Ultrasonic Spectral Information Promotes Specificity and Visualization of Breast Cancer in Deep Learning

Computer-aided diagnosis is important for breast ultrasound evaluation, as it can enhance diagnostic efficiency and objectivity. However, the performance of such systems, particularly deep learning methods, has been hindered by the limited resolution and interpretability of traditional B-mode ultrasound. We hypothesized that ultrasonic spectral information can serve as a representative feature for differentiating tissues. By leveraging machine learning to recognize tissue spectral patterns, we differentiated tissues in breast ultrasound with an AUCROC of 0.832 and enabled the color enhancement of the malignant probability in lesion areas. Moreover, we explored the integration of tissue spectral pattern recognition with deep learning vision models, resulting in a specificity of 0.896 and an AUCROC of 0.873 for lesion classification, representing a significant improvement.