to select the top three core difference features. When the evaluation value is greater than or equal to a preset probability decision threshold, text information is generated; The absolute scores corresponding to the core differential features are mapped into bar charts that show the direction and intensity of the feature's influence. The spatial texture sub-regions corresponding to the core differential features on the testicular ultrasound images acquired at the second preset time period after surgery are identified, and a heat map based on marginal contribution is superimposed on the testicular ultrasound images acquired at the second preset time period after surgery. The evaluation value, text information, bar chart and heat map are output.

[0032] Specifically, a game theory-based attribution algorithm is used to calculate the marginal contribution of each feature in the dynamic difference feature set to the evaluation value. In this embodiment, the attribution algorithm adopts the SHAP algorithm, and its calculation method is as follows: ;in Let i represent the Shapley value of feature i, F be the set of all features, S be the subset of features that does not contain feature i, and |S|! and |F|! represent the factorial of the cardinality of the corresponding sets, respectively. The model's predicted output is obtained by using only a subset of features S. The calculated marginal contributions are sorted in descending order to select the top three core differential features. Text information is generated when the evaluation value is greater than or equal to a preset probability decision threshold. The Shapley values ​​corresponding to the core differential features are mapped to bar charts showing the direction and magnitude of feature contributions. For deep features in the core differential features, gradient-weighted class activation mapping is used to calculate the gradient weight of the feature map of the last convolutional layer of the ResNet50 architecture convolutional neural network model for that deep feature dimension. Class activation maps are generated by weighted summation and ReLU activation to determine the spatial texture sub-regions corresponding to the core differential features on the testicular ultrasound images collected in the second preset time period after surgery. A heatmap based on Shapley values ​​is superimposed on the testicular ultrasound images collected in the second preset time period after surgery, and the evaluation value, text information, bar chart, and heatmap are output simultaneously. This embodiment achieves a non-invasive assessment of semen recovery status at 6 months post-surgery based on testicular ultrasound images within 1 month post-surgery through the above steps. By employing temporal difference analysis of radiomics features and depth features, combined with physiologically anchored feature screening and a predictive model training strategy for imbalanced data, it can effectively identify patients with extremely low semen recovery probability while ensuring a high negative predictive value. At the same time, SHAP attribution analysis and Grad-CAM heatmap visualization provide interpretive support for the assessment results, assisting clinicians in making decisions to avoid unnecessary follow-up examinations.

[0033] Example 2: In the second embodiment of the present invention, the present invention provides a postoperative semen quality dynamic assessment system based on ultrasound radiomics, such as... Figure 6 As shown, it includes the following modules: The image acquisition module is used to acquire testicular ultrasound images of the target object before surgery, during the first preset time period after surgery, and during the second preset time period after surgery; The feature extraction module is used to segment the testicular parenchyma region in the testicular ultrasound image, extract and output the radiomics feature set and depth feature set corresponding to each time point; The differential feature generation module is used to calculate the differences between the image omics feature sets and the differences between the depth feature sets at different time points, and to generate dynamic differential feature sets. The quality prediction module is used to input the dynamic difference feature set into the prediction model and calculate an assessment value indicating the likelihood of postoperative semen quality recovery. The visualization attribution module is used to determine the contribution of each feature in the dynamic difference feature set to the evaluation value, and simultaneously output the evaluation value and visualization information representing the contribution.

