Classification of the Clinical Images for Benign and Malignant Cutaneous Tumors Using a Deep Learning Algorithm

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate classification of skin diseases is crucial for timely treatment.
• Deep learning models show promise in analyzing medical images.
• Existing algorithms require validation across diverse datasets.

Purpose of the Study:

• To evaluate a deep learning algorithm's efficacy in classifying 12 common skin diseases.
• To compare the algorithm's diagnostic performance against dermatologists.
• To identify areas for improvement in the convolutional neural network model.

Main Methods:

• A convolutional neural network (Microsoft ResNet-152) was trained on 19,398 skin lesion images from multiple datasets.
• The model was fine-tuned using training data and validated on separate testing sets (Asan, Hallym, Edinburgh).
• Performance was assessed using area under the curve (AUC) and sensitivity metrics.

Main Results:

• The algorithm achieved high AUC values for diagnosing basal cell carcinoma (0.96) and melanoma (0.96) in the Asan dataset.
• Comparable AUCs were observed in the Edinburgh dataset, with strong performance for squamous cell carcinoma (0.91).
• Sensitivity for basal cell carcinoma reached 87.1% in the Hallym dataset, with overall performance rivaling dermatologists.

Conclusions:

• Deep learning algorithms can effectively classify multiple skin diseases from clinical images.
• The ResNet-152 model demonstrates significant potential in dermatological diagnostics.
• Expanding training datasets with diverse demographics is essential for future performance enhancement.