Retrieval-Augmented Generation-Enabled Multimodal Large Language Model for Histopathologic Grading of Cutaneous...

Area of Science:

• Dermatopathology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Histopathologic grading is crucial for cutaneous lesion management.
• Inter-observer variability in grading cutaneous squamous cell carcinoma (cSCC) and melanocytic nevus dysplasia poses a challenge.
• Retrieval-augmented generation (RAG) AI may assist in dermatopathology education and diagnosis.

Purpose of the Study:

• To evaluate the performance of RAG-assisted AI in grading cSCC differentiation and melanocytic nevus dysplasia.
• To compare AI performance against expert dermatopathologist grading.
• To assess the potential of AI in reducing inter-observer variability in histopathologic grading.

Main Methods:

• Two RAG pathways utilizing Claude 4.5 Opus were developed, grounded with ChromaDB vector retrieval.
• AI models graded 60 cSCC cases for differentiation and 67 nevus cases for dysplasia.
• Concordance and inter-rater reliability (Cohen's kappa) were calculated against reference dermatopathologist grades.

Main Results:

• High concordance (90.0%) and excellent agreement (κ = 0.85) were achieved for cSCC differentiation.
• Low concordance (44.8%) and chance-level agreement (κ = 0.072) were observed for nevus dysplasia grading.
• AI misclassifications for nevi skewed towards moderate dysplasia, with no severe dysplasia cases correctly identified.

Conclusions:

• RAG-assisted AI performance in dermatopathology grading is task-dependent.
• AI demonstrated high accuracy for cSCC differentiation but low accuracy for nevus dysplasia.
• Task-specific validation is essential before implementing AI tools in clinical or educational dermatopathology settings.