[0034] In a medical center specializing in male reproductive health, clinicians face the daunting challenge of predicting the long-term semen recovery outcome in patients with non-obstructive azoospermia after microsurgical sperm retrieval. Current postoperative assessments largely rely on routine semen analysis or static ultrasound at a single postoperative time point, failing to capture the dynamic physiological processes of testicular tissue evolution and wound repair over time. This makes it difficult to accurately predict the likelihood of long-term semen quality recovery in the early postoperative period. To address this issue, this invention employs a dynamic postoperative semen quality assessment system based on ultrasound radiomics, such as... Figure 1 As shown, the specific implementation process of this system is as follows: First, the system's image acquisition module acquires testicular ultrasound images of the target subject before surgery, during the first preset time period after surgery, and during the second preset time period after surgery; this step effectively ensures that subsequent analysis can be based on continuous and uninterrupted time-series image data. Then, the system's feature extraction module is used to accurately segment the testicular parenchyma region in each acquired testicular ultrasound image, extracting and outputting the corresponding radiomics feature set and depth feature set for each time point. In this process, the module performs strict stability filtering on the two types of features through intragroup correlation coefficients, thereby improving the problem that existing feature extraction methods mostly use full-image features or unfiltered single-modal features, which are easily affected by equipment noise, and significantly improving the robustness of feature representation of changes in tissue microstructure. Next, the system's differential feature generation module calculates the differences between the radiomics feature sets and the differences between the depth feature sets at different time points, and combines them with the baseline clinical history information of the target object to perform matrix stitching to generate a dynamic differential feature set. This step successfully transforms static image data into quantitative indicators that reflect the evolution trend of tissue repair at different stages from preoperative to postoperative, and fully supplements the data link information. Subsequently, the generated dynamic difference feature set is input into the optimized gradient boosting tree classifier prediction model through the system's quality prediction module to calculate the evaluation value indicating the possibility of postoperative semen quality recovery. This prediction model makes decisions based on the dynamic evolution features at multiple time points, thereby improving the shortcomings of traditional static examinations that cannot capture the dynamic repair process, leading to inaccurate early predictions. It enables highly accurate predictions of the possibility of long-term semen quality recovery in the early postoperative period. Finally, the system's visualization attribution module determines the marginal contribution of each feature in the dynamic difference feature set to the evaluation value, and simultaneously outputs the evaluation value, follow-up guidance text information, a bar chart showing the strength of feature influence, and a contribution heatmap superimposed on ultrasound images at the second preset time period after surgery. This step improves upon the problem that existing deep model predictions mostly provide cold probability scores or labels without interpretable means of model decision-making basis, enabling clinicians to intuitively understand the core ultrasound imaging indicators that lead to poor recovery in the target subject. This provides transparent and reliable white-box clinical decision support for doctors to develop personalized and precise follow-up and treatment strategies.

[0035] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic assessment of postoperative semen quality based on ultrasound radiomics, characterized in that, Includes the following steps: S1. Obtain testicular ultrasound images of the target subject before surgery, during the first preset time period after surgery, and during the second preset time period after surgery; S2. The testicular parenchyma region in each of the testicular ultrasound images is segmented, and radiomics features are extracted and deep features are extracted using a deep learning model to obtain the radiomics feature set and deep feature set corresponding to each time point. S3. By calculating the differences between radiomics features and depth features at different time points, a dynamic difference feature set is generated. The dynamic difference feature set reflects the changes from the first preset time period before surgery to the first preset time period after surgery and the changes from the first preset time period after surgery to the second preset time period after surgery. S4. Input the dynamic difference feature set into the trained prediction model to obtain an assessment value indicating the likelihood of postoperative semen quality recovery. S5. Determine the contribution of each feature in the dynamic difference feature set to the evaluation value, and output the evaluation value and visualization information representing the contribution.

2. The method for dynamic assessment of postoperative semen quality based on ultrasound radiomics according to claim 1, characterized in that, S1 includes the following steps: Acquire preoperative testicular ultrasound images of the target subject within 48 hours before surgery. The preoperative testicular ultrasound images include DICOM format images with three different sections. The B-mode gain of the preoperative testicular ultrasound images is set to 50%, and the scanning depth is set to 4 cm. The testicular ultrasound images of the target subject were acquired during the first preset time period after surgery, from day 14 to day 21 after surgery, and during the second preset time period after surgery, from day 28 to day 35 after surgery. If the area of ​​hematoma or effusion in the testicular ultrasound image during the first preset time period after surgery or the testicular ultrasound image during the second preset time period after surgery is greater than or equal to 30% of the testicular parenchyma area, the corresponding testicular ultrasound image of the target object will not be included in the subsequent processing steps.

3. The method for dynamic assessment of postoperative semen quality based on ultrasound radiomics according to claim 1, characterized in that, In S2, the step of segmenting the testicular parenchyma region in each of the said testicular ultrasound images includes: A region of interest is drawn along the edge of the testicular parenchyma in each of the aforementioned testicular ultrasound images, avoiding the tunica albuginea region during the drawing process, to obtain the corresponding testicular parenchyma region of interest at each time point. Randomly select 20% of the testicular ultrasound images of the target subjects, and perform secondary region of interest delineation to obtain the corresponding contrast region of interest; The overlap coefficient between the testicular parenchyma region of interest and the contrasting region of interest is calculated. When the overlap coefficient is greater than 0.85, the testicular parenchyma region of interest is determined to have passed the consistency verification and is allowed to be included in subsequent feature extraction.

