Development and validation of an intratumoral-peritumoral deep transfer learning fusion model for differentiating BI-RADS 3-4 breast nodules.

Background: The Breast Imaging Reporting and Data System (BI-RADS) 3-4 breast nodules present a diagnostic challenge, as some benign lesions lead to unnecessary biopsies. Traditional imaging modalities like mammography and ultrasound often yield false positives due to limited specificity. While radiomics and machine learning show potential for improving accuracy, most studies focus on intratumoral features, neglecting the diagnostic value of peritumoral regions (PTRs). This study aimed to develop a non-invasive tool integrating intratumoral and peritumoral deep transfer learning (DTL) features to enhance risk stratification.

Methods: Clinical data (age, tumor size), ultrasound images, and parameters [calcification, color Doppler flow imaging (CDFI), BI-RADS] were retrospectively collected from 555 patients with BI-RADS 3-4 nodules confirmed by pathology at two Shanghai medical centers. Patients from Center 1 (Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine) were split into training (n=291) and internal validation sets (n=125) at a 7:3 ratio, while those from Center 2 (Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine) formed an external validation set (n=139). Radiomics features from intratumoral and PTRs (5, 10, 20 voxels) were extracted using PyRadiomics, and DTL features were derived using a pre-trained ResNet-18 network. Combined features from DTL, radiomics, and clinical data were selected via least absolute shrinkage and selection operator (LASSO) regression. Machine learning models, including logistic regression (LR), random forest (RF), naive Bayes, K-nearest neighbors (KNN), and light gradient boosting machine (LightGBM), were constructed and compared using metrics like area under the curve (AUC). Ultrasound physicians independently reviewed images, and their performance was compared with the models.

Results: The cohort included 555 female patients (mean age: 48.11±14.83 years), with 72.07% of nodules lacking calcifications and 61.08% without CDFI signals. The naive Bayes model based on intratumoral and 10-voxel peritumoral DTL features performed best. In the training set, it achieved an AUC of 0.911 (accuracy: 0.852, sensitivity: 0.852, specificity: 0.852). In the internal and external validation sets, AUCs were 0.909 and 0.910, respectively, outperforming physicians' AUCs of 0.722 and 0.745. The model also surpassed physicians in accuracy, sensitivity, specificity, and efficiency.

Conclusions: The DTL feature model integrating intratumoral and PTRs effectively predicts BI-RADS 3-4 nodule malignancy, outperforming ultrasound physicians. It aids in reducing unnecessary biopsies and improving treatment decisions.