Objective: We evaluated whether a deep-learning model could predict the response to neoadjuvant chemotherapy (NAC) in breast cancer using the pre-treatment B-mode ultrasound. Methods: This retrospective study included 245 female patients (253 lesions) treated with NAC between 2017 and 2019. Lesions were categorized as complete response (CR; 103) or non-CR (150) based on postoperative pathology. We trained ResNet18-based models using pre-treatment B-mode ultrasound images (Image) and clinical features. Three training strategies were evaluated: training from scratch (SC); transfer learning (TL); and domain-specific pretraining (USP). Predictive performance was assessed using descriptive statistics. Results: The best-performing model (USP Image) achieved 0.76 accuracy (specificity: 0.80; sensitivity: 0.72), significantly outperforming all other models, including those that used additional clinical features (p<0.05). USP improved performance across most model types compared to SC and TL, highlighting the value of domain-specific pretraining. Clinical features added value with SC or TL, but not with USP, suggesting that pretrained models can extract the most relevant information directly from images. Grad-CAM analysis revealed that non-CR predictions focused on the tumor and posterior shadowing—features linked to chemoresistant subtypes. CR predictions focused mainly on more heterogeneous, peritumoral regions. Conclusion: This finding underscores ultrasound’s potential as a low-cost, accessible tool for predictive oncology in personalized, AI-driven treatment planning.
Fürböck et al. (Thu,) studied this question.