Automatic tumor lesion analysis method and platform, storage medium and computer equipment

Abstract

The invention discloses an automatic tumor lesion analysis method and platform, a storage medium and computer equipment, and the method comprises the steps: determining a tumor minimum coverage graph region based on a tumor region marked by an ultrasonic image and a minimum coverage graph corresponding to a tumor type, and subtracting the tumor region from the tumor minimum coverage graph region to obtain a minimum coverage graph region; obtaining a peritumoral region; the method comprises the following steps: sequentially extracting features of a tumor region, a tumor minimum coverage pattern region and a peritumor region, constructing four feature input modes, respectively inputting the four feature input modes into each ML classifier for model training and evaluation, determining an optimal region selection mode on the basis of the feature input modes used when the ML classifiers are in optimal performance, and carrying out model evaluation on the optimal region selection mode. Therefore, a doctor can obtain a lesion analysis result through an optimal performance ML classifier based on the initially marked tumor area and the optimal area selection mode. Through multi-region feature extraction and classifier optimization, accurate analysis of benign and malignant tumors is realized, and ultrasonic image diagnosis efficiency and accuracy are improved.

Description

Technical Field

[0001] This application relates to the field of medical image recognition technology, and in particular to an automated method and platform for tumor lesion analysis, a storage medium, and a computer device. Background Technology

[0002] Cancer is a major challenge to global public health and remains one of the leading causes of death worldwide. Breast cancer and thyroid cancer, in particular, have garnered significant attention due to their substantial health impacts and epidemiological characteristics. Increasing research efforts are focused on achieving early, accurate diagnosis and developing personalized treatment plans, which are crucial for improving patient outcomes and reducing mortality.

[0003] Early diagnosis and treatment planning for cancer are typically achieved through routine screening procedures that rely on advanced technologies, with medical imaging being the most common approach. Ultrasound, in particular, is widely used due to its non-invasiveness, lack of ionizing radiation, convenience, and cost-effectiveness; these characteristics make it especially suitable for screening and diagnosing breast and thyroid cancers. However, effective analysis of imaging data remains hampered by several challenges, including a limited number of trained physicians and poor reproducibility due to inter- and intra-observer variability. Summary of the Invention

[0004] In view of this, this application provides an automated tumor lesion analysis method and platform, storage medium, and computer equipment. Compared with the current methods that mostly use morphological dilation operations for lesion identification, which require prior confirmation of dilated pixels, this application only needs to identify the tumor region and adaptively select its corresponding surrounding area for radiomics feature extraction and selection, as well as ML-based classification and evaluation. This integrated process achieves the standardization of peritumoral information, is independent of morphological utilization, and is more concise and efficient, thereby ultimately achieving more accurate and repeatable lesion identification.

[0005] According to one aspect of this application, an automated method for tumor lesion analysis is provided, the method comprising: For ultrasound images containing tumors, multiple ultrasound images involving the same type of tumor are acquired. The ultrasound images are labeled with tumor type, tumor area pre-labeled by the doctor, and lesion analysis results. The lesion analysis results include benign and malignant tumors. Different tumor types correspond to preset minimum coverage graphics. For any ultrasound image, based on the minimum coverage pattern corresponding to the tumor region and tumor type marked on the ultrasound image, the minimum coverage pattern region of the tumor is determined, and the tumor region is subtracted from the minimum coverage pattern region of the tumor to obtain the peritumoral region. After extracting features from the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region in sequence, four feature input methods are constructed based on the extracted features. These methods are then input into the ML classifier for training. The performance of the ML classifier in outputting lesion analysis results under different feature input methods is evaluated. The trained ML classifier outputs lesion analysis results for the input features. The feature input methods include inputting only the features of the tumor region, inputting only the features of the minimum coverage area of ​​the tumor, inputting only the features of the peritumoral region, and inputting the features of both the tumor region and the peritumoral region in parallel. Based on the feature input method used to achieve optimal performance of the ML classifier, the optimal region selection method is determined. After the doctor performs initial tumor region annotation on the newly input ultrasound image, the target region extraction features are formed based on the initially annotated tumor region and the optimal region selection method. The features are then input into the trained ML classifier based on the feature input method under the optimal region selection method to obtain the lesion analysis results. The target region includes the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region.

[0006] According to another aspect of this application, an automated tumor lesion analysis platform is provided, the platform comprising: The ultrasound image acquisition module is used to acquire multiple ultrasound images involving the same type of tumor from ultrasound images containing tumors. The ultrasound images are labeled with tumor type, tumor area pre-marked by the doctor, and lesion analysis results. The lesion analysis results include benign and malignant tumors. Different tumor types correspond to preset minimum coverage graphics. The region of interest identification module, for any ultrasound image, determines the minimum coverage area of ​​the tumor based on the minimum coverage area and tumor type marked on the ultrasound image, and subtracts the tumor area from the minimum coverage area of ​​the tumor to obtain the peritumoral region; The classifier performance evaluation module is used to extract features from the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region in sequence. Based on the extracted features, four feature input methods are constructed and input into the ML classifier for training. The module evaluates the performance of the ML classifier in outputting lesion analysis results under different feature input methods. The trained ML classifier outputs lesion analysis results for the input features. The feature input methods include inputting only the features of the tumor region, inputting only the features of the minimum coverage area of ​​the tumor, inputting only the features of the peritumoral region, and inputting the features of the tumor region and the peritumoral region in parallel. The automated tumor lesion analysis module is used to determine the optimal region selection method based on the feature input method that enables the ML classifier to perform at its best. After the doctor performs initial tumor region annotation on the newly input ultrasound image, the module extracts features from the target region based on the initially annotated tumor region and the optimal region selection method. The features are then input into the pre-trained ML classifier based on the feature input method under the optimal region selection method to obtain the lesion analysis results. The target region includes the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region.

[0007] According to another aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described automated tumor lesion analysis method.

