Abstract

Artificial intelligence (AI) is poised to revolutionize healthcare, with pathology at the forefront of this digital transformation. The high-volume field of dermatopathology stands to benefit significantly from AI-driven advancements, potentially reshaping diagnostic processes. This study investigated whether a customized vision-language model could accurately diagnose and describe dermatopathology biopsies, aiming to complement human efforts with differential diagnoses and text descriptions of histopathologic findings. We developed a dermatopathology AI assistant trained on over 40,000 custom tailored images and captions, using a state-of-the-art open-source vision-language model. The model's performance was evaluated on a standardized set of histopathologically confirmed diagnoses, focusing on benign and malignant epidermal and melanocytic neoplasms. The baseline model achieved 50% accuracy, improving to 68% post-training. All trained models consistently identified common diagnoses such as basal cell carcinoma, squamous cell carcinoma, actinic keratosis, and melanoma and were able to provide accurate and detailed text descriptions of these diagnoses. However, less common inflammatory and neoplastic dermatoses were often misdiagnosed. This pilot study demonstrates the potential of advanced AI in dermatopathology through a sophisticated model capable of both interpreting images for diagnoses and generating detailed histopathological descriptions. Our results highlight both the promise and challenges of AI in dermatopathological practice. The ability to provide accurate textual descriptions alongside diagnoses represents a significant advancement, potentially streamlining workflow and enhancing diagnostic accuracy. Future work will focus on improving performance for less common conditions and integrating this technology into real-world dermatopathology settings.