Abdominal drainage ultrasound contrast small sample segmentation method and system based on meta learning

Abstract

The invention belongs to an abdominal drainage ultrasound contrast small sample segmentation method and system based on meta-learning. The method comprises the steps of collecting multi-source data and images, analyzing target patient data to obtain patient state difference data, and preprocessing an ultrasound contrast image to obtain a standard contrast image. Performing feature extraction on the historical clinical data to obtain clinical operation difference data, and obtaining physiological influence drainage data according to the influence of the physiological state change of the target patient on the drainage area. The patient state difference data, the clinical operation difference data and the physiological influence drainage data are analyzed to obtain general segmentation features, and a meta-learning feature set is formulated in combination with a standard angiography image; constructing a meta-learning feature segmentation model, inputting a to-be-segmented ultrasound contrast image, and outputting a drainage region segmentation result; the method can adapt to physiological and pathological states and scenes of different patients, significantly reduces the segmentation boundary deviation, the false drop rate and the omission rate, and improves the accuracy of drainage region segmentation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sample segmentation, in particular to an abdominal drainage ultrasound contrast small sample segmentation method and system based on meta-learning. BACKGROUND

[0002] In the postoperative diagnosis and treatment of abdominal surgery, the placement of abdominal drainage tube is a key means to prevent abdominal effusion and infection, and ultrasound contrast technology has become a core imaging method for evaluating the state of the drainage area due to its non-invasive and real-time advantages. The ultrasound contrast image segmentation technology based on deep learning can accurately locate the drainage tube and leakage area, providing data support for clinical decision-making, but the application of this technology is limited by the two core pain points of medical annotation data scarcity and large clinical scene differences.

[0003] Existing small sample segmentation methods mostly use meta-learning framework, which uses a small amount of labeled data to complete model training through the task paradigm of support set-query set, but there are still significant defects in the abdominal drainage ultrasound contrast scene. On the one hand, existing methods only rely on the bottom features of images for segmentation, without considering the influence of individual differences of patients - the abdominal fat thickness, organ position and pathological state of different patients are significantly different, which will cause significant differences in image gray scale and boundary features of the drainage area, and the model is easy to misjudge the tissue attenuation artifact of obese patients as a leakage area, or miss small leakage points due to blurred boundaries caused by ascites, which seriously reduces the segmentation accuracy.

[0004] On the other hand, the existing meta-learning model lacks generalization ability, and the task set used in the training process does not cover the operational differences and physiological dynamic changes of real clinical scenes. In clinical ultrasound examination, there are subjective differences in the probe pressure, scanning angle and contrast agent acquisition timing of the physician, and the patient's respiratory movement and body position change will also cause real-time changes in the shape of the drainage area. However, the traditional meta-learning task set is mostly based on standardized acquisition of static images, and the model training is disconnected from the actual application scene, resulting in large fluctuations in the segmentation accuracy of the model in cross-patient, cross-device and cross-operation scenes, making it difficult to meet the actual needs of clinical diagnosis and treatment.

[0005] In addition, the existing technology lacks an effective integration mechanism for multi-source difference data, and the patient's physiological and pathological data, clinical operation parameters, and physiological dynamic influence data are mutually isolated from image features, which cannot provide differentiated segmentation strategy guidance for the model, further restricting the adaptation ability of the small sample segmentation model in clinical scenes.

[0006] Therefore, it is a key problem to be solved in the current medical imaging technology field to develop an abdominal drainage ultrasound contrast small sample segmentation method and system based on meta-learning. SUMMARY

[0007] To overcome the shortcomings of existing technologies, this invention proposes a small-sample segmentation method and system for abdominal drainage ultrasound contrast imaging based on meta-learning. This invention primarily addresses the impact of individual differences on segmentation accuracy and the insufficient generalization ability of the segmentation model.

[0008] The technical solution adopted by this invention to solve its technical problem is: the small sample segmentation method for abdominal drainage ultrasound contrast imaging based on meta-learning provided by this invention, comprising:

[0009] Collect target patient data and historical clinical data, acquire ultrasound contrast images in real time, analyze the target patient data to obtain patient status difference data, and preprocess the ultrasound contrast images to obtain standard contrast images.

[0010] Clinical operation difference data are obtained by feature extraction from historical clinical data, and physiological impact drainage data are obtained by analyzing the effect of changes in the physiological state of the target patient on the drainage area.

[0011] A differential perception feature purification method was used to analyze patient state difference data, clinical operation difference data, and physiological influence drainage data to obtain general segmentation features, and a meta-learning feature set was formulated in combination with standard contrast images.

[0012] A meta-learning segmentation model is constructed, and a meta-learning feature segmentation model is obtained by training based on the meta-learning feature set. The input is the ultrasound contrast image to be segmented, and the output is the segmentation result of the drainage area.

[0013] The present invention provides a meta-learning-based method for small-sample segmentation of abdominal drainage ultrasound contrast imaging, the steps of which include analyzing patient state difference data:

[0014] The target patients' abdominal fat thickness, organ position offset, contrast agent metabolism rate, pathological condition, and drainage tube implantation duration parameters were collected as target patient data.

[0015] Cases with the same disease and drainage type as the target patient were selected from historical clinical data, and corresponding parameter statistical features were extracted to construct benchmark reference data.

[0016] The target patient data is compared with the baseline reference data, the difference value is calculated, and the difference value is graded and labeled according to the preset standard to highlight the key difference characteristics.

[0017] By combining clinical experience and correlation analysis with historical clinical data, weights are assigned to each difference value, a comprehensive difference value is calculated, and the patient status difference data is obtained by associating and binding it with the original information of the target patient.

