A Deep Learning Based Framework for Diagnosing Multiple Skin Diseases in a Clinical Environment.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• AI in dermatology primarily focuses on classification and segmentation, with limited research in real clinical settings.
• A need exists for AI tools adapted to clinical environments, especially for diverse patient populations like Asians.

Purpose of the Study:

• To construct a novel deep learning framework for classifying dermatological conditions.
• To train the AI model using a dataset reflecting real-world clinical data from a tertiary hospital in China.
• To enhance AI adaptability in clinical practice for Asian patients.

Main Methods:

• Utilized a dataset of 13,603 dermatologist-labeled dermoscopic images across 14 disease categories.
• Employed Google's EfficientNet-b4 architecture with auxiliary classifiers, retrained using PyTorch.
• Visualized model attention using saliency maps and analyzed learned features with t-SNE.

Main Results:

• The framework achieved high classification performance: 0.948 overall accuracy, 0.934 sensitivity, and 0.950 specificity.
• Outperformed existing Convolutional Neural Network (CNN) models with an Area Under the Curve (AUC) of 0.985.
• Demonstrated comparable diagnostic performance to 280 board-certified dermatologists in an 8-class task.

Conclusions:

• A deep learning framework trained on real-world clinical data can accurately classify common dermatoses.
• The model effectively handles infectious, inflammatory, benign, and malignant skin conditions encountered in outpatient settings.
• This AI framework shows potential for integration into clinical practice to aid dermatological diagnosis.