High-fidelity detection, subtyping, and localization of five skin neoplasms using supervised and semi-supervised...

Area of Science:

• Dermatopathology and Artificial Intelligence
• Computational Pathology
• Digital Health

Background:

• Skin cancers are common worldwide malignancies.
• Early detection of skin lesions improves outcomes, but pathology workflows face challenges like pathologist shortages and diagnostic discordance.
• Existing AI models for classifying skin lesions from whole slide images (WSIs) often do not surpass expert pathologist performance.

Purpose of the Study:

• To develop an AI model for detecting and classifying skin lesions with enhanced sensitivity.
• To create a system that potentially matches and surpasses expert pathologists' diagnostic capabilities.
• To improve clinical pathology workflows through advanced AI.

Main Methods:

• Combined supervised learning (SL) and semi-supervised learning (SSL) to create a multi-level skin detection system.
• Trained the system on a large dataset of 2188 supervised WSIs and 5161 weakly supervised WSIs.
• Validated and tested the AI model on curated and non-curated WSIs, including 'mimickers', assessing detection, subtyping, localization, and margin status.

Main Results:

• The AI model (Mihm) achieved high sensitivity in classifying key skin lesions: melanocytic lesions (98.91%), basal cell carcinoma (97.24%), atypical squamous lesions (95.26%), verruca vulgaris (93.50%), and seborrheic keratosis (86.91%).
• The multi-level algorithm accurately subtypes lesions, localizes diagnostic regions of interest (ROIs) via AI overlay, and reports margin status.
• The system demonstrated robust performance on a large testing set of 3821 WSIs.

Conclusions:

• The developed AI model, created with dermatopathologists, surpasses previous AI models in detecting five major skin lesion types.
• It offers critical information including subtyping, localization, and margin status through a digital display.
• This end-to-end system promises to enhance pathology workflows, improve diagnostic accuracy, expedite patient care, and improve outcomes.