[0018] The present invention provides a small sample segmentation method for abdominal drainage ultrasound contrast imaging based on meta-learning, the preprocessing steps for obtaining standard contrast images include:

[0019] The ultrasound contrast images were quality controlled and screened to remove invalid images, and the speckle noise and stripe artifacts were differentially eliminated to obtain the preliminary processed images.

[0020] Histogram equalization is performed on the pre-processed image, adaptive contrast enhancement is added, and pixel value normalization is performed to obtain a normalized image.

[0021] By combining patient condition difference data, the location of the drainage area is determined from the normalized image, and the area is cropped to obtain standard angiographic image data.

[0022] The present invention provides a meta-learning-based method for small-sample segmentation of abdominal drainage ultrasound contrast imaging, which obtains clinical operational difference data through the following steps:

[0023] Historical clinical data were structured and classified from two dimensions: operational parameters and image standards, to obtain ultrasound examination operational parameters and clinical annotation data.

[0024] The ultrasound examination operation parameters were grouped according to the principle of single variable. The standard operation parameter combination corresponding to the clinical guidelines was selected as the baseline group, and the other operation parameter combinations were selected as the control group.

[0025] For each combination of operating parameters, the ultrasound contrast images corresponding to the data are used to extract image features related to the segmentation of the drainage area from the clinical annotation data as the target image examination features.

[0026] The target image inspection features of each comparison group are compared with the target image inspection features of the benchmark group, and the operational difference feature values ​​are calculated.

[0027] The operational differences are classified according to the magnitude of the operational difference characteristic values ​​to obtain the classification results, which are then linked and bound with the ultrasound examination operation parameters and operational difference characteristic values ​​to form clinical operational difference data.

[0028] The present invention provides a meta-learning-based method for small-sample segmentation of abdominal drainage ultrasound contrast imaging, the steps of which include obtaining physiologically influential drainage data:

[0029] Physiological state changes were collected from three dimensions: dynamic physiological changes, static physiological characteristics, and contrast agent metabolism. Corresponding data were then collected in conjunction with clinical practice to form an associated dataset.

[0030] Based on the associated dataset, the location features, morphological features, image grayscale features, and boundary features of the drainage area were extracted as physiological impact assessment indicators for ultrasound contrast images under different physiological conditions.

[0031] The control variable method was used to analyze the impact of changes in a single physiological state on physiological impact assessment indicators and to calculate the characteristic change rate.

[0032] Multiple linear regression analysis was used to calculate the contribution rate of each physiological factor to the physiological impact assessment index in the case of coordinated changes in multiple physiological states, and the correlation analysis results were obtained.

[0033] Based on the correlation analysis results and the characteristic change rate, the degree of influence of different physiological state changes was classified to obtain physiological influence drainage data.

[0034] The present invention provides a meta-learning-based method for small-sample segmentation of abdominal drainage ultrasound contrast imaging, the steps of which to obtain general segmentation features include:

[0035] Using standard contrast images as a carrier, corresponding patient status difference data, clinical operation difference data, and physiological impact drainage data are tagged and bound to form a multidimensional correlation set.

[0036] For patient condition differences, clinical operation differences, and physiological impact on drainage data, the degree of influence on drainage area segmentation features is quantified, and the mapping relationship of differences is determined.

[0037] A weighted attention mechanism is used to integrate the differential feature mapping relationship and select the fusion feature vectors that are relevant to the segmentation task.

[0038] A feature clustering algorithm is used to cluster the fused feature vectors, and the features corresponding to the cluster centers are used as segmentation features.

[0039] Calculate the Pearson correlation coefficient between different segmentation features, and retain the segmentation features that reach the preset coefficient threshold as general segmentation features.

[0040] The present invention provides a method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning. The steps for formulating the meta-learning feature set include:

[0041] The general segmentation features are decomposed to determine the feature types and quantification indicators to form feature codes.

[0042] For each standard contrast image, a convolutional neural network is used to extract the underlying morphological features, which are then compared with general segmentation features to calculate feature similarity. The underlying morphological features with a preset matching degree are selected as general image matching features.

[0043] The general matching features of the image are dimensionally aligned and normalized, and then bound to the feature encoding to form the image feature encoding.

[0044] Image feature encoding is divided into support set and query set according to a preset ratio, and meta-learning subtasks are constructed based on different combinations of patient differences, operational differences and physiological effects.

[0045] For each meta-learning subtask, geometric transformation enhancement and noise enhancement are performed to form a meta-learning feature set.

[0046] The present invention provides a small-sample segmentation method for abdominal drainage ultrasound contrast imaging based on meta-learning. The steps for training the meta-learning feature segmentation model include:

[0047] A three-level architecture consisting of a feature extraction network, differential feature fusion, and prototype matching segmentation is adopted to perform parameter initialization and training hyperparameter configuration to obtain a meta-learning segmentation model.

[0048] Multiple meta-learning sub-tasks are randomly sampled from the meta-learning feature set. The general image matching features of the support set are input into the meta-learning segmentation model to extract the prototype features of the diversion region and generate the support set segmentation mask.

[0049] Calculate the loss values ​​of the support set segmentation mask and the support set label mask, and update the model parameters using gradient descent.

[0050] The query set is processed using the updated model parameters to generate a query set segmentation mask, and the query set loss value is calculated.

[0051] The mean loss is calculated by summing the query set loss values ​​of all meta-learning subtasks. The initial parameters of the meta-learning segmentation model are then updated using gradient descent until the preset number of training rounds is reached, resulting in the meta-learning feature segmentation model.

