Hybrid Vision Transformer-Boosting Model for Skin Lesion Classification on SLICE-3D

Early detection of cutaneous malignancies remains a critical challenge in healthcare. This study proposes an approach to skin lesion classification using the SLICE-3D dataset, containing over 400,000 smartphone-quality dermatological images from more than 1,000 patients. Several models, such as the Dataefficient Image Transformer 3 (DeIT3) and Light Gradient Boosting Machine (LGBM), are employed in the proposed methodology. An ensemble approach is also utilized to capitalize on their collective capabilities and strengths. To address class imbalance, a multi-phase preprocessing protocol involving stratified sampling, Synthetic Minority Over-sampling Technique (SMOTE), and data augmentation is implemented. Quantitative analysis indicates that the combined model exhibits improved performance, achieving an accuracy of 89% and a recall of 90% for class 1. This outperforms the individual models: DeIT3 achieved an accuracy of 73 % with a recall of 91% for class 1, while LGBM achieved a higher accuracy of 96% but with a lower recall of 71% for class 1. These findings underscore the potential of ensemble methodologies in enhancing the accuracy and reliability of dermatological diagnostics for cutaneous malignancy detection.