[0008] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described automated tumor lesion analysis method.

[0009] By employing the above technical solution, this application provides an automated tumor lesion analysis method, platform, storage medium, and computer equipment. Based on the minimum coverage pattern corresponding to the tumor region and tumor type annotated by ultrasound images, the minimum coverage pattern region of the tumor is determined. The tumor region is then subtracted from the minimum coverage pattern region to obtain the peritumoral region. Features of the tumor region, the minimum coverage pattern region, and the peritumoral region are extracted sequentially, and four feature input methods are constructed. These are then input into an ML classifier for training. Based on the feature input method that optimizes the ML classifier's performance, the optimal region selection method is determined, enabling physicians to obtain lesion analysis results based on the initially annotated tumor region and the optimal region selection method. Through multi-region feature extraction and classifier optimization, accurate analysis of tumor benign and malignant tumors is achieved, improving the efficiency and accuracy of ultrasound image diagnosis.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating an automated tumor lesion analysis method provided in an embodiment of this application is shown. Figure 2 This illustration shows a schematic diagram of an automated cancer lesion analysis framework provided in an embodiment of this application; Figure 3 This illustration shows a schematic diagram of a region of interest generation process provided in an embodiment of this application; Figure 4 This illustration shows a schematic diagram of a rectangular and circular minimum coverage graphic region for a tumor, provided in an embodiment of this application. Figure 5 A schematic diagram of the framework of an automated tumor lesion analysis platform provided in an embodiment of this application is shown. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] This embodiment provides an automated method for tumor lesion analysis, such as Figure 1 As shown, the method includes: Step 101: For ultrasound images containing tumors, acquire multiple ultrasound images involving the same type of tumor. The ultrasound images are labeled with tumor type, tumor area pre-labeled by the doctor, and lesion analysis results. The lesion analysis results include benign and malignant tumors. Different tumor types correspond to preset minimum coverage graphics. Step 102: For any ultrasound image, based on the minimum coverage pattern corresponding to the tumor region and tumor type marked on the ultrasound image, determine the minimum coverage pattern region for the tumor, and subtract the tumor region from the minimum coverage pattern region to obtain the peritumoral region. Step 103: After extracting the features of the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region in sequence, four feature input methods are constructed based on the extracted features. These methods are then input into the ML classifier for training. The performance of the ML classifier in outputting lesion analysis results under different feature input methods is evaluated. The trained ML classifier outputs lesion analysis results for the input features. The feature input methods include inputting only the features of the tumor region, inputting only the features of the minimum coverage area of ​​the tumor, inputting only the features of the peritumoral region, and inputting the features of both the tumor region and the peritumoral region in parallel. Step 104: Based on the feature input method used to achieve optimal performance of the ML classifier, determine the optimal region selection method so that after the doctor performs initial tumor region annotation on the newly input ultrasound image, the target region extraction features are formed based on the initially annotated tumor region and the optimal region selection method. The features are then input into the trained ML classifier based on the feature input method under the optimal region selection method to obtain the lesion analysis results. The target region includes the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region.

[0014] In the above embodiments of this application, the minimum coverage pattern method can be widely applied to the identification of various tumor types, including but not limited to tumors in the breast and thyroid regions. Correspondingly, ultrasound images include breast ultrasound images and thyroid ultrasound images. It can also identify other imaging modalities, such as CT and MRI.

[0015] Specifically, in the process of automated tumor lesion analysis, the novel "adaptive peritumoral region selection method" constructed in this application, named APASA, can be used. APASA is implemented through a clinically meaningful, AI-driven automated cancer lesion analysis framework, such as... Figure 2 As shown, Regions of Interest (ROIs) are extracted from medical images (ultrasound images) for feature extraction and selection. ROIs include the tumor region, the tumor minimum coverage area, and the peritumoral region. Multiple machine learning (ML) classifiers are then used for classification, and optimal parameters are found through cross-validation. Finally, various evaluation metrics are used to assess the performance of the ML classifiers. Figure 2In this document, ROIGenerationProcess refers to the ROI generation process, APASA is an abbreviation for the "Adaptive Peritumoral Region Selection Method" named in this application, Image refers to an ultrasound image, Mask refers to a black-and-white image after the doctor marks the tumor region on the ultrasound image, and APASAMask is an APASA mask, which includes the minimum coverage area of ​​the tumor and the peritumoral region. The peritumoral region is obtained by subtracting the tumor region from the minimum coverage area of ​​the tumor. ImageandMaskLogicAndOperationbyPixel performs pixel-wise logical operations on the image and mask. FeatureUtilization refers to feature utilization, FeatureExtraction to feature extraction, Shape to shape, specifically the shape of the smallest covering graphic, which can include rectangles and circles, Histogram to histogram, Texture to texture, WaveletTransform to wavelet transform, CombinedROI40Features to combine 40 features of an ROI, FeatureStandardization to feature standardization, FeatureSelection to feature selection, IndependentSamplesT-test to independent samples T-test, LASSO to feature selection, Top20Features to the top 20 features, and Classify[r](SelectionandPerformance) to classify features. eEvaluation is the classifier, Trainsets is the training set, LR is Logistic Regression, SVM is Support Vector Machine, KNN is K-Nearest Neighbors, NaiveBayes is Naive Bayes, Perceptron is Perceptron, FindingOptimalParameters is for finding optimal parameters, GridSearchCV is for grid search cross-validation, Testsets is the test set, EvaluationIndexes is the evaluation metric, ROCCurve is the ROC curve, Accuracy is the accuracy, Sensitivity is the sensitivity, Specificity is the specificity, and F1-score is the F1 score.

[0016] Specifically, data preparation plays a crucial role in the successful implementation and reliability of AI-based automated frameworks. In cancer lesion analysis, for example, it includes two main steps: (1) tumor region identification, which identifies the area within the tumor (the tumor itself) and the peritumoral area; and (2) lesion type classification, which determines whether the lesion is benign or malignant.