[0052] The present invention provides a meta-learning-based method for small-sample segmentation of abdominal drainage ultrasound contrast imaging, the steps of which include outputting the drainage area segmentation results are as follows:

[0053] The ultrasound contrast image to be segmented is input into the feature extraction network of the meta-learning feature segmentation model to extract low-level detail features and high-level semantic features of the image as the features of the image to be segmented.

[0054] The difference data of the target patient status are analyzed by difference feature fusion to obtain the feature to be differentiated, and the feature to be image is fused with the feature to be differentiated to obtain the fused feature vector to be segmented.

[0055] Select a reference support set that is the same as the difference scene to be segmented from the meta-learning feature set, and input the corresponding standard contrast images and patient state difference data into the meta-learning feature segmentation model to extract the baseline prototype features.

[0056] Calculate the cosine similarity between the feature vector to be segmented and the baseline prototype feature vector, generate a similarity heatmap, and mark candidate diversion regions according to a preset similarity threshold to obtain a preliminary segmentation mask.

[0057] The initial segmentation mask is processed by removing noise points through morphological opening operations and then filling the holes through morphological closing operations to obtain the processed segmentation mask.

[0058] The jagged boundaries of the drainage region are corrected by the processing segmentation mask based on the edge smoothing algorithm, and the drainage region segmentation result is output.

[0059] The present invention provides a small sample segmentation system for abdominal drainage ultrasound contrast imaging based on meta-learning, comprising:

[0060] The heterogeneous data processing module is used to collect target patient data and historical clinical data, acquire ultrasound contrast images in real time, analyze the target patient data to obtain patient status difference data, and preprocess the ultrasound contrast images to obtain standard contrast images.

[0061] The differential diagnosis analysis module is used to extract features from historical clinical data to obtain clinical operation difference data, and to obtain physiological impact drainage data based on the influence of changes in the target patient's physiological state on the drainage area.

[0062] The feature set construction module is used to analyze patient state difference data, clinical operation difference data and physiological influence drainage data using the difference perception feature purification method to obtain general segmentation features, and to formulate a meta-learning feature set in combination with standard contrast images.

[0063] The segmentation output data module is used to construct a meta-learning segmentation model. It is trained based on the meta-learning feature set to obtain the meta-learning feature segmentation model. The input is the ultrasound contrast image to be segmented, and the output is the segmentation result of the drainage area.

[0064] The beneficial effects of this invention are as follows:

[0065] 1. This invention extracts universal segmentation features by fusing three types of differential data, enabling the model to adapt to different patient physiological and pathological states, different clinical operation parameters, and different dynamic physiological scenarios. This significantly reduces segmentation boundary deviation, false positive rate, and false negative rate, improving the accuracy of drainage area segmentation. A diverse set of meta-learning tasks is constructed based on differential scenarios, combined with a dual-loop training mechanism, allowing the model to learn universal segmentation capabilities that take into account multiple clinical differences. This effectively solves the problem of segmentation accuracy fluctuations across devices, physicians, and patients, adapting to complex clinical scenarios such as obesity, ascites, and respiratory movements. The model can automatically identify key differential factors affecting segmentation and assign them high weights, making the model's segmentation strategy more closely aligned with actual clinical patterns. Through morphological post-processing and boundary smoothing optimization, the output segmentation mask has low noise, clear boundaries, and a morphological fit to the actual drainage area, assisting clinicians in making diagnostic and treatment decisions regarding drainage patency and leakage risk. Attached Figure Description

[0066] The invention will now be further described with reference to the accompanying drawings.

[0067] Fig. 1 This is a flowchart illustrating the small sample segmentation method for abdominal drainage ultrasound contrast imaging based on meta-learning provided in this embodiment of the invention.

[0068] Fig. 2This is a flowchart illustrating the process of obtaining physiologically influential drainage data in the small sample segmentation method for abdominal drainage ultrasound contrast imaging based on meta-learning provided in this embodiment of the invention.

[0069] Fig. 3 This is a schematic diagram of the module of the abdominal drainage ultrasound contrast imaging small sample segmentation system based on meta-learning provided in the embodiments of the present invention. Detailed Implementation

[0070] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0071] like Figs. 1 to 3 As shown in the embodiment of the present invention, the method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning includes:

[0072] Collect target patient data and historical clinical data, acquire ultrasound contrast images in real time, analyze the target patient data to obtain patient status difference data, and preprocess the ultrasound contrast images to obtain standard contrast images.

[0073] The steps involved in analyzing patient status difference data include:

[0074] The target patients' abdominal fat thickness, organ position offset, contrast agent metabolism rate, pathological condition, and drainage tube implantation duration parameters were collected as target patient data.

[0075] Cases with the same disease and drainage type as the target patient were selected from historical clinical data, and corresponding parameter statistical features were extracted to construct benchmark reference data.

[0076] The target patient data is compared with the baseline reference data, the difference value is calculated, and the difference value is graded and labeled according to the preset standard to highlight the key difference characteristics.

[0077] By combining clinical experience and correlation analysis with historical clinical data, weights are assigned to each difference value, a comprehensive difference value is calculated, and the patient status difference data is obtained by associating and binding it with the original information of the target patient.

[0078] The steps for preprocessing to obtain standard contrast images include:

[0079] The ultrasound contrast images were quality controlled and screened to remove invalid images, and the speckle noise and stripe artifacts were differentially eliminated to obtain the preliminary processed images.

[0080] Invalid images removed include: images with missing frames or blurred images; images where the probe is not aligned with the drainage area; images with background noise exceeding 50%; and images with timing errors indicating insufficient contrast agent coverage or excessive contrast agent fade.

[0081] For speckle noise: A directional controllable filter is used for filtering. This filter can adjust the filtering kernel according to the direction of the image edge. While eliminating noise, it retains key details such as the drainage tube wall and the edge of the leakage area, avoiding edge blurring caused by traditional Gaussian filtering.

