A method for classifying benign and malignant breast intraductal lesions based on MRI

Abstract

This invention relates to an MRI-based method for classifying benign and malignant intraductal lesions (IDLs) of the breast, comprising the following steps: Step 1) Extracting a set of clinical features from MRI images: age, ADC value, BI-RADS category, lesion shape, margin clarity, and TIC curve type; Step 2) Extracting radiomics features from the 3D lesion segments of the T1 DCE-MRI sequence; Step 3) Reducing the dimensionality of the radiomics features using the Lasso algorithm; Step 4) Fusing the dimensionality-reduced radiomics features with the clinical features into an input vector; Step 5) Outputting benign / malignant probability values ​​through a logistic regression model; Step 6) Applying stratification rules to BI-RADS 4A lesions: older patients or those with low ADC values ​​are marked as high-risk and biopsy is recommended; younger patients with high ADC values ​​are marked as low-risk and follow-up is recommended. This method solves the problem of current dynamic contrast-enhanced MRI's difficulty in accurately identifying BI-RADS categories, especially the BI-RADS 4A high-risk subgroup, thereby reducing missed diagnoses and over-biopsy. It is of urgent significance for overcoming the bottleneck of preoperative risk stratification in IPLs and guiding individualized clinical decision-making.

Description

Technical Field

[0001] This invention relates to a medical artificial intelligence diagnostic technology, and particularly to a method and system for classifying benign and malignant intraductal papillary lesions (IPLs) of the breast based on the fusion of MRI radiomics features and clinical features, especially for risk stratification management of BI-RADS4A lesions. Background Technology

[0002] Intraductal papillary lesions (IPLs) are a class of epithelial lesions occurring in the mammary ductal system, characterized by papillary structures. Their spectrum ranges from benign intraductal papillomas (IDPs) to atypical hyperplasia with malignant potential, and even carcinoma in situ and invasive papillary carcinoma. Due to the high degree of overlap in histological morphology and imaging manifestations between benign and malignant IPLs, traditional diagnostic methods (including ultrasound, mammography, and MRI combined with biopsy) are insufficient for accurate differentiation, often leading to the clinical dilemma of overtreatment of benign lesions or delayed intervention for malignant lesions. Although dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), with its clear visualization of ductal structures and sensitive capture of lesion hemodynamic characteristics, has become the core tool for assessing IPLs, it still has diagnostic limitations when dealing with poorly defined or small lesions.

[0003] The emerging radiomics technology, through high-throughput extraction of quantitative features such as grayscale distribution, texture patterns, and morphology from medical images, has provided a new pathway for the precise diagnosis of breast cancer. However, its value in differentiating benign from malignant intravascular coagulation (IPLs) remains unclear, especially lacking systematic research on its integration with key clinical features such as age, ADC value (apparent diffusion coefficient), and BI-RADS classification (a classification system used for standardized evaluation and management of breast imaging results). Summary of the Invention To address the current challenge of accurately identifying BI-RADS classifications, especially the high-risk subgroup of BI-RADS 4A, using dynamic contrast-enhanced magnetic resonance imaging (MRI), thus reducing missed diagnoses and excessive biopsies, a new method for classifying benign and malignant intraductal lesions (IDLs) is proposed. This method is of urgent significance for overcoming the bottleneck of preoperative risk stratification for IDLs and guiding individualized clinical decision-making.