[0017] The effectiveness of the framework proposed in this application has been validated on two cancer types: breast cancer and thyroid cancer. The thyroid cancer dataset can be obtained from public resources, where all lesion type labels (tumor types) and tumor region divisions are provided. The breast cancer dataset is private data, and its preparation can be completed by two experienced radiologists to obtain lesion type labels and tumor region divisions.

[0018] Once both datasets (including images, tumor masks, and lesion type labels) are ready, multiple Regions of Interest (ROIs) are identified through specific operations. For each image and its tumor mask, a Minimal Coverage Pattern (MCG) is constructed using the proposed APASA, which adaptively covers the tumor and its surrounding peritumoral region. This process requires no morphological operations and relies solely on the tumor's boundary coordinates. The peritumoral region is obtained by subtracting the tumor mask (i.e., the tumor region) from the Minimal Coverage Pattern. Therefore, three masks are generated for each case: (1) the tumor mask (i.e., the tumor region); (2) the MCG mask (MCC, circular minimum coverage area of ​​the tumor + MCR, rectangular minimum coverage area of ​​the tumor); and (3) the peritumoral region mask (i.e., the peritumoral region). Each mask is multiplied pixel-by-pixel with the corresponding image to produce three distinct ROIs, which are then used for radiomics feature extraction. Notably, the MCG construction in the proposed APASA framework offers flexibility in shape selection. In the embodiments described above, two shape configurations are listed: rectangular and circular. These two shapes were chosen to explore the impact of the geometry of the tumor periphery on the extracted radiomics features and subsequent classification performance, thereby evaluating the robustness and adaptability of the proposed method under different spatial representations. The overall process of region of interest generation is as follows: Figure 3 As shown.

[0019] Figure 3In this framework, ROIGenerationProcess refers to the ROI generation process, APASA stands for "Adaptive Peritumoral Region Selection Method," Image represents the image, Mask represents the image mask, InverseMask represents the inverse mask, MCR-Peri represents the rectangular minimum coverage area of ​​the tumor, MCC-Peri represents the circular minimum coverage area of ​​the tumor, MCGMask represents the minimum coverage area mask, PeritumorMask represents the peritumoral mask, and ImageandMaskLogicAndOperationbyPixel represents the pixel-wise logical operation between the image and the mask. TumorROI represents the region of interest of the tumor, i.e., the labeled tumor region, MCGROI represents the MCG region of interest, referring to the minimum coverage area of ​​the tumor, and PeritumorROI represents the peritumoral region of interest. Specifically, the input medical image (Image) and mask (Mask) generate the inverse mask (InverseMask). Through APASA processing, combining the mask and the inverse mask, different mask regions are generated, including MCR, MCR-Peri, MCC, MCC-Peri, MCGMask, and PeritumorMask. By performing pixel-by-pixel logical operations, the tumor region of interest (TumorROI), the MCG region of interest (MCGROI), and the peritumor region of interest (PeritumorROI) are finally obtained.

[0020] Specifically, tumor types include, for example, breast cancer and thyroid cancer. When performing lesion analysis on breast cancer, the minimum coverage pattern can be circular, and when performing lesion analysis on thyroid cancer, the minimum coverage pattern can be rectangular. In addition, the minimum coverage pattern can be any shape other than circular and rectangular to suit different cancer types.

[0021] About ML classifiers: ML classifiers, or Machine Learning Classifiers, are algorithms or models used in machine learning to assign input data to different categories or labels. Common types of ML classifiers include: SVM (Support Vector Machine): A widely used classification algorithm suitable for high-dimensional data and non-linear classification problems.

[0022] KNN (K Nearest Neighbors): An instance-based learning algorithm that classifies samples by calculating the distance between the sample to be classified and known samples.

[0023] RF (Random Forest): An ensemble learning method that improves classification accuracy by constructing multiple decision trees and combining their predictions.

[0024] Naive Bayes (NB): A classification algorithm based on Bayesian decision theory, which assumes that the features of samples are independent of each other.

[0025] For ML classifiers, during the training phase, the ML classifier is trained using a labeled training dataset, learning patterns and features in the data to accurately assign new data to the correct category. During the prediction phase, for new input data, the ML classifier utilizes the patterns and features learned during training to assign it to the most probable category.

[0026] Application scenarios for ML classifiers, for example: Image recognition: classifying images into different categories, such as animals, plants, and landscapes.

[0027] Text classification: Classifying text data (such as emails, news articles) into different themes or sentiment categories.

[0028] Medical diagnosis: Classifying a patient's illness into different types based on their symptoms and medical records.

[0029] Financial risk control: Based on the customer's credit history and transaction behavior, their risk level is classified as low, medium, or high.

[0030] Optionally, the minimum coverage pattern includes rectangles and circles. In step 102, based on the minimum coverage pattern corresponding to the tumor region and tumor type marked on the ultrasound image, the minimum coverage pattern region for the tumor is determined, including: Step 1021: When the minimum coverage shape determined based on the tumor type is a rectangle, the tumor region marked by the ultrasound image is converted into a discrete set of coordinate points. Based on the coordinate points corresponding to the minimum horizontal coordinate, maximum horizontal coordinate, minimum vertical coordinate, and maximum vertical coordinate in the set of coordinate points, the minimum coverage shape region of the tumor region is constructed. Step 1022: When the minimum coverage pattern determined based on the tumor type is a circle, the tumor region marked on the ultrasound image is converted into a discrete point set. The Weltz algorithm is used to solve for the minimum coverage circle of the discrete point set to obtain the center and radius. A circle is drawn on the ultrasound image based on the solved center and radius as the minimum coverage pattern region of the tumor region.

