<![CDATA[The increasing global incidence of skin cancer, particularly melanoma, coupled with a scarcity of dermatologists, necessitates the development of accurate and accessible AI-driven diagnostic tools. However, deep learning models often struggle with severe class imbalance in public dermoscopic datasets, leading to poor performance on minority lesion types. This research aims to enhance the diagnostic precision of a lightweight MobileNetV2 architecture for multi-class skin lesion classification by integrating the Convolutional Block Attention Module (CBAM) and employing Focal Loss. The methodology involves evaluating four model variants (Baseline, Baseline+CBAM, Baseline+Focal Loss, and Baseline+CBAM+Focal Loss) on the HAM10000 dataset, with performance measured by accuracy, precision, recall, and F1-score. The optimal model (M4) successfully achieved convergence without overfitting, demonstrating exceptional F1-scores for six of seven classes, including near-perfect classification for melanoma (0.96) and dermatofibroma (0.97). The primary limitation was the actinic keratosis class (F1-score 0.60) due to high morphological similarity with other lesions. In conclusion, the synergistic combination of CBAM and Focal Loss effectively mitigates class imbalance and enhances feature representation in a computationally efficient model, providing a robust and interpretable solution for skin cancer screening.]]>
Copyrights © 2026