[0082] For electromagnetic interference stripe artifacts: Wavelet transform denoising method is used to decompose the image into different frequency components, remove high-frequency artifact components and reconstruct the image to restore the normal texture of the image.

[0083] Histogram equalization is performed on the pre-processed image, adaptive contrast enhancement is added, and pixel value normalization is performed to obtain a normalized image.

[0084] Perform histogram equalization: stretch the dynamic range of image grayscale to solve the problem of concentrated grayscale distribution and low contrast in ultrasound images, making the contrast agent enhancement signal in the drainage area more prominent.

[0085] Supplemental adaptive contrast enhancement: In response to the situation indicated by the patient condition difference data, such as severe image attenuation in obese patients and weak signal due to ascites, adaptive gain adjustment is performed on local areas of the image to enhance the gray-scale difference in the drainage area and avoid background noise amplification caused by overall enhancement.

[0086] By combining patient condition difference data, the location of the drainage area is determined from the normalized image, and the area is cropped to obtain standard angiographic image data.

[0087] Based on the implantation location parameters of the drainage tube, the approximate coordinate range of the drainage area in the image is determined.

[0088] For patients with thicker fat deposits, the cutting area should be appropriately expanded to avoid truncation of the drainage area edges due to tissue attenuation. For patients with displaced organs, the cutting center coordinates should be adjusted according to the direction of displacement.

[0089] The cropped images are uniformly scaled to a fixed size to eliminate the differences in image size among different patients, resulting in standard angiographic images.

[0090] Clinical operation difference data are obtained by feature extraction from historical clinical data, and physiological impact drainage data are obtained by analyzing the effect of changes in the physiological state of the target patient on the drainage area.

[0091] The steps to obtain clinical operational difference data include:

[0092] Historical clinical data were structured and classified from two dimensions: operational parameters and image standards, to obtain ultrasound examination operational parameters and clinical annotation data.

[0093] Ultrasound examination operating parameters may include: ultrasound probe pressure parameters, scanning angle parameters, contrast agent injection dosage parameters, image acquisition timing parameters, equipment gain and dynamic range parameters, etc.

[0094] Clinical annotation data may include: ultrasound contrast images matching the above operating parameters, manually segmented annotation masks, and clinical diagnostic conclusions.

[0095] The ultrasound examination operation parameters were grouped according to the principle of single variable. The standard operation parameter combination corresponding to the clinical guidelines was selected as the baseline group, and the other operation parameter combinations were selected as the control group.

[0096] For each combination of operating parameters, the ultrasound contrast images corresponding to the data are used to extract image features related to the segmentation of the drainage area from the clinical annotation data as the target image examination features.

[0097] The steps for extracting features from the target image may include: calculating the average gray value, gray standard deviation, and gray histogram peak of the drainage area and surrounding tissue, and quantifying the changes in image brightness and contrast caused by differences in the quantification process.

[0098] An edge detection algorithm is used to extract the edge contours of the drainage tube wall and the leakage area, and the smoothness, continuity, and average gradient of the contours are calculated to quantify the changes in the clarity of the target boundary caused by differences in operation.

[0099] Based on manually labeled masks, the area, perimeter, roundness, and centroid coordinates of the diversion area are calculated to quantify the morphological deviation of the target area caused by differences in operation.

[0100] The target image inspection features of each comparison group are compared with the target image inspection features of the baseline group, and the operational difference feature value is calculated. The formula is as follows:

[0101]

[0102] In the formula, It is an operation on the difference feature value. It involves examining feature values ​​in the target images of the comparison group. These are the feature values ​​for checking the target image of the benchmark group.

[0103] The operational differences are classified according to the magnitude of the operational difference characteristic values ​​to obtain the classification results, which are then linked and bound with the ultrasound examination operation parameters and operational difference characteristic values ​​to form clinical operational difference data.

[0104] Operational differences are categorized into four levels: no difference, slight difference, moderate difference, and severe difference.

[0105] The range of values ​​for each operating parameter and the corresponding characteristic differences, such as the decrease in the smoothness of the drainage area boundary by 15% for every 10 kPa increase in probe pressure.

[0106] The different operational differences affect the segmentation results in different directions. For example, delayed acquisition can lead to the fading of contrast agent and a smaller measured value of the drainage area.

[0107] Prioritize operational differences, such as the influence of acquisition timing on image features > probe pressure > scanning angle.

[0108] The steps to obtain physiological impact data on drainage include:

[0109] Physiological state changes were collected from three dimensions: dynamic physiological changes, static physiological characteristics, and contrast agent metabolism. Corresponding data were then collected in conjunction with clinical practice to form an associated dataset.

[0110] Dynamic physiological changes include: respiratory movement amplitude, such as organ displacement distance during quiet / deep breathing; heart rate fluctuations, such as the effect of heart rate changes on abdominal blood flow and contrast agent distribution; and changes in body position, such as the angle of deviation of the drainage tube position corresponding to supine / lateral / semi-recumbent positions.

[0111] Static physiological characteristics include: abdominal fat thickness, displacement of the original position of organs (such as compression of the abdominal cavity caused by hepatosplenomegaly), pathological conditions (such as tissue edema caused by abdominal infection), and the degree of blurring of the drainage area boundary caused by ascites.

[0112] Dimensions related to contrast agent metabolism: metabolic rate after contrast agent injection, blood circulation status (e.g., delayed contrast agent filling due to hypotension). Simultaneously, abdominal drainage ultrasound contrast images and manually labeled data of the drainage area under different physiological states were collected to form a correlated dataset.