[0031] In the above embodiments of this application, when the minimum coverage shape is a rectangle, specifically: 1. Obtain the coordinate point set of the tumor region: The tumor region marked by the doctor (such as the outline or boundary of the tumor in the ultrasound image) is transformed into a discrete set of coordinate points (denoted as point set P). These points are the "boundary feature points" of the tumor region (such as the vertices of the tumor outline, edge sampling points, etc.), because the minimum coverage rectangle only needs to enclose the boundary points of the tumor, and the internal points are naturally included.

[0032] 2. Calculate the extreme values ​​of the coordinates of the point set: Iterate through the horizontal coordinates (x, corresponding to the "width direction" of the ultrasound image) and vertical coordinates (y, corresponding to the "height direction" of the ultrasound image) of all points in the point set P, and find the four key extrema, i.e., P={a,b,c,d}, specifically: Minimum horizontal coordinate x min The minimum x-coordinate of all points in the point set (corresponding to the leftmost boundary of the tumor region, such as...). Figure 4 (a) The x-coordinate of the midpoint b); Maximum horizontal coordinate x max : The maximum x-coordinate of all points in the point set (corresponding to the rightmost boundary of the tumor region, such as...) Figure 4 (x-coordinates of the midpoint c in (a)) Minimum vertical coordinate y min The minimum y-coordinate of all points in the point set (corresponding to the lowest boundary of the tumor region, such as...). Figure 4 (y-coordinate of midpoint d in (a)) Maximum vertical coordinate y max : The maximum y-coordinate of all points in the point set (corresponding to the uppermost boundary of the tumor region, such as...) Figure 4 (y-coordinate of the midpoint a in (a)).

[0033] 3. Define the vertices of the minimum covering rectangle: Using the above four extreme values, we can directly define the four vertices of the smallest axis-aligned rectangle that can completely enclose the tumor region (i.e., the boundary vertices of the smallest tumor-covered graphical region): Top left vertex: (x min ,y max ), the leftmost and topmost, corresponding to Figure 4 In (a), vertex a; Top right corner vertex: (x max ,y max ), the far right and the top, corresponding to Figure 4 (a) Middle vertex b; bottom right corner vertex: (x max ,y min ), the rightmost and bottommost, corresponding to Figure 4 In (a) the vertex c; Bottom left vertex: (x min ,y min ), the leftmost and bottommost, corresponding to Figure 4 In (a), vertex d.

[0034] 4. Determine the minimum coverage area of ​​the tumor in the graphic representation: Therefore, the rectangular area formed by connecting the above four vertices clockwise or counterclockwise is the "minimum tumor coverage area" for the tumor region. The determined minimum tumor coverage area completely includes the tumor region marked by the doctor (all tumor boundary points are within the rectangle). At the same time, it is the "minimum axis-aligned rectangle" that can surround the tumor region, that is, it is impossible to completely surround the tumor region by reducing the width or height of the rectangle. In addition, the sides of the rectangle are parallel to the coordinate axes (horizontal / vertical direction) of the ultrasound image.

[0035] More specifically, if the tumor region marked by the doctor corresponds to the point set P={a,b,c,d}, where: The coordinates of point a are (x) a ,y a ), y a The largest, corresponding to the top; The coordinates of point b are (x) b ,y b ), x b The smallest, corresponding to the leftmost side; The coordinates of point c are (x) c ,y c ), x c The largest, corresponding to the far right; The coordinates of point d are (x) d ,y d ), y d The smallest, corresponding to the bottom.

[0036] The minimum coverage area of ​​the tumor is: with (x b ,y a (top left), (x) c ,y a (top right), (x) c ,y d (bottom right), (x) b ,y d The rectangle with vertices (bottom left) completely encloses the point set P and is the smallest axis-aligned rectangle.

[0037] By following the steps above, the minimum coverage area of ​​the tumor can be accurately obtained based on the "tumor area marked by the doctor" and the "minimum coverage rectangle corresponding to the tumor type".

[0038] Furthermore, when the minimum coverage shape is circular (i.e., the "minimum coverage circle, MCC"), the tumor area marked by the doctor can be completely enclosed by the minimum circumscribed circle. Specifically, for example... Figure 4 (b) 1. Obtain the discrete point set of the tumor region: The tumor area marked by the doctor (such as the outline or boundary of the tumor in an ultrasound image) will be converted into a discrete set of coordinate points (denoted as point set P={p1,p2,…,p...). n These points represent the boundary feature points of the tumor region (such as the vertices of the tumor contour, edge sampling points, etc.), because the minimum covering circle only needs to enclose the boundary points of the tumor, and the internal points are naturally included.

[0039] 2. Determine the solution method for the minimum covering circle (MCC): The minimum covering circle is the circle with the smallest area that completely encloses all points in a point set P. According to mathematical theory: If a point set P contains 1 to 2 points, the smallest covering circle is the circle (or single-point circle) with these two points as the endpoints of its diameter. If point set P has 3 or more points, the smallest covering circle could be: The circumcircle bounded by three points (if the three points are not collinear); A circle whose diameter endpoints are two points (if the distance between the two points is the maximum and all other points are inside the circle). In special cases, it may also degenerate into a circle centered on a certain point (but this rarely occurs in actual tumor annotation because the tumor area usually has a certain range).

[0040] The solution can be obtained using the Welzl's algorithm, whose core idea is the stochastic incremental method, which finds the minimum covering circle by gradually expanding the boundary circle.

[0041] 3. Run the Weltz algorithm to determine the parameters of the minimum covering circle: Following the Weltz algorithm, the points in the point set P are processed step by step to finally obtain the following parameters of the minimum covering circle: Center coordinates: (x c ,y c ( ), that is, the location of the center point of the covering circle in the ultrasound image; Radius: r, which is the size of the covering circle, is determined by the boundary points in the point set P.

[0042] Specific steps are as follows: Initialize an empty boundary circle, and traverse the point set P in random order; For each new point p i If it is not inside or on the current circle, then update the minimum covering circle so that it passes through p. i and 1-2 points from the current boundary points; Repeat the above steps until all points have been processed. The resulting circle is the minimum covering circle (MCC).

