Application and prospects of artificial intelligence-enhanced multimodal ultrasound in diagnosing breast non-mass lesions

This is the authors' abstract. We don't add key points for this paper.

Non-mass lesions (NMLs) of the breast lack the typical mass effect, which makes conventional ultrasound diagnosis difficult and often leads to missed or incorrect diagnoses. Multimodal ultrasound (MUS) integrates gray-scale ultrasound, elastography, contrast-enhanced ultrasound, and automated breast volume scanning to assess lesions from multiple angles—morphology, stiffness, blood flow perfusion, and spatial structure—greatly improving diagnostic accuracy. In recent years, rapid advances in AI have further driven the intelligent evolution of MUS. Studies show that multimodal ultrasound fusion strategies can achieve an AUC above 0.930 for distinguishing benign from malignant non-mass lesions. Adding deep learning pushes the AUC to 0.960, and machine learning–based multimodal models reach an AUC of 0.947 for molecular subtyping prediction. Moreover, ultrasound imaging features correlate to some extent with molecular subtypes such as Luminal, HER2-overexpressing, and triple-negative breast cancer. Emerging techniques like explainable artificial intelligence and multi-view visual Transformers also show promise. This paper systematically reviews the progress of integrating multimodal ultrasound with artificial intelligence for diagnosing non-mass breast lesions, focusing on deep learning models, radiomics, and molecular subtyping prediction, aiming to guide future research and clinical translation.