From Label-Free Multiphoton Imaging to Pathological Reports: A Vision-Language Breast Cancer Margin Pathological Diagnosis System.

Margin pathological assessment provides critical feedback for breast-conserving surgery, whereas H&E-stained histopathology focuses on residual tumor and may cause unnecessary resections. Label-free multiphoton microscopy (MPM) reveals tumor-associated collagen signatures (TACS) at the margin, offering a prognostically relevant diagnostic information. However, limited familiarity of novel MPM images by pathologists has prevented its integration into diagnostic workflows. Here, we introduce MarginPath, a Margin Pathological diagnosis system built on an MPM-language model that requires only a single label-free section. By integrating MPM-derived TACS with virtual H&E diagnostic information, MarginPath provides a multimodal diagnostic report, including: (i) a MPM image and corresponding virtual H&E image, (ii) a TACS-based pixel-level margin-status heatmap, and (iii) detailed natural-language diagnostic descriptions of tumor margin microenvironment. Validated on 158 invasive breast cancer specimens, MarginPath outperforms existing pathology vision-language models in margin diagnosis and can be extended into a question-answering system, enhancing both clinical decision-support and patient communication.