[0043] 4. Define the minimum coverage area of ​​the tumor in the graphic: Using the minimum covering circle parameter (center (x) obtained above) c ,y c The radius r) is used to draw a circular area on the ultrasound image. This circular area is the "minimum tumor coverage area". This area completely includes the tumor area marked by the doctor. That is, all tumor boundary points are inside or on the circle. It is the "minimum circle" that can surround the tumor area. It is impossible to maintain complete enclosure of the tumor area by reducing the radius. The center and radius are determined by the spatial distribution of the tumor boundary points and may not coincide with the geometric center of the tumor (e.g. when the tumor shape is irregular).

[0044] Specifically, the minimum coverage circle can be visually verified to ensure it completely surrounds the tumor area labeled by the doctor. If some boundary points are found to be uncovered, the Weltz algorithm can be rerun or the point set can be adjusted (e.g., by increasing the density of boundary sampling points) to improve coverage accuracy.

[0045] Therefore, when the minimum coverage shape is circular, based on "the tumor region annotated by the doctor" and "the circular minimum coverage shape corresponding to the tumor type", further: Transform the tumor regions marked by doctors into a discrete point set P; Weltz's algorithm is used to find the minimum covering circle (MCC) of the point set P, and the center and radius of the circle are obtained. The circle is drawn on the ultrasound image as the smallest coverage area of ​​the tumor. This circle completely surrounds the tumor area and is the smallest coverage circle.

[0046] Using the above method, the "minimum coverage area of ​​the tumor" can be accurately obtained for the tumor region, providing a basis for subsequent extraction of the peritumoral region (minimum coverage area minus the tumor region) and feature analysis.

[0047] Optionally, in step 103, features of the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region are extracted sequentially, including: Step 1031: For any region among the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region, apply the feature extractor function of the PyRadiomics library to extract the features of the region.

[0048] Optionally, in step 1031, the feature extractor function of the PyRadiomics library is applied to extract features of the region, including: Step 10311: Apply the feature extractor function of the PyRadiomics library to extract the initial features of the region; Step 10312: Based on the initial features, determine the initial candidate set for minimum absolute shrinkage and selection operator regression; Step 10313: After optimizing the feature selection by combining the initial candidate set and five-fold cross-validation, the optimized features are sorted according to the absolute value of the feature weights, and the features of the region are determined based on the sorting results.

[0049] In the above embodiments of this application, using the APASA strategy proposed in this application and the initially generated tumor mask (tumor region marked by the doctor), three different mask types were obtained: tumor mask (tumor region), MCG mask (minimum coverage graphic region of the tumor covering the tumor and peritumoral area), and peritumoral mask (peritumoral region). For each mask, the corresponding ROI was obtained by performing a pixel-by-pixel logical AND operation between the mask and its associated image. Subsequently, the feature extractor function of the PyRadiomics library was applied to each region of interest, for example, resulting in 1032 conventional radiomics features.

[0050] Specifically, to eliminate redundancy and remove features with low information content, a two-stage feature selection process can be employed. First, an independent samples t-test is used for initial screening. Next, the features retained by the t-test are used as the initial candidate set for Least Absolute Shrinkage and Selection Operator (LASSO) regression, and five-fold cross-validation is combined to optimize feature selection.

[0051] Furthermore, the selected features are ranked according to the absolute value of their weights, reflecting their impact on the target variable in the training set. Then, the top 20 features from each ROI are retained as the most informative subset for subsequent analysis, including features from individual ROIs and combined feature sets obtained by merging features from multiple ROIs. In the case of combined ROIs, the total number of representative features used for model development can be up to 40.

[0052] Optionally, performance includes accuracy. In step 103, the performance of the ML classifier in outputting lesion analysis results under different feature input methods is evaluated, including: Step 1032: Based on various classifiers, evaluate the accuracy of the lesion analysis results output under different feature input methods, and assess the performance of the ML classifier in outputting lesion analysis results under different feature input methods. Accordingly, in step 104, based on the feature input method used to achieve optimal performance of the ML classifier, the optimal region selection method is determined, including: Step 1041: Determine the optimal region selection method based on the feature input method used by the ML classifier when the accuracy is highest.

[0053] In the above embodiments of this application, the area under the feature curve (AUC) is used as the core evaluation index to systematically optimize the medical image lesion analysis process. Focusing on the key index of AUC, step 1032 quantifies the classifier performance under different feature input methods to ensure that the evaluation results are intuitive and reliable, providing a clear basis for subsequent optimization. Step 1041 uses the highest AUC to deduce the best feature input method, thereby determining the optimal region selection method, realizing a precise mapping from feature performance to region selection and ensuring the scientific nature of region positioning. To this end, a closed-loop optimization logic of "feature input - classification performance - region selection" is formed, reducing reliance on human experience, promoting the standardization and automation of the lesion analysis process, effectively improving the efficiency and accuracy of medical image diagnosis, and providing more reliable technical support for clinical decision-making.

[0054] Optionally, the ML classifier includes logistic regression, perceptron, support vector machine, K-nearest neighbors, and Naive Bayes. In step 104, based on the feature input method that enables the ML classifier to achieve optimal performance, the optimal region selection method is determined, including: Step 1042: Evaluate the types of ML classifiers and the feature input methods used to achieve the best performance of the ML classifier, and determine the optimal region selection method based on the feature input methods used to achieve the best performance. Accordingly, based on the feature input method under the optimal region selection approach, the features are input into the pre-trained ML classifier to obtain the lesion analysis results, including: Step 1043: Based on the feature input method under the optimal region selection method, input the features into the ML classifier of the type with the best performance to obtain the lesion analysis results.

