Comparative evaluation of ResNet-Based AI models for breast cancer detection in digital mammography and digital breast tomosynthesis

Introduction Early detection of breast cancer is crucial for better treatment options; however, physicians usually encounter challenges when reading breast mammograms because some tissues can overlap or look similar. New computer technologies can help by highlighting areas that may contain cancer. In this study, we assessed computer systems to identify whether a breast image is normal, benign or malignant and compared their performance on standard digital mammography (DM) and digital breast tomosynthesis (DBT) images. Methods This retrospective study included 31 (100 after augmentation) images with malignant breast lesions, 175 with benign findings, and 475 normal images. We developed a deep learning convolutional neural network (CNN) model to classify images using ResNet-18 and ResNet-50. We trained the model via 10-fold cross-validation and Holdout validation. Performance was measured for each model. We computed performance metrics on original (non-augmented) images from 10-fold cross-validated predictions, with 95% confidence intervals from a stratified patient-level bootstrap (2,000 replicates), precision–recall analysis with average precision (AP), and paired DeLong and exact McNemar tests with Holm correction. We increased the malignant class from 31 to 100 images in the training set via data augmentation for the stage-2 classifier. Validation used non-augmented images. Results In normal versus abnormal classification (679 original images per view; ResNet-18, 10-fold cross-validation), DM CC, DM MLO and DBT MLO achieved an AUC-ROC of 0.995 (95% CI 0.991–0.998), 0.984 (0.975–0.992) and 0.994 (0.989–0.998), sensitivity of 95.1%, 95.6% and 95.1%, and specificity of 96.6%, 94.1% and 98.5%, respectively. DBT had significantly higher specificity than DM in selected comparisons (ResNet-18 vs. DM MLO and ResNet-50 vs. DM CC; Holm-adjusted p ≤ 0.002), with no significant AUC-ROC differences between modalities or architectures (all Holm-adjusted p ≥ 0.13). In benign versus malignant classification (131 original images, 31 malignant), AUC-ROC ranged from 0.977 to 0.999 and AP from 0.946 to 0.996, with no significant differences between DM and DBT or between ResNet-18 and ResNet-50 (all Holm-adjusted p ≥ 0.50). The detection and classification of malignant, benign, and normal images in DBT and DM using an artificial intelligence (AI) training model achieved comparable performance to the medical field. Conclusion Computer systems can detect and classify breast findings with high accuracy, helping to reduce errors and support physicians’ decisions. The original number of malignant images was relatively small (n = 31), which may affect the model’s generalizability and highlights the need for validation on larger, multi-institutional datasets. These findings may encourage more use of computer-assisted detection in breast imaging and contribute to developing better tools for early breast cancer diagnosis.