Adaptive diagnostic reasoning framework for pathology with multimodal large language models - PubMed
4 days ago
- #Diagnostic Reasoning
- #AI in Pathology
- #Multimodal LLMs
- Developed an adaptive diagnostic reasoning framework for pathology using multimodal large language models.
- Aims to enhance transparency in AI-driven pathology by generating evidence-linked reasoning for diagnostic auditing.
- Utilizes a two-phase self-learning process with small labeled datasets from breast and prostate cancer without updating model weights.
- Incorporates expert feedback from board-certified pathologists to align with medical standards.
- Achieves over 90% accuracy in distinguishing normal tissue from invasive carcinoma and differentiates complex subtypes like ductal carcinoma in situ.
- Generates audit-ready rationales that match expert assessments, improving over conventional baselines.
- Demonstrates adaptability across diverse tissue types and independent foundation models.
- Provides a transparent, understandable reasoning process for doctors to verify automated results, enhancing patient safety.