[0055] In the embodiments described above, after feature extraction and filtering, the retained features are used to construct various ML models for identifying cancer lesions, thereby evaluating the diagnostic capability of the proposed APASA strategy. Five commonly used ML classifiers covering various algorithm types can be selected to more comprehensively evaluate the sensitivity of different modeling approaches to feature information obtained from images. These ML classifiers include: Logistic Regression (LR), Perceptron, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Naïve Bayes. All models employ grid search combined with 10-fold cross-validation to select optimal parameters. This systematic evaluation, encompassing classification algorithms from simple to complex, assesses the generalization and robustness of multi-tumor region features in differentiating between benign and malignant lesions, highlighting their inherent discriminative power.

[0056] Optionally, the performance of the ML classifier is evaluated using classification metrics, including area under the curve, accuracy, specificity, sensitivity, and F1 score. The area under the curve represents the overall discriminative ability of the ML classifier, accuracy represents the correct classification rate of the ML classifier, specificity represents the ability of the ML classifier to correctly identify negative cases, sensitivity is represented by recall, which represents the ability of the ML classifier to correctly identify positive cases, and the F1 score represents the balance between precision and recall of the ML classifier. Negative cases represent benign cases, and positive cases represent malignant cases.

[0057] In the embodiments described above, the performance of the implemented ML model is evaluated using five widely adopted classification metrics: Area Under the Curve (AUC), Accuracy, Specificity, Sensitivity, and F1 score. These metrics provide a comprehensive evaluation of model performance, covering multiple dimensions: overall discriminative power (AUC), correct classification rate (Acc), ability to correctly identify negative cases (Specificity), ability to correctly identify positive cases (Sensitivity), and the balance between precision and recall (F1 score). The F1 score is particularly crucial in scenarios with data imbalance. The comprehensive use of these five metrics enables a multi-dimensional evaluation of experimental results, encompassing overall performance, class discrimination, and clinical risk management, thereby validating the effectiveness of the proposed AI-driven framework in automated cancer lesion analysis.

[0058] Therefore, the APASA application proposed in the above embodiments of this application generates three ROIs: tumor ROI (tumor region), peritumoral ROI (peritumoral region), and MCGROI (minimum coverage graphic region of the tumor). For MCGROI, two shape types (circular and rectangular) were used, resulting in five main ROIs: tumor ROI, MCRROI, MCR-peri-ROI, MCCROI, and MCC-peri-ROI. Furthermore, for each MCG shape type, a combined feature set was generated by combining the features of the peritumoral ROI with the features of the corresponding tumor ROI, resulting in two additional cases: combined MCR (tumor ROI + MCR-peri-ROI) and combined MCC (tumor ROI + MCC-peri-ROI). Therefore, for each of the two cancer types considered, performance was evaluated in seven scenarios: tumor ROI, MCRROI, MCR-peri-ROI, MCCROI, MCC-peri-ROI, combined MCR, and combined MCC.

[0059] Specifically, the core concept of the adaptive peritumoral region selection method is the minimum coverage map (MCG). In the embodiments described above, two MCG shapes are implemented: rectangular (MCR) and circular (MCC). For the same type of cancer lesion, the optimal ML model may differ depending on the selected MCG shape. For example, in breast lesion identification, the combination of MCR and LR models performs best, while the combination of MCC and Perceptron performs best. Similarly, for thyroid lesions, the optimal model for MCR is SVM, and the optimal model for MCC is LR.

[0060] Specifically, the geometry of the peritumoral region plays a crucial role in shaping the extracted feature spectrum, thus affecting classification performance. Rectangular and circular MCGs capture different spatial patterns and texture variations in the tumor-peripheral environment, resulting in differences in the quantity and nature of the acquired radiomics features. Since shape descriptors are essential in malignancy assessment, these differences can significantly impact the specificity and sensitivity of the predictive model. Therefore, optimizing the combination of MCG shape and ML model according to specific cancer lesion types is crucial for achieving higher diagnostic accuracy.

[0061] Furthermore, the impact of peritumoral features on the differential diagnosis of lesions across different cancer types varies, and the characteristics of the peritumoral microenvironment provide valuable supplementary information for malignant tumor assessment and diagnostic decisions. Traditionally, peritumoral features have been considered a secondary, rather than primary, source of information. However, depending on the cancer type, sometimes relying solely on peritumoral features can produce more accurate predictions than using only intratumoral features, or even a combination of both. This phenomenon may occur when the tumor is small or its internal features do not clearly indicate whether it is malignant or benign. In these cases, the surrounding tissue may provide richer information, thus better distinguishing the lesion type. Including intratumoral features along with peritumoral features may introduce noisy or conflicting data, which can interfere with the ML classifier and reduce predictive accuracy. Moreover, the effectiveness of combining these features depends on the specific biases and principles of the machine learning model used. Therefore, feature selection should be tailored to each specific cancer type, as in some cases, the peritumoral microenvironment may be the most informative region for accurately distinguishing lesions.

[0062] The shape of the Region of Interest (ROI) and the type of lesion also affect model performance. The best-performing model varies depending on the selected ROIs (tumor, peritumoral, or a combination of both), shape strategy (MCC or MRC), and cancer type (breast cancer or thyroid cancer). For example, in some ROI configurations, SVM achieves the highest AUC and F1-score under both MCC and MRC, but other models perform better in different ROI and shape combinations, highlighting that no single model can consistently dominate under all conditions. This specificity highlights two key points. First, it reflects that different ROI configurations capture complementary but not identical information: tumor regions provide direct lesion-specific cues, peritumoral regions capture context and microenvironment patterns, and combined ROIs integrate both. Therefore, the discriminative power of each ROI may have varying degrees of alignment with the learning strengths of different models. Second, the inconsistency between MCC and MRC suggests that the geometric representation of a region can affect feature distribution, thereby altering the applicability of the model. These results highlight that the shape of ROIs and the type of lesion can influence which model achieves the highest performance, underscoring the need for careful model selection based on specific scenarios.