[0113] Based on the associated dataset, the location features, morphological features, image grayscale features, and boundary features of the drainage area were extracted as physiological impact assessment indicators for ultrasound contrast images under different physiological conditions.

[0114] Location characteristics: the coordinate offset of the center point of the drainage tube, and the change in the relative distance between the centroid of the drainage area and the reference point of the abdominal cavity.

[0115] Morphological characteristics: the rate of change of the area of ​​the drainage region, the rate of change of the perimeter, the roundness, and the thickness deviation of the drainage tube wall.

[0116] Image grayscale characteristics: average grayscale value of contrast agent in drainage area, grayscale uniformity, and contrast difference with surrounding tissue.

[0117] Boundary features: Clarity score of the edge of the drainage area, 1-5 points, where 1 point is a completely blurred boundary and 5 points is a clearly distinguishable boundary with continuous edge contour.

[0118] The control variable method was used to analyze the impact of changes in a single physiological state on physiological impact assessment indicators and to calculate the characteristic change rate.

[0119] Multiple linear regression analysis was used to calculate the contribution rate of each physiological factor to the physiological impact assessment index in the case of coordinated changes in multiple physiological states, and the correlation analysis results were obtained.

[0120] Based on the correlation analysis results and the characteristic change rate, the degree of influence of different physiological state changes was classified to obtain physiological influence drainage data.

[0121] Physiological impact data for drainage may include: the specific values ​​of various physiological parameters of the target patient and their differences from the baseline state.

[0122] The characteristics of the drainage area change according to the changes in various physiological states. For example, during deep breathing, the center point of the drainage tube shifts by about 8 mm, and the area increases by 12%.

[0123] Prioritization and quantification of physiological effects.

[0124] For segmentation tasks, prompts may indicate that the patient's abdominal infection has resulted in blurred boundaries, requiring enhanced edge feature extraction during segmentation.

[0125] A differential perception feature purification method was used to analyze patient state difference data, clinical operation difference data, and physiological influence drainage data to obtain general segmentation features, and a meta-learning feature set was formulated in combination with standard contrast images.

[0126] The steps to obtain general segmentation features include:

[0127] Using standard contrast images as a carrier, corresponding patient status difference data, clinical operation difference data, and physiological impact drainage data are tagged and bound to form a multidimensional correlation set.

[0128] For patient condition differences, clinical operation differences, and physiological impact on drainage data, the degree of influence on drainage area segmentation features is quantified, and the mapping relationship of differences is determined.

[0129] Feature impact analysis based on patient condition differences: Cases with significant differences in patient condition, such as obesity vs. thinness, infection vs. non-infection, were selected. Gray-scale, boundary, and morphological features of the drainage area were extracted, and the fluctuation range of features under different difference levels was calculated. For example, patients with a difference in fat thickness ≥30% showed an average decrease of 25% in contrast of the drainage area and a decrease of 1.8 points in boundary clarity score. A mapping table of patient difference level and feature attenuation coefficient was established.

[0130] Feature impact analysis based on clinical operational variation data: Referring to the operational parameter-feature variation patterns in the clinical operational variation data, the interference weight of different operational deviations on segmentation features is quantified. For example, the weight of the impact of acquisition timing delay on contrast agent grayscale features is determined to be: > the weight of the impact of excessive probe pressure on regional morphological features > the weight of the impact of scanning angle deviation on boundary features.

[0131] Feature impact analysis based on physiological influence drainage data: Based on the degree of influence in the physiological influence drainage data, the sensitivity of dynamic physiological changes and static physiological characteristics is calibrated. For example, it is calibrated that severe respiratory motion will increase the sensitivity of the drainage tube position displacement feature to 80%, and moderate ascites will increase the sensitivity of the boundary blurring feature to 65%.

[0132] A weighted attention mechanism is used to integrate the differential feature mapping relationship and select the fusion feature vectors that are relevant to the segmentation task.

[0133] Set fusion weights: Combine clinical priority and characteristic influence quantification results to assign fusion weights to the three types of difference data, such as patient status difference weight 0.4, physiological influence drainage weight 0.35, and clinical operation difference weight 0.25.

[0134] Construct a difference-feature attention matrix: with the core segmentation features of the traffic-driving area (such as edge gradient, gray-level mean, and morphological moment) as rows and the key indicators of the three types of difference data as columns, calculate the attention coefficient of each difference indicator to each segmentation feature, highlighting the core features corresponding to high-impact differences.

[0135] Generate a fused feature vector: Multiply the single-dimensional features with the corresponding attention coefficients and sum them to obtain a feature vector that fuses multi-source differential information. This vector contains both the segmentation features of the image itself and the influence information of the differential data.

[0136] A feature clustering algorithm is used to cluster the fused feature vectors, and the features corresponding to the cluster centers are used as segmentation features. For example, clustering yields general features such as high-contrast boundary features, morphological robustness features, and gray-scale anti-interference features. These features can be adapted to images under different patient, operational, and physiological conditions.

[0137] Calculate the Pearson correlation coefficient between different segmentation features, and retain the segmentation features that reach the preset coefficient threshold as general segmentation features.

[0138] The steps for developing a meta-learning feature set include:

[0139] The general segmentation features are decomposed to determine the feature types and quantification indicators to form feature codes.

[0140] Classified by feature attributes: General segmentation features are divided into three categories: low-level image features, such as edge gradient, gray-level mean, texture entropy, target morphology features, such as the circularity of the drainage area, area ratio, and tube wall thickness ratio, and differential adaptation features, such as patient differential attenuation coefficient, operational deviation interference weight, and physiological influence sensitivity.