4. The method for dynamic assessment of postoperative semen quality based on ultrasound radiomics according to claim 3, characterized in that, In S2, the step of extracting radiomics features includes: From the testicular parenchyma region of interest that has passed consistency verification, 100 to 150 hand-designed radiomics features, including first-order statistical features, shape features, gray-level co-occurrence matrix features, and gray-level run-length matrix features, are extracted. Calculate the intragroup correlation coefficient of the extracted hand-designed image omics features; The hand-designed imagemic features with an intragroup correlation coefficient greater than 0.8 are retained to form the imagemic feature set corresponding to each time point.

5. The method for dynamic assessment of postoperative semen quality based on ultrasound radiomics according to claim 3, characterized in that, In S2, the step of extracting deep features using a deep learning model includes: The image blocks of the region of interest of the testicular parenchyma at each corresponding time point are uniformly scaled to a preset image size; The scaled region of interest image block of the testicular parenchyma is input into a pre-trained convolutional neural network model for forward feature mapping; The deep feature vector before the last fully connected layer in the convolutional neural network model is extracted, and after calculating the intra-group correlation coefficient of the deep feature vector, the feature dimensions with an intra-group correlation coefficient greater than 0.8 are retained to obtain the deep feature set corresponding to each time point.

6. The method for dynamic assessment of postoperative semen quality based on ultrasound radiomics according to claim 1, characterized in that, In S3, the step of calculating the differences between image omics features and the differences between depth features at different time points includes: From the image omics feature set, the entropy features, homogeneity features, and correlation features of the gray-level co-occurrence matrix, as well as the short run emphasis features and long run emphasis features of the gray-level run matrix, are extracted as anchored handcrafted features. The difference between the radiomics feature set corresponding to the first preset time period after surgery and the radiomics feature set corresponding to the preoperative period is calculated to obtain the first phase of manual differential features. The difference between the radiomics feature set corresponding to the second preset time period after surgery and the radiomics feature set corresponding to the first preset time period after surgery is calculated to obtain the second phase of manual differential features. The difference between the depth feature set corresponding to the first preset time period after surgery and the depth feature set corresponding to the preoperative period is calculated to obtain the first-stage depth difference feature. The difference between the depth feature set corresponding to the second preset time period after surgery and the depth feature set corresponding to the first preset time period after surgery is calculated to obtain the second-stage depth difference feature. The first phase manual differential feature, the second phase manual differential feature, the first phase deep differential feature, the second phase deep differential feature, and the target subject's baseline clinical history information are spliced ​​together to generate the dynamic differential feature set. The baseline clinical history information includes age, testicular volume, and serum FSH level.

7. The method for dynamic assessment of postoperative semen quality based on ultrasound radiomics according to claim 1, characterized in that, In S4, the step of inputting the dynamic difference feature set into the trained prediction model includes: The dynamic difference feature set is input into a pre-trained gradient boosting tree classifier, which divides the sample data into a training set, a validation set, and an independent test set in a 7:1:2 ratio during the training phase. During the training phase, the gradient boosting tree classifier uses an oversampling algorithm to oversample only the samples in the training set that have a positive semen recovery outcome, while maintaining the original sample ratio between the validation set and the independent test set. During the training phase, the gradient boosting tree classifier performs grid search on the validation set to optimize hyperparameters using the area under the precision-recall curve as the metric for hyperparameter tuning, and establishes a probability decision threshold under the constraint that the negative prediction value is greater than 95%. The gradient boosting tree classifier with optimized hyperparameters is used to perform classification inference on the input dynamic differential feature set, and the evaluation value is calculated and output.

8. The method for dynamic assessment of postoperative semen quality based on ultrasound radiomics according to claim 1, characterized in that, In S5, the step of outputting the evaluation value and the visualization information representing the contribution includes: The marginal contribution of each feature in the dynamic difference feature set to the evaluation value is calculated using a game theory-based attribution algorithm. The calculated marginal contribution values ​​are then sorted in descending order to select the top three core difference features. When the evaluation value is greater than or equal to a preset probability decision threshold, text information is generated; The absolute scores corresponding to the core difference features are mapped into bar charts that show the direction and intensity of the feature's influence. The spatial texture sub-regions corresponding to the core difference features on the testicular ultrasound images acquired during the second preset time period after the operation are determined, and a heat map based on the marginal contribution is superimposed on the testicular ultrasound images acquired during the second preset time period after the operation. The evaluation value, the text information, the bar chart, and the heat map are output.

Images