LightSkinNet: Protocol-Aware Skin Lesion Classification with Segmentation-Guided Networks

Automated dermoscopic analysis is commonly approached as two separate tasks: lesion segmentation and skin lesion classification. However, the wide range of reported accuracies on the HAM10000 dataset is difficult to interpret because different studies often use different evaluation protocols. In this work, we introduce LightSkinNet, a two-stage framework in which a lightweight segmentation model first localizes the lesion, and the resulting cropped image is then passed to a classifier. We investigate two encoder designs consistently across segmentation and classification: a convolutional encoder incorporating squeeze-and-excitation and bottleneck attention mechanisms, and a hybrid architecture with a transformer-based bottleneck. On ISIC 2018 Task 1, the convolutional model achieves a Dice coefficient of 0.8927, while on HAM10000, our best configuration achieves 0.9271 accuracy under an image-level split, with the convolutional design again outperforming the hybrid model (p < 10^-12, McNemar's test). We further examine the evaluation protocol and find that the 10,015 HAM10000 images correspond to 7,470 distinct lesions, with 661 lesions crossing the training-test boundary under the commonly used image-level split. We also show that melanoma detection is strongly influenced by the decision threshold rather than solely by the learned representation: validation-only threshold tuning reduces missed melanomas from 49 to 22 without modifying the model parameters. To assess whether the classifier focuses on clinically relevant regions, we quantify the overlap between Grad-CAM activations and the predicted lesion masks. Finally, we report both the strengths and limitations of the proposed framework, including the finding that a low parameter count does not necessarily translate into low computational cost.