[0063] By applying the technical solution of this embodiment, automated analysis of breast and thyroid cancer lesions can be performed. The Adaptive Peritumoral Region Selection (APASA) method proposed in the above embodiments of this application addresses the challenge of defining effective Regions of Interest (ROIs) for cancer lesion analysis. By adaptively identifying the surrounding region that best captures contextual information, APASA can extract more discriminative features, thereby improving diagnostic results. The effectiveness of this method is demonstrated by integrating it into AI-based frameworks for breast and thyroid cancer lesion identification, where it consistently outperforms commonly used morphological dilation operations, achieving significant performance improvements across different ROI settings. Beyond improvements in quantitative metrics, it provides important insights into model selection and ROI configuration, offering practical guidance for designing more clinically relevant decision support systems. By reducing reliance on heuristic ROI selection, APASA establishes a principled mechanism for incorporating tumor and peritumoral information into cancer diagnosis. These findings highlight the promise of the proposed strategy as a reliable ROI definition tool, with the potential to be extended to other imaging modalities and clinical applications where contextual information plays a crucial role, ultimately contributing to improved clinical decision-making.

[0064] Furthermore, as Figure 1 In terms of specific implementation, this application provides an automated tumor lesion analysis platform, such as... Figure 5 As shown, the platform includes: The ultrasound image acquisition module 201 is used to acquire multiple ultrasound images involving the same type of tumor for ultrasound images containing tumors. The ultrasound images are labeled with tumor type, tumor area pre-labeled by the doctor and lesion analysis results. The lesion analysis results include benign and malignant, and different tumor types correspond to preset minimum coverage graphics. The region of interest identification module 202, for any ultrasound image, determines the minimum coverage area of ​​the tumor based on the minimum coverage area of ​​the tumor region and the tumor type marked on the ultrasound image, and subtracts the tumor region from the minimum coverage area of ​​the tumor to obtain the peritumoral region. The classifier performance evaluation module 203 is used to extract features from the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region in sequence. Based on the extracted features, four feature input methods are constructed and input into the ML classifier for training. The performance of the ML classifier in outputting lesion analysis results under different feature input methods is evaluated. The trained ML classifier outputs lesion analysis results for the input features. The feature input methods include inputting only the features of the tumor region, inputting only the features of the minimum coverage area of ​​the tumor, inputting only the features of the peritumoral region, and inputting the features of the tumor region and the peritumoral region in parallel. The automated tumor lesion analysis module 204 is used to determine the optimal region selection method based on the feature input method used to achieve the best performance of the ML classifier. After the doctor performs the initial tumor region annotation on the newly input ultrasound image, the module extracts features of the target region based on the initially annotated tumor region and the optimal region selection method. The features are then input into the trained ML classifier based on the feature input method under the optimal region selection method to obtain the lesion analysis results. The target region includes the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region.

[0065] It should be noted that other corresponding descriptions of the functional units involved in the automated tumor lesion analysis platform provided in this application embodiment can be found in the following references. Figure 1 The corresponding descriptions in the method will not be repeated here.

[0066] Based on the above, Figure 1 Accordingly, this application also provides a storage medium storing a computer program, which, when executed by a processor, implements the above-described method. Figure 1 The automated tumor lesion analysis method shown is illustrated.

[0067] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive) and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in the various implementation scenarios of this application.

[0068] Based on the above, Figure 1 The method shown, and Figure 5 To achieve the above objectives, the virtual platform embodiment shown in this application also provides a computer device, specifically a personal computer, server, network device, etc. This computer device includes a storage medium and a processor; the storage medium stores computer programs; the processor executes the computer programs to achieve the above-described objectives. Figure 1 The automated tumor lesion analysis method shown is illustrated.

[0069] Optionally, the computer device may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB ports, card reader ports, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Bluetooth interfaces, Wi-Fi interfaces), etc.

[0070] Those skilled in the art will understand that the computer device structure provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0071] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages and stores the hardware and software resources of a computer device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software within the physical device.

[0072] Through the above description of the implementation methods, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented using hardware to determine the minimum coverage area of ​​the tumor region and tumor type based on the tumor region and tumor type annotated by ultrasound images. The minimum coverage area of ​​the tumor is then determined, and the tumor region is subtracted from the minimum coverage area to obtain the peritumoral region. After sequentially extracting the features of the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region, four feature input methods are constructed and input into a trained ML classifier. Based on the feature input method that maximizes the performance of the ML classifier, the optimal region selection method is determined, allowing doctors to obtain lesion analysis results based on the initially annotated tumor region and the optimal region selection method. Through multi-region feature extraction and classifier optimization, accurate analysis of tumor benignity and malignancy is achieved, improving the efficiency and accuracy of ultrasound image diagnosis.

[0073] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the platform within the embodiment can be distributed within the platform as described in the embodiment, or they can be modified to reside in one or more platforms different from this embodiment. The modules in the above-described embodiment can be merged into one module, or further divided into multiple sub-modules.

[0074] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any modifications that can be made by those skilled in the art should fall within the protection scope of this application.

Claims

1. An automated method for analyzing tumor lesions, characterized in that, The method includes: For ultrasound images containing tumors, multiple ultrasound images involving the same type of tumor are acquired. The ultrasound images are labeled with tumor type, tumor area pre-labeled by the doctor, and lesion analysis results. The lesion analysis results include benign and malignant tumors. Different tumor types correspond to preset minimum coverage graphics. For any ultrasound image, based on the minimum coverage pattern corresponding to the tumor region and tumor type marked on the ultrasound image, the minimum coverage pattern region of the tumor is determined, and the tumor region is subtracted from the minimum coverage pattern region of the tumor to obtain the peritumoral region. After extracting features from the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region in sequence, four feature input methods are constructed based on the extracted features. These methods are then input into the ML classifier for training. The performance of the ML classifier in outputting lesion analysis results under different feature input methods is evaluated. The trained ML classifier outputs lesion analysis results for the input features. The feature input methods include inputting only the features of the tumor region, inputting only the features of the minimum coverage area of ​​the tumor, inputting only the features of the peritumoral region, and inputting the features of both the tumor region and the peritumoral region in parallel. Based on the feature input method used to achieve optimal performance of the ML classifier, the optimal region selection method is determined. After the doctor performs initial tumor region annotation on the newly input ultrasound image, the target region extraction features are formed based on the initially annotated tumor region and the optimal region selection method. The features are then input into the trained ML classifier based on the feature input method under the optimal region selection method to obtain the lesion analysis results. The target region includes the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region.