[0004] The technical solution of this invention is as follows: A method for classifying benign and malignant intraductal lesions of the breast based on MRI, comprising the following steps: Step 1) Extract clinical feature set from MRI images: age, ADC value, BI-RADS category, lesion shape, margin clarity, and TIC curve type; Step 2) Extract radiomics features from the 3D lesion segments of the T1 DCE-MRI sequence; Step 3) Use the Lasso algorithm to reduce the dimensionality of the image omics features; Step 4) Fuse the dimensionality-reduced radiomics features with clinical features into an input vector; Step 5) Output the probability values ​​of benign or malignant tumors using a logistic regression model; Step 6) Apply stratification rules to BI-RADS 4A lesions: elderly patients or those with low ADC values ​​are marked as high-risk and biopsy is recommended; younger patients with high ADC values ​​are marked as low-risk and further follow-up is recommended, specifically including: Step 1) Read the raw DICOM format data acquired by the MRI equipment, select the T1 DCE sequence and ADC map from it, and convert it into an initial image in NIfTI format; Step 2) Use 3D Slicer software to perform three-dimensional segmentation of the lesions in the initial image and generate a segmentation mask file; In this process, multiple radiologists independently performed lesion segmentation, and the intersection area was taken as the final segmentation volume; the segmentation layer thickness was 1mm, covering the entire three-dimensional volume of the lesion; Step 3) Extract radiomics features based on segmentation mask, and simultaneously obtain age, BI-RADS category, and morphological features from clinical records; Clinical features extracted include: age, lesion size, ADC value, BI-RADS category, lesion shape, edge clarity, duct dilation, nipple discharge, and TIC curve type. Radiomics Feature Extraction: Features were extracted from the segmentation mask using the PyRadiomics library. Feature extraction techniques included wavelet transform features and gray-level co-occurrence matrix features to obtain clinical features. Step 4) Integrate radiomics features and clinical features, output the probability of benign or malignant lesions using a logistic regression model, and perform risk stratification on BI-RADS 4A lesions; Step 4) specifically involves: Step 41) Feature fusion: The Lasso algorithm was used to reduce the dimensionality of radiomics features, with a regularization coefficient λ=0.01, retaining the top 10 features with non-zero weights; clinical features and radiomics features were concatenated into an input vector. Step 42) Classification model: Use a logistic regression model, set the L2 regularization strength C=0.1; output the malignancy probability value, and judge it as malignancy if the threshold is >0.5; Step 43) BI-RADS 4A Hierarchical Rules: If age > 50 years or ADC value ≤ 0.98 × 10 - ³mm² / s, marked as high-risk group, upgraded to biopsy management; If the age is ≤50 years and the ADC value is >0.98×10 - ³mm² / s, marked as low-risk group, 6-month follow-up recommended; Step 44) Diagnostic report generation: The output includes benign and malignant probability values, risk stratification suggestions, and key decision-making basis.

[0005] The beneficial effects of this invention are as follows: The fusion of clinical features (ADC weight 30.41%) and radiomics features (such as wavelet texture) resulted in a classification AUC of 0.88, a 35.4% improvement over the single radiomics model (AUC 0.65). A dual-parameter quantification rule for age-ADC was established to address the underestimation of malignancy rates in BI-RADS4A lesions (actual 20.97% vs. guideline ≤10%). Attached Figure Description

[0006] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the 3D Slicer used in this invention to delineate lesions; Figure 3 This is a schematic diagram of the image data analysis process of the present invention; wherein Figure 3 T1WI-DCE of a showed a high signal intensity nodule in the upper outer quadrant of the right breast (white arrow). Figure 3 In b, contour1 is the outlined lesion, and the average ADC value shown in the figure is 0.75×10-3 mm² / s; Figure 3 d represents the TIC curve. In the semi-quantitative breast map, the region of interest is selected and placed, and the time-signal intensity curve, i.e., the TIC curve, is observed to be type III. Detailed Implementation

[0007] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0008] A method for classifying intraductal lesions of the breast includes the following steps: (1) Extract clinical feature set from MRI images: age, ADC value, BI-RADS category, lesion shape, edge clarity, TIC curve type (time-signal intensity curve). (2) Extract radiomics features from 3D lesion segments in T1 DCE-MRI sequences; (3) Use the Lasso algorithm to reduce the dimensionality of image omics features; (4) The reduced radiomics features and clinical features are fused into an input vector; (5) Output the probability values ​​of benign and malignant diseases through the logistic regression model; (6) Apply stratification rules to BI-RADS 4A lesions: When age > 50 years or ADC value ≤ 0.98 × 10 - When the flow rate is 3 mm² / s, it is marked as high-risk and a biopsy is recommended; When age ≤ 50 years and ADC value > 0.98 × 10 - When the flow rate is 3 mm² / s, the individual is classified as low-risk and follow-up is recommended.

[0009] 1. For example Figure 1 As shown, the classification method for benign and malignant intraductal lesions of the breast includes the following steps: Step 1) Read the raw DICOM format data acquired by the MRI equipment, select the T1 DCE (Dynamic Contrast Enhanced) sequence and ADC map from it, and convert it into an initial image in NIfTI format; Step 2) Use 3D Slicer software to perform three-dimensional segmentation of the lesions in the initial image and generate a segmentation mask file; The lesion segmentation was performed independently by two radiologists, and the intersection area was taken as the final segmentation body; the segmentation layer thickness was 1mm, covering the entire three-dimensional volume of the lesion.

