Specialized Diagnosis of Cutaneous Squamous Cell Carcinoma from Similar Lesions Using Artificial Intelligence

This study investigates the most significant contribution of lesion detection towards enhancing artificial intelligence-based classification of dermoscopic images of cutaneous squamous cell carcinoma (SCC). The research differentiates SCC from lesions with comparable appearance such as actinic keratosis (AK) and basal cell carcinoma (BCC). Two architectures in deep learning were contrasted using ۵-fold cross-validation: an EfficientNet-B۱ model trained on raw, unprocessed dermoscopic images and one utilizing pre-processed images through lesion detection and cropping. The unprocessed-image model had ۹۳% accuracy, ۹۱% sensitivity, and ۹۵% specificity. On the other hand, the cropped-image model reached ۹۲% accuracy, ۹۰% sensitivity, and ۹۵% specificity. Although with similar overall accuracy, the cropped-image approach showed improved stability in detecting non-SCC lesions through lessening background noise. EfficientNet-B۱ performed better than ResNet۵۰ and VGG۱۶ because of its computational efficiency and scalability, which lends itself to medical imaging. These findings validate the importance of image preprocessing and propose combining lesion detection with classification to further improve diagnostic accuracy for SCC in clinical care.

کلیدواژه ها:

Squamous cell carcinoma ، Artificial Intelligence ، Lesion detection ، Dermoscopy ، Image classification ، EfficientNet ، Skin cancer ، Preprocessing