Establishing dermatopathology encyclopedia DermpathNet with Artificial Intelligence-Based Workflow.
This study addresses the scarcity of high-quality, openly accessible image datasets in dermatopathology, which hinders both clinical education and machine learning research. To overcome this limitation, the authors propose a hybrid AI workflow that integrates deep learning–based image classification with textual analysis of figure captions to automatically retrieve, filter, and annotate dermatopathological images from PubMed Central. The pipeline incorporates expert review to establish a semi-automated dataset curation process, resulting in the release of DermpathNet—an open-access dataset comprising 7,772 images spanning 166 diagnostic categories. The hybrid filtering approach achieves an F-score of 90.4%, demonstrating high precision and recall. Beyond providing a valuable resource for educational and algorithmic development purposes, this work also highlights the performance limitations of general-purpose AI models in the specialized domain of dermatopathology.