[0010] Step 3) Extract radiomics features based on segmentation mask, and simultaneously obtain age, BI-RADS category, and morphological features from clinical records; Clinical feature extraction includes: age (numerical value), lesion size (unit: mm), and ADC value (unit: ×10). - ³mm² / s, measuring the mean ROI of the lesion), BI-RADS category (2 / 3 / 4A / 4B / 4C / 5 / 6), lesion shape (0=regular, 1=irregular), edge clarity (0=clear, 1=blurred), duct dilation (0=none, 1=present), nipple discharge (0=none, 1=present), TIC curve type (1=type I, 2=type II, 3=type III). Radiomics Feature Extraction: The PyRadiomics library (an open-source Python package designed for radiomics feature extraction of medical images, version 3.0.1) was used to extract features from the segmentation mask. Feature extraction techniques such as wavelet transform features (e.g., wavelet-HHH_glszm_SmallAreaLowGrayLevelEmphasis) and gray-level co-occurrence matrix features (e.g., original_glcm_Imc1) were used to obtain clinical features.

[0011] Step 4) Integrate radiomics features and clinical features, output the probability of benign or malignant lesions through a logistic regression model, and perform risk stratification on BI-RADS 4A lesions. Step (4) specifically involves: Step 41) Feature fusion: The Lasso algorithm (regularization coefficient λ=0.01) was used to reduce the dimensionality of radiomics features, retaining the top 10 features with non-zero weights; the clinical features and radiomics features were concatenated into an input vector. Step 42) Classification model: Use a logistic regression model, set the L2 regularization strength C=0.1; output the malignancy probability value, and judge it as malignancy if the threshold is >0.5; Step 43) BI-RADS 4A Hierarchical Rules: If age > 50 years or ADC value ≤ 0.98 × 10 - ³mm² / s, marked as high-risk group (malignancy risk 32.7%-61.9%), upgraded to biopsy management; If the age is ≤50 years and the ADC value is >0.98×10 - ³mm² / s, marked as low-risk group (8% risk of malignancy), 6-month follow-up is recommended; Step 44) Diagnostic report generation: The output includes benign and malignant probability values, risk stratification suggestions, and key decision-making criteria (such as ADC contribution weight).

[0012] The effectiveness of the method for classifying benign and malignant lesions in intraductal breast disease of the present invention can be further illustrated by the following experiments.

[0013] Experimental design and results: Data source: 341 cases of intraductal breast lesions (227 benign, 114 malignant) treated at Xinhua Hospital affiliated with Shanghai Jiao Tong University School of Medicine from 2022 to 2024. Inclusion criteria: 1) Female patients, aged 20-80 years; 2) Preoperative breast MRI examination, with clear image quality of T1-weighted dynamic contrast-enhanced MRI (T1-DCE), diffusion-weighted imaging (DWI) sequences, and apparent diffusion coefficient (ADC) maps; 3) The ADC values ​​of the lesions corresponding to the postoperative pathological results can be obtained from the DWI sequence; 4) Possesses complete clinical, imaging, and postoperative paraffin pathology data; 5) Undergo surgical treatment; Data annotation: completed independently by two breast imaging specialists. Figure 2 ): Manually delineate the three-dimensional boundaries of the lesion on T1 DCE-MRI. The ADC value is taken as the mean of the lesion ROI (accurate to 0.01×10). - ³mm² / s). The specific image data analysis process is as follows: Figure 3 As shown, this image is of a typical patient with intraductal malignancy in the breast. Figure 3 T1WI-DCE showed a high signal intensity nodule (white arrow) in the upper outer quadrant of the right breast, measuring 8 mm in size. Contrast-enhanced CT showed heterogeneous enhancement, indistinct borders, and an irregular shape. Diffusion-weighted imaging (DWI) showed high signal intensity. Figure 3 b. Apparent diffusion coefficient (ADC) The average ADC value is 0.75 × 10⁻³ mm² / s. Figure 3 d. Select and place the region of interest in the semi-quantitative breast map, and observe the time-signal intensity curve, i.e., the type III TIC curve.

[0014] Experimental Design: The dataset was divided into 5 equal parts, and 5-fold cross-validation was used. In each round, 4 parts were used to train the logistic regression model, and 1 part was used for testing. This was repeated 5 times until each part was tested.

[0015] Experimental results: Table 1 shows the accuracy, sensitivity, specificity, and AUC of this method in the five-fold cross-validation.

[0016]

[0017] Table 2. Feature importance ranking and percentage for each feature in the logistic regression model.

