인공지능을 통한 피부암 진단

초록

This study aims to verify the performance of an AI-based Convolutional Neural Network (CNN)-based EfficientNetB0 model in binary classification of skin cancer into malignant and benign categories using the HAM10000 skin lesion image dataset, and to exploratorily compare its diagnostic accuracy with the visual reading results of four plastic surgeons. During data preprocessing, image integrity verification was performed using the PIL library, removing one corrupted image, and a final dataset of 10,014 images was used. To address class imbalance, 50:50 balanced sampling was applied only to the training data, and the model was trained using transfer learning with ImageNet pre-trained weights and full-layer fine-tuning with 10-Fold cross-validation. The experimental results show that the EfficientNetB0 model achieved a mean accuracy of 77.95% (maximum 82.48%, minimum 74.09%, standard deviation 2.36%), with sensitivity 54.68%, specificity 82.46%, F1-Score 44.59%, and AUC-ROC 0.7777 on the full HAM10000 dataset of 10,014 images via 10-Fold cross-validation. Furthermore, to ensure comparability, the AI model was additionally evaluated on the same 40 images reviewed by four plastic surgeons using 10-Fold ensemble inference, achieving an accuracy of 82.50% and sensitivity of 95.00%, which is 19.40%p higher than the mean accuracy of four plastic surgeons (63.10%). This study exploratorily demonstrates the potential of AI-based skin cancer screening systems to be utilized as auxiliary tools in general health checkup programs in non-specialist medical environments.

키워드