CNN-Based AI Versus Physicians for Solitary Skin Lesion Diagnosis

This study is a retrospective, observational diagnostic accuracy study designed to evaluate the performance of a convolutional neural network (CNN)-based artificial intelligence model in the assessment of solitary skin lesions using macroscopic clinical images.

Macroscopic clinical images of solitary skin lesions with histopathological or clinically confirmed diagnoses will be retrospectively retrieved from the dermatology image archive of Istanbul Training and Research Hospital. All images and associated clinical documents will be anonymized prior to analysis, and any identifying visual or textual information will be removed. Data processing and analysis will be conducted in a secure, institution-based environment with restricted access limited to the study team.

A CNN-based artificial intelligence model will be developed using supervised learning techniques. Image preprocessing steps will include resizing to standardized input dimensions, color normalization, and removal of regions containing potentially identifiable information. The dataset will be partitioned into training, validation, and test subsets to enable model development, hyperparameter optimization, and independent performance evaluation. Model training and evaluation will be implemented using the PyTorch deep learning framework.

The diagnostic performance of the CNN-based model will be evaluated using standard classification metrics and will be compared with the independent assessments of dermatologists, dermatology residents, and non-dermatologist physicians who evaluate the same set of anonymized images without access to additional clinical or histopathological information. Comparative analyses will be performed to assess differences in diagnostic performance and agreement between the artificial intelligence model and physician groups.

Inclusion criteria

• Patients who have provided informed consent for the use of their clinical images in scientific research.
• Clinical images with a resolution exceeding 224x224 pixels, ensuring compatibility with the artificial intelligence architecture.
• Retrospective records of solitary skin lesions with confirmed diagnoses.

Exclusion criteria

• Patients who have not consented to the use of their clinical photographs for research purposes.
• Images containing potentially identifiable personal information or visual features that compromise patient anonymity.
• Images with a resolution lower than 224x224 pixels or poor diagnostic quality (e.g., blurring, significant occlusion).
• Duplicate images or entries for the same lesion.