Establishing dermatopathology encyclopedia DermpathNet with Artificial Intelligence-Based Workflow

Area of Science:

• Dermatopathology
• Medical Informatics
• Computational Biology

Background:

• Clinicians and trainees face challenges accessing high-quality, open-access dermatopathology image datasets for education and research.
• Existing resources often lack comprehensive annotation or are not readily available for machine learning applications.

Purpose of the Study:

• To establish a comprehensive, open-access dermatopathology image dataset for educational, cross-referencing, and machine-learning purposes.
• To develop and validate a robust workflow for curating and categorizing dermatopathology images.

Main Methods:

• A hybrid workflow combining deep learning-based image modality classification and figure caption analysis was employed.
• Images were curated and categorized from the PubMed Central (PMC) repository using specific keywords.
• The workflow's robustness was validated on 651 manually annotated images.

Main Results:

• The hybrid approach achieved an F-score of 90.4%, significantly outperforming deep learning (89.6%) and keyword retrieval (61.0%) alone.
• Over 7,772 images covering 166 diagnoses were retrieved and annotated.
• The developed dataset, DermpathNet, was found to be a challenging benchmark for current image analysis algorithms.

Conclusions:

• A large, peer-reviewed, open-access dermatopathology image dataset (DermpathNet) has been successfully developed.
• The semi-automated curation workflow provides a scalable method for creating valuable medical image datasets.
• DermpathNet serves as a crucial resource for dermatopathology education, research, and the advancement of AI in medical image analysis.