Computer Vision for Predicting the Efficacy of Neoadjuvant Therapy in Breast Cancer

Neoadjuvant therapy (NAT) is a standard component of breast cancer treatment, yet response rates vary substantially across patients. Accurate prediction of pathological complete response remains an unmet clinical need to improve patient selection for NAT. This review summarizes current approaches of using computer vision to predict breast cancer response to NAT from histopathological slides. We examined studies employing computer vision and machine learning models on hematoxylin and eosin and immunohistochemically stained whole-slide images, focusing on morphological features of tumor cells, stroma and tumor-infiltrating lymphocytes associated with pathological complete response. Key morphological predictors of therapy resistance included low tumor cell density with cord-like patterns, necrosis, predominance of collagenous and fibroblast-rich stroma and tumor vascularization, while therapy sensitivity was associated with high nuclear staining intensity, high tumor cell density and lymphocyte infiltration. We highlighted the advantages of incorporating multimodal data to enhance predictive performance. Our analysis demonstrates that computer vision models can detect subtle morphological patterns that may be difficult for pathologists to evaluate, providing valuable insights for personalized therapy planning in breast cancer. Further development of cross-modal, interpretable artificial intelligence solutions may improve prediction accuracy and deepen our understanding of tumor biology relevant to NAT response.