[0141] Feature quantization encoding: Numerical encoding is performed on each type of feature. For example, the edge gradient is quantized into the gradient value in the interval [0,1], the influence weight in the difference adaptation feature is directly used as the encoding value, and the qualitative feature is converted into a numerical level of 1–5.

[0142] Construct feature encoding: Establish a feature dictionary with feature name, feature type, quantification index, and encoding rules to ensure consistency in subsequent feature embedding.

[0143] For each standard contrast image, a convolutional neural network is used to extract the underlying morphological features, which are then compared with general segmentation features to calculate feature similarity. The underlying morphological features with a preset matching degree are selected as general image matching features.

[0144] The general matching features of the image are dimensionally aligned and normalized, and then bound to the feature encoding to form the image feature encoding.

[0145] Image feature encoding is divided into support set and query set according to a preset ratio. Meta-learning subtasks are constructed based on different combinations of patient differences, operational differences, and physiological influences. For example, there are subtasks such as obese patients + high probe pressure + deep breathing, and ascites patients + standard operation + supine position, to ensure that subtasks cover feature combinations of different clinical scenarios.

[0146] For each meta-learning subtask, geometric transformation enhancement and noise enhancement are performed to form a meta-learning feature set.

[0147] Geometric transformation enhancement: The standard angiographic image corresponding to the feature vector is flipped, rotated, and scaled, and the morphological features (such as area and position offset) in the feature vector are updated synchronously.

[0148] Noise enhancement: A small amount of Gaussian noise is added to the feature vector to simulate noise interference during image acquisition and improve the feature's anti-interference ability.

[0149] A meta-learning segmentation model is constructed, and a meta-learning feature segmentation model is obtained by training based on the meta-learning feature set. The input is the ultrasound contrast image to be segmented, and the output is the segmentation result of the drainage area.

[0150] The steps to train a meta-learning feature segmentation model include:

[0151] A three-level architecture consisting of a feature extraction network, differential feature fusion, and prototype matching segmentation is adopted to perform parameter initialization and training hyperparameter configuration to obtain a meta-learning segmentation model.

[0152] Feature extraction network: A variant of U-Net is selected as the backbone network. A standard contrast image is input to extract low-level detail features, such as the edge of the drainage tube and the grayscale distribution of the contrast agent, as well as high-level semantic features, such as the overall morphology of the drainage area. Simultaneously, a feature mapping layer is added to the backbone network to transform the differentially adapted features in the meta-learning feature set into vectors of the same dimension as the image features.

[0153] Differential Feature Fusion: An attention mechanism is introduced to calculate the weighting coefficients of differential fitting features on image features—high weights are assigned to differential features that have a significant impact on segmentation, such as boundary blurring caused by severe ascites, while low weights are assigned to features with a minor impact. Deep fusion of image features and differential fitting features is achieved through weighted summation, generating a fused feature vector that includes both the segmentation features of the image itself and information on the impact of clinical differences.

[0154] Prototype Matching Segmentation Head: Borrowing from the core idea of ​​prototype networks, the fused feature vector is input into the segmentation head. Prototype features of the target region are extracted from the fused features of the support set through masked average pooling. The cosine similarity between the fused features of the query set and the prototype features is calculated to generate an initial segmentation mask. Finally, a 1×1 convolutional layer optimizes the mask boundaries, outputting the final segmentation result.

[0155] Parameter initialization: The feature extraction network uses pre-trained weights from medical images, such as U-Net weights pre-trained on the abdominal organ ultrasound dataset. The differential feature fusion and prototype matching segmentation heads are randomly initialized using a normal distribution to ensure that the model has basic feature extraction capabilities in its initial state.

[0156] Hyperparameter configuration: Set the core hyperparameters for meta-learning—inner loop learning rate, such as 0.001, used for fine-tuning parameters within a single task; outer loop learning rate, such as 0.0001, used for global parameter updates; meta-task batch size, such as selecting 16 meta-learning sub-tasks for training each time; inner loop fine-tuning steps, such as 3 steps, meaning fine-tuning 3 times within each sub-task. Also configure training-related parameters—number of training epochs, such as 100 epochs; loss function weights, Dice loss weight 0.7, boundary loss weight 0.3; early stopping threshold, stopping training if there is no improvement in accuracy on the validation set after 10 consecutive epochs.

[0157] Multiple meta-learning sub-tasks are randomly sampled from the meta-learning feature set. The general image matching features of the support set are input into the meta-learning segmentation model to extract the prototype features of the diversion region and generate the support set segmentation mask.

[0158] Calculate the loss values ​​of the support set segmentation mask and the support set label mask, and update the model parameters using gradient descent.

[0159] The query set is processed using the updated model parameters to generate a query set segmentation mask, and the query set loss value is calculated.

[0160] The mean loss is calculated by summing the query set loss values ​​of all meta-learning subtasks. The initial parameters of the meta-learning segmentation model are then updated using gradient descent until the preset number of training rounds is reached, resulting in the meta-learning feature segmentation model.

[0161] The steps for outputting the drainage region segmentation results include:

[0162] The ultrasound contrast image to be segmented is input into the feature extraction network of the meta-learning feature segmentation model to extract low-level detail features and high-level semantic features of the image as the features of the image to be segmented.

[0163] The difference data of the target patient status are analyzed by difference feature fusion to obtain the feature to be differentiated, and the feature to be image is fused with the feature to be differentiated to obtain the fused feature vector to be segmented.

[0164] Select a reference support set that is the same as the difference scene to be segmented from the meta-learning feature set, and input the corresponding standard contrast images and patient state difference data into the meta-learning feature segmentation model to extract the baseline prototype features.

[0165] Calculate the cosine similarity between the feature vector to be segmented and the baseline prototype feature vector, generate a similarity heatmap, and mark candidate diversion regions according to a preset similarity threshold to obtain a preliminary segmentation mask.

