Classification of large-scale image database of various skin diseases using deep learning

Area of Science:

• Artificial Intelligence
• Dermatology
• Medical Imaging

Background:

• Skin disease diagnosis relies heavily on visual inspection, which can be subjective.
• Computer-aided diagnosis (CADx) systems offer potential for objective and accurate classification.
• Classifying a large number of diverse skin conditions from varied photographic images presents a significant challenge.

Purpose of the Study:

• To develop a deep learning-based CADx system for classifying 59 distinct skin diseases.
• To enhance classification accuracy and reliability using photographic patient images.
• To address challenges posed by diverse disease appearances and imaging conditions.

Main Methods:

• Utilized ResNet-18 as a baseline classification model.
• Incorporated metric learning to improve model generalization and prevent overfitting.
• Implemented patient-wise classification through aggregation of inference vectors from multiple images per patient.

Main Results:

• Achieved significant improvements in patient classification accuracy: 0.579 (Top-1), 0.793 (Top-3), and 0.863 (Top-5).
• Demonstrated statistical significance of improvements using the McNemar test.
• Validated effectiveness on a large dataset of 70,196 images from 13,038 patients.

Conclusions:

• The developed deep learning system effectively classifies 59 skin diseases using multiple patient images.
• Metric learning and image aggregation significantly enhance classification performance.
• The system shows promise for aiding dermatologists in diagnosing diverse skin conditions under various imaging scenarios.