Channel attention of the image for distinguishing ductal carcinoma in situ from benign mimics based on ultrasound images

Background: Ductal carcinoma in situ (DCIS) is widely recognized as a precursor to invasive breast cancer, with an estimated 20–50% of cases progressing to invasive disease. The early detection of DCIS could save doctors and patients more time to deal with the breast cancer, which has posed a great threat to the life safety of women worldwide. This study aimed to develop a deep learning algorithm based on the image-level attention to distinguish DCIS from the benign mimics on ultrasound images, assisting doctors in screening for benign breast lesions.

Methods: By screening the biopsy results of 3,327 patients, the ultrasound images of 612 patients were retrospectively collected, including DCIS and benign mimics. The dataset was randomly split at the patient level into training (70%, n=430) and test (30%, n=182) sets. Using ResNet18 as the backbone architecture, we compared and developed models with image-level attention mechanisms. Model performance was evaluated on the test set using the area under the curve (AUC), accuracy, sensitivity, specificity, precision, and F1-score. To assess the statistical significance of performance differences between models, the Holm-Bonferroni correction was employed to account for multiple comparisons.

Results: Through quantitative comparison with 6 shallow benchmark algorithms such as AlexNet, VGG16, GoogleNet, ResNet18, CoAtNet and Swin Transformer, the effectiveness of our method was demonstrated, which achieved the best performance with AUC, accuracy, sensitivity, specificity, precision and F1-score of 0.868±0.009, 0.835±0.016, 0.767±0.056, 0.916±0.057, 0.916±0.059, 0.835±0.04, respectively. After Holm-Bonferroni correction, these improvements were statistically significant (corrected P<0.05) compared to all other models. The effect size of AUC, specificity, precision, F1-score were 0.69, 0.68, 0.89, 0.58, respectively, revealing large practical improvement. Furthermore, Gradient-weighted Class Activation Mapping (GradCAM) visualization confirmed that our model focused more accurately on lesion regions, providing interpretability for its superior feature extraction.

Conclusions: In this paper, we proposed the ResNet18-CA for the classification of DCIS and confused cases, which was good at learning the ultrasound image features. The proposed network achieved higher accuracy than the other methods being compared.