2. The method according to claim 1, characterized in that, Minimum coverage patterns include rectangles and circles. The determination of the minimum coverage pattern region for a tumor, based on the tumor region and tumor type annotated by the ultrasound image, includes: When the minimum coverage shape determined based on the tumor type is a rectangle, the tumor region marked by the ultrasound image is transformed into a discrete set of coordinate points. Based on the coordinate points corresponding to the minimum horizontal coordinate, maximum horizontal coordinate, minimum vertical coordinate, and maximum vertical coordinate in the set of coordinate points, the minimum coverage shape region of the tumor region is constructed. When the minimum coverage pattern determined based on the tumor type is a circle, the tumor region marked on the ultrasound image is converted into a discrete point set. The Weltz algorithm is used to solve for the minimum coverage circle of the discrete point set to obtain the center and radius. Then, a circle is drawn on the ultrasound image based on the solved center and radius as the minimum coverage pattern region of the tumor region.

3. The method according to claim 1, characterized in that, Performance includes accuracy, and the evaluation of the ML classifier's output of lesion analysis results under different feature input methods includes: Based on various classifiers, the accuracy of outputting lesion analysis results under different feature input methods is evaluated to assess the performance of the ML classifier in outputting lesion analysis results under different feature input methods. Accordingly, determining the optimal region selection method based on the feature input method used to achieve the best performance of the ML classifier includes: Based on the feature input method used by the ML classifier that achieves the highest accuracy, the optimal region selection method is determined.

4. The method according to claim 3, characterized in that, The ML classifiers include logistic regression, perceptron, support vector machine, K-nearest neighbors, and Naive Bayes. The method for determining the optimal region selection based on the feature input method used by the ML classifier to achieve the highest accuracy includes: The evaluation assesses the types of ML classifiers and the feature input methods used to achieve optimal performance, and determines the optimal region selection method based on the feature input methods used to achieve optimal performance. Accordingly, based on the feature input method under the optimal region selection approach, the features are input into the pre-trained ML classifier to obtain the lesion analysis results, including: Based on the feature input method under the optimal region selection method, the feature input is the ML classifier of the type with the best performance, and the lesion analysis results are obtained.

5. The method according to claim 1, characterized in that, The sequential extraction of features from the tumor region, the minimum tumor coverage area, and the peritumoral region includes: For any region among the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region, the feature extractor function of the PyRadiomics library is applied to extract the features of the region.

6. The method according to claim 5, characterized in that, The feature extractor functions of the PyRadiomics library are used to extract features from the region, including: The initial features of the region are extracted using the feature extractor functions of the PyRadiomics library. Based on the initial features, determine the initial candidate set for minimum absolute shrinkage and selection operator regression; By combining the initial candidate set and five-fold cross-validation, the feature selection is optimized. The optimized features are then sorted according to their absolute weights, and the features of the region are determined based on the sorting results.

7. The method according to claim 1, characterized in that, The performance of the ML classifier is evaluated using classification metrics, including area under the curve (AUC), accuracy, specificity, sensitivity, and F1 score. The AUC represents the overall discriminative ability of the ML classifier, accuracy represents the correct classification rate, specificity represents the ability of the ML classifier to correctly identify negative cases, sensitivity is represented by recall, which represents the ability of the ML classifier to correctly identify positive cases, and the F1 score represents the balance between precision and recall. Negative cases represent benign cases, and positive cases represent malignant cases.

8. An automated tumor lesion analysis platform, characterized in that, The platform includes: The ultrasound image acquisition module is used to acquire multiple ultrasound images involving the same type of tumor from ultrasound images containing tumors. The ultrasound images are labeled with tumor type, tumor area pre-marked by the doctor, and lesion analysis results. The lesion analysis results include benign and malignant tumors. Different tumor types correspond to preset minimum coverage graphics. The region of interest identification module, for any ultrasound image, determines the minimum coverage area of ​​the tumor based on the minimum coverage area and tumor type marked on the ultrasound image, and subtracts the tumor area from the minimum coverage area of ​​the tumor to obtain the peritumoral region; The classifier performance evaluation module is used to extract features from the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region in sequence. Based on the extracted features, four feature input methods are constructed and input into the ML classifier for training. The module evaluates the performance of the ML classifier in outputting lesion analysis results under different feature input methods. The trained ML classifier outputs lesion analysis results for the input features. The feature input methods include inputting only the features of the tumor region, inputting only the features of the minimum coverage area of ​​the tumor, inputting only the features of the peritumoral region, and inputting the features of the tumor region and the peritumoral region in parallel. The automated tumor lesion analysis module is used to determine the optimal region selection method based on the feature input method that enables the ML classifier to perform at its best. After the doctor performs initial tumor region annotation on the newly input ultrasound image, the module extracts features from the target region based on the initially annotated tumor region and the optimal region selection method. The features are then input into the pre-trained ML classifier based on the feature input method under the optimal region selection method to obtain the lesion analysis results. The target region includes the tumor region, the minimum coverage area of ​​the tumor, and the peritumoral region.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method for automated tumor lesion analysis as described in any one of claims 1 to 7.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for automated tumor lesion analysis as described in any one of claims 1 to 7.

Images