A Two-Stage Attention-Enhanced ConvNeXt Framework for Skin Cancer Detection and Classification

Abstract

The rising global incidence of skin cancer has positioned automated dermatological screening as a critical technology for modern healthcare diagnostics. This study introduces a two-stage deep learning framework for skin cancer detection and classification, demonstrating strong performance across diverse dermatoscopic imaging conditions. The proposed approach employs a ConvNeXt-Tiny backbone enhanced with Squeeze and Excitation (SE) and Convolutional Block Attention Module (CBAM) to improve feature recalibration and spatial channel attention. In Stage 1, a binary classifier distinguishes cancerous lesions from normal skin, achieving near-perfect validation accuracy of approximately 100% and an F1-Macro score consistently exceeding 95%, confirming robust generalization with minimal overfitting. In Stage 2, a multi-class classifier categorizes lesions into three clinically significant classes Melanoma (MEL), Basal Cell Carcinoma (BCC), and Vascular Lesions (VASC) achieving a validation accuracy of 94.62% with a weighted F1-score of 95.16%. On the held-out test set, Stage 2 attained an overall accuracy of 90.64%, a macro F1-score of 88.14%, and a weighted F1-score of 90.68%.The proposed two-stage framework was evaluated on a combined dataset sourced from Harvard Dataverse and Kaggle, achieving a testing accuracy of 100% on the Stage 1 binary screening task (cancer vs. normal skin) and a Stage 2 multi-class test accuracy of 90.64% with a macro F1-score of 88.14% across melanoma, basal cell carcinoma, and vascular lesions. These results highlight the framework's potential for reliable, automated skin cancer screening, providing a valuable tool for clinical decision support, early diagnosis, and accessible dermatological care.