[0018] The implementation of this protocol revealed a key clinical issue: the current BI-RADS system significantly underestimates the malignancy risk of class 4A IPL, with an actual malignancy rate of 20.97% (26 / 124), far exceeding the guideline expectation (≤10%). Follow-up alone would result in a 21% missed diagnosis of malignant lesions. Fusion model analysis revealed that age and ADC value were the most important among all characteristics, with age >50 years and ADC ≤0.98×10⁻⁶ being the most significant. - ³mm² / s is an independent high-risk factor for class 4A lesions—the malignancy risk rises to 32.7% (20 / 61) in the age > 50 years group, as high as 53.8% (14 / 26) in the ADC ≤ 0.98 group, and the malignancy risk surges to 61.9% (13 / 21) in the double positive group.

[0019] Based on this, this solution proposes a hierarchical management strategy: High-risk group (age > 50 years or ADC ≤ 0.98): malignancy risk 32.7%-61.9%, upgraded to biopsy management (treated as 4B / 4C category); Low-risk group (age ≤ 50 years and ADC > 0.98): malignancy risk is only 8% (5 / 59), and 6-month imaging follow-up is recommended.

[0020] 3. Experimental Analysis Based on the above results, we can draw the following conclusions: This method exhibits high performance in differentiating between benign and malignant tumors, with an accuracy of 0.82 and an AUC of 0.88, which are significantly higher than methods that rely solely on radiomics.

[0021] By using age-ADC dual-parameter stratification, the biopsy positivity rate in the high-risk group increased to 43.9% (36 / 82 validation set), which is 2.2 times higher than the traditional 4A management strategy; the follow-up safety of the low-risk group reached 92.9% (39 / 42 no malignant progression was observed), which confirms the dual advantages of this protocol in avoiding over-biopsy and reducing the rate of missed diagnoses.

[0022] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention should be determined by the appended claims.

Claims

1. A method for classifying benign and malignant intraductal lesions of the breast based on MRI, characterized in that, Includes the following steps: Step 1) Extract clinical feature set from MRI images: age, ADC value, BI-RADS category, lesion shape, margin clarity, and TIC curve type; Step 2) Extract radiomics features from the 3D lesion segments of the T1 DCE-MRI sequence; Step 3) Use the Lasso algorithm to reduce the dimensionality of the image omics features; Step 4) Fuse the dimensionality-reduced radiomics features with clinical features into an input vector; Step 5) Output the probability values ​​of benign or malignant tumors using a logistic regression model; Step 6) Apply stratification rules to BI-RADS 4A lesions: elderly patients or those with low ADC values ​​are marked as high-risk and biopsy is recommended; young patients with high ADC values ​​are marked as low-risk and follow-up is recommended.

2. The MRI-based method for classifying benign and malignant intraductal lesions of the breast according to claim 1, characterized in that, Specifically, it includes: Step 1) Read the raw DICOM format data acquired by the MRI equipment, select the T1 DCE sequence and ADC map from it, and convert it into an initial image in NIfTI format; Step 2) Use 3D Slicer software to perform three-dimensional segmentation of the lesions in the initial image and generate a segmentation mask file; In this process, multiple radiologists independently performed lesion segmentation, and the intersection area was taken as the final segmentation volume; the segmentation layer thickness was 1mm, covering the entire three-dimensional volume of the lesion; Step 3) Extract radiomics features based on segmentation mask, and simultaneously obtain age, BI-RADS category, and morphological features from clinical records; Clinical features extracted include: age, lesion size, ADC value, BI-RADS category, lesion shape, edge clarity, duct dilation, nipple discharge, and TIC curve type. Radiomics Feature Extraction: Features were extracted from the segmentation mask using the PyRadiomics library. Feature extraction techniques included wavelet transform features and gray-level co-occurrence matrix features to obtain clinical features. Step 4) Integrate radiomics features and clinical features, output the probability of benign or malignant lesions using a logistic regression model, and perform risk stratification on BI-RADS 4A lesions; Step 4) specifically involves: Step 41) Feature fusion: The Lasso algorithm was used to reduce the dimensionality of radiomics features, with a regularization coefficient λ=0.01, retaining the top 10 features with non-zero weights; clinical features and radiomics features were concatenated into an input vector. Step 42) Classification model: Use a logistic regression model, set the L2 regularization strength C=0.1; output the malignancy probability value, and judge it as malignancy if the threshold is >0.5; Step 43) BI-RADS 4A Hierarchical Rules: If age > 50 years or ADC value ≤ 0.98 × 10 - ³mm² / s, marked as high-risk group, upgraded to biopsy management; If the age is ≤50 years and the ADC value is >0.98×10 - ³mm² / s, marked as low-risk group, 6-month follow-up recommended; Step 44) Diagnostic report generation: The output includes benign and malignant probability values, risk stratification suggestions, and key decision-making basis.

Images