[0166] The initial segmentation mask is processed by removing noise points through morphological opening operations and then filling the holes through morphological closing operations to obtain the processed segmentation mask.

[0167] The jagged boundaries of the drainage region are corrected by the processing segmentation mask based on the edge smoothing algorithm, and the drainage region segmentation result is output.

[0168] The results of drainage area segmentation can include: a visualized segmentation mask image, such as marking the location of the drainage tube, the extent of the drainage fluid area, the leakage area, the area of ​​the drainage area, the perimeter, and the centroid coordinates.

[0169] Based on the same general inventive concept, this invention also protects a meta-learning-based small-sample segmentation system for abdominal drainage ultrasound contrast imaging, comprising:

[0170] The heterogeneous data processing module is used to collect target patient data and historical clinical data, acquire ultrasound contrast images in real time, analyze the target patient data to obtain patient status difference data, and preprocess the ultrasound contrast images to obtain standard contrast images.

[0171] The differential diagnosis analysis module is used to extract features from historical clinical data to obtain clinical operation difference data, and to obtain physiological impact drainage data based on the influence of changes in the target patient's physiological state on the drainage area.

[0172] The feature set construction module is used to analyze patient state difference data, clinical operation difference data and physiological influence drainage data using the difference perception feature purification method to obtain general segmentation features, and to formulate a meta-learning feature set in combination with standard contrast images.

[0173] The segmentation output data module is used to construct a meta-learning segmentation model. It is trained based on the meta-learning feature set to obtain the meta-learning feature segmentation model. The input is the ultrasound contrast image to be segmented, and the output is the segmentation result of the drainage area.

[0174] In summary, the meta-learning-based small-sample segmentation method and system for abdominal drainage ultrasound contrast imaging provided in this embodiment, through multi-source differential feature fusion and attention mechanisms, enables the model to accurately identify the boundary of the drainage area, significantly improving segmentation accuracy. The meta-learning framework allows the model to quickly adapt to new patients, maintaining stable segmentation performance even on unseen patient data, enhancing generalization ability and effectively solving the overfitting problem. Through differential feature quantification analysis and weight allocation, the influence of different factors on the segmentation results can be clearly demonstrated, enhancing the model's interpretability and facilitating clinical understanding. Combined with individual patient difference data, the model can adaptively adjust for different patients, achieving personalized and accurate segmentation, providing technical support for precision medicine.

[0175] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0176] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A small-sample segmentation method for abdominal drainage ultrasound contrast imaging based on meta-learning, characterized in that, include: Collect target patient data and historical clinical data, acquire ultrasound contrast images in real time, analyze the target patient data to obtain patient status difference data, and preprocess the ultrasound contrast images to obtain standard contrast images. Clinical operation difference data are obtained by feature extraction from the historical clinical data, and physiological impact drainage data are obtained by considering the influence of changes in the physiological state of the target patient on the drainage area. The patient state difference data, the clinical operation difference data, and the physiological influence drainage data were analyzed using the differential perception feature purification method to obtain general segmentation features, and a meta-learning feature set was formulated in combination with the standard contrast images. A meta-learning segmentation model is constructed, and a meta-learning feature segmentation model is obtained by training based on the meta-learning feature set. The input is the ultrasound contrast image to be segmented, and the output is the segmentation result of the drainage area.

2. The method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning according to claim 1, characterized in that: The steps for analyzing the patient status difference data include: The abdominal fat thickness, organ position offset, contrast agent metabolism rate, pathological condition, and drainage tube implantation duration parameters of the target patients were collected as the target patient data. Cases with the same disease and drainage type as the target patient were screened from the historical clinical data, and corresponding parameter statistical features were extracted to construct benchmark reference data. The target patient data is compared with the benchmark reference data, the difference value is calculated, and the difference value is graded and labeled according to the preset standard to highlight the key difference characteristics. By combining clinical experience and correlation analysis of the historical clinical data, weights are assigned to each difference value, a comprehensive difference value is calculated, and the patient status difference data is obtained by associating and binding it with the original information of the target patient.

3. The method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning according to claim 1, characterized in that: The steps for preprocessing to obtain the standard contrast image include: The ultrasound contrast images are subjected to quality control screening to remove invalid images, and speckle noise and stripe artifacts are differentially eliminated to obtain preliminary processed images. Histogram equalization is performed on the pre-processed image, adaptive contrast enhancement is added, and pixel value normalization is performed to obtain a normalized image. The drainage area is located in the normalized image by combining the patient status difference data, and the area is cropped to obtain the standard angiography image data.

4. The method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning according to claim 1, characterized in that: The steps to obtain the clinical operational difference data include: The historical clinical data is structured and classified from two dimensions: operating parameters and image standards, to obtain ultrasound examination operating parameters and clinical annotation data; The ultrasound examination operation parameters were grouped according to the single variable principle. The standard operation parameter combination corresponding to the clinical guidelines was selected as the baseline group, and the other operation parameter combinations were used as the control group. For each set of operating parameters, image features related to the segmentation of the drainage area are extracted from the clinical annotation data as target image examination features for the ultrasound contrast images. The target image inspection features of each comparison group are compared with the target image inspection features of the benchmark group, and the operation difference feature value is calculated. The operational differences are classified according to the magnitude of the operational difference feature values ​​to obtain the classification results, and then associated and bound with the ultrasound examination operation parameters and the operational difference feature values ​​to form the clinical operational difference data.

