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

Area of Science:

• Pathology
• Medical Imaging
• Artificial Intelligence

Background:

• Pathological margin assessment is crucial for breast-conserving surgery.
• Current H&E staining may lead to unnecessary resections.
• Label-free multiphoton microscopy (MPM) detects tumor-associated collagen signatures (TACS) with prognostic value.

Purpose of the Study:

• To develop an AI system, MarginPath, for improved pathological diagnosis of tumor margins using MPM.
• To integrate MPM-derived TACS with virtual H&E imaging for multimodal diagnostics.
• To overcome pathologists' limited familiarity with novel MPM images.

Main Methods:

• Developed MarginPath, an MPM-language model system for margin diagnosis.
• Integrated MPM TACS with virtual H&E data for multimodal reports.
• Generated TACS-based margin-status heatmaps and natural-language descriptions.

Main Results:

• Validated MarginPath on 158 invasive breast cancer specimens.
• MarginPath demonstrated superior performance compared to existing pathology vision-language models in margin diagnosis.
• The system can be extended into a question-answering tool.

Conclusions:

• MarginPath offers a novel, accurate approach to pathological margin assessment in breast cancer.
• The system enhances clinical decision-support and patient communication.
• AI integration with advanced imaging techniques like MPM holds significant potential for diagnostic pathology.