5. The method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning according to claim 1, characterized in that: The steps for obtaining the physiological effects of drainage data include: Physiological state changes were collected from three dimensions: dynamic physiological changes, static physiological characteristics, and contrast agent metabolism. Corresponding data were collected in conjunction with clinical practice to form an associated dataset. Based on the associated dataset, for ultrasound contrast images under different physiological conditions, the location features, morphological features, image grayscale features, and boundary features of the drainage area are extracted as physiological impact assessment indicators. The influence of a single physiological state change on the physiological impact assessment index was analyzed using the controlled variable method, and the characteristic change rate was calculated. Multiple linear regression analysis was used to calculate the contribution rate of each physiological factor to the physiological impact assessment index in the case of coordinated changes in multiple physiological states, and the correlation analysis results were obtained. Based on the correlation analysis results and the characteristic change rate, the degree of influence of different physiological state changes is classified to obtain the physiological influence drainage data.

6. The method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning according to claim 1, characterized in that: The steps to obtain the general segmentation features include: Using the standard contrast images as a carrier, the corresponding patient status difference data, clinical operation difference data, and physiological impact drainage data are tagged and bound to form a multidimensional correlation set; For the patient status difference data, the clinical operation difference data, and the physiological influence drainage data, the degree of influence on the drainage area segmentation features is quantified respectively, and the difference feature mapping relationship is determined. A weighted attention mechanism is used to integrate the differential feature mapping relationships and filter the fusion feature vectors that are relevant to the segmentation task. The fused feature vectors are clustered using a feature clustering algorithm, and the features corresponding to the cluster centers are used as segmentation features. Calculate the Pearson correlation coefficient between different segmentation features, and retain the segmentation features that reach the preset coefficient threshold as the general segmentation features.

7. The method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning according to claim 1, characterized in that: The steps for formulating the meta-learning feature set include: The general segmentation features are decomposed to determine the feature types and quantification indicators to form feature codes; For each standard imaging image, a convolutional neural network is used to extract the underlying morphological features, which are then compared with the general segmentation features to calculate the feature similarity. The underlying morphological features with a preset matching degree are then selected as the general matching features of the image. The general matching features of the image are subjected to dimensional alignment and normalization, and then bound to the feature encoding to form image feature encoding; The image feature encoding is divided into a support set and a query set according to a preset ratio, and a meta-learning subtask is constructed based on different combinations of patient differences, operational differences and physiological effects. For each meta-learning subtask, geometric transformation enhancement and noise enhancement are performed to form the meta-learning feature set.

8. The method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning according to claim 1, characterized in that: The steps for training the meta-learning feature segmentation model include: A three-level architecture consisting of a feature extraction network, differential feature fusion, and prototype matching segmentation is adopted to initialize parameters and configure training hyperparameters to obtain the meta-learning segmentation model. Multiple meta-learning sub-tasks are randomly sampled from the meta-learning feature set, and the general image matching features of the support set are input into the meta-learning segmentation model to extract the prototype features of the diversion region and generate a support set segmentation mask. Calculate the loss values ​​of the support set segmentation mask and the support set label mask, and update the model parameters using gradient descent; The query set is processed using the updated model parameters to generate a query set segmentation mask, and the query set loss value is calculated. The mean loss is calculated by summing the query set loss values ​​of all meta-learning subtasks, and the initial parameters of the meta-learning segmentation model are updated by gradient descent until a preset number of training rounds is reached, thus obtaining the meta-learning feature segmentation model.

9. The method for small sample segmentation of abdominal drainage ultrasound contrast imaging based on meta-learning according to claim 1, characterized in that: The steps for outputting the drainage area segmentation results include: The ultrasound contrast image to be segmented is input into the feature extraction network of the meta-learning feature segmentation model to extract low-level detail features and high-level semantic features of the image as features of the image to be segmented. The target patient status difference data is analyzed by difference feature fusion to obtain the feature to be differentiated, and the image feature to be differentiated is fused with the feature to be differentiated to obtain the segmentation fusion feature vector; Select a reference support set that is the same as the difference scene to be segmented from the meta-learning feature set, input the corresponding standard contrast image and patient state difference data into the meta-learning feature segmentation model, and extract the baseline prototype features; Calculate the cosine similarity between the feature vector to be segmented and the baseline prototype feature vector, generate a similarity heatmap, and mark candidate traffic-driving regions according to a preset similarity threshold to obtain a preliminary segmentation mask; The initial segmentation mask is processed by removing noise points through morphological opening operations and filling holes through morphological closing operations to obtain the processed segmentation mask. The jagged boundaries of the drainage region are corrected by the processing segmentation mask according to the edge smoothing algorithm, and the drainage region segmentation result is output.

10. A meta-learning-based small-sample segmentation system for abdominal drainage ultrasound contrast imaging, applied to the meta-learning-based small-sample segmentation method for abdominal drainage ultrasound contrast imaging as described in any one of claims 1 to 9, characterized in that, The segmentation system includes: The heterogeneous data processing module is used to collect target patient data and historical clinical data, acquire ultrasound contrast images in real time, analyze the target patient data to obtain patient status difference data, and preprocess the ultrasound contrast images to obtain standard contrast images. The differential diagnosis and analysis module is used to extract features from the historical clinical data to obtain clinical operation difference data, and to obtain physiological impact drainage data based on the influence of changes in the target patient's physiological state on the drainage area. The feature set construction module is used to analyze the patient state difference data, the clinical operation difference data, and the physiological influence drainage data using the difference perception feature purification method to obtain general segmentation features, and to formulate a meta-learning feature set in combination with the standard contrast images; The segmentation output data module is used to construct a meta-learning segmentation model. It is trained based on the meta-learning feature set to obtain a meta-learning feature segmentation model. The input is the ultrasound contrast image to be segmented, and the output is the segmentation result of the drainage area.

Images