An innovative deep learning framework for skin cancer detection employing ConvNeXtV2 and focal self-attention mechanisms

• Published in: Results in engineering Vol. 25; p. 103692
• Main Authors: Ozdemir, Burhanettin, Pacal, Ishak
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2025; Elsevier
• Subjects: ConvNeXtv2; Focal self-attention; Medical image analysis; Skin cancer detection; Vision transformer; Vision transformer; Skin cancer detection; ConvNeXtv2; Focal self-attention; Medical image analysis
• ISSN: 2590-1230; 2590-1230
• DOI: 10.1016/j.rineng.2024.103692

•This study presents a hybrid model that integrates ConvNeXtV2 blocks and focal self-attention, enhancing feature extraction and detail recognition in skin cancer images.•By replacing traditional convolutional blocks with ConvNeXtV2 in the initial stages, the model effectively identifies subtle and complex skin anomalies.•The implementation of focal self-attention allows the model to prioritize diagnostically significant areas, improving accuracy and sensitivity in skin cancer classification.•The proposed model achieved 93.60 % accuracy, 91.69 % precision, 90.05 % recall, and a 90.73 % F1-score, surpassing numerous existing CNN and ViT models.•With only 36.44 million parameters, the model is optimized for real-time applications, addressing data imbalance and ensuring reliable diagnoses across all skin cancer classes. The skin, the body's largest organ, plays a critical role in protection and regulation, making its health essential. Skin cancer, one of the most prevalent malignancies, continues to rise globally and presents significant risks when diagnosis is delayed. Accurate detection is challenging due to the subtle and overlapping features of skin lesions, often leading to diagnostic errors. Deep learning has emerged as a powerful tool, capable of analyzing complex dermatological data and improving diagnostic accuracy through precise pattern recognition. This study proposes a novel lightweight and mobile-friendly hybrid model that combines ConvNeXtV2 blocks and focal self-attention mechanisms, addressing challenges such as data imbalance and model complexity. The Proposed Model employs ConvNeXtV2 in the first two stages for superior local feature extraction, while focal self-attention in the subsequent stages enhances sensitivity by focusing on diagnostically relevant regions. The Proposed Model was evaluated on the ISIC 2019 dataset, encompassing eight skin cancer classes with significant class imbalances, such as the Melanocytic Nevus class having 51 times more images than the Vascular Lesion class. Despite these disparities, the Proposed Model achieved robust performance across all classes, with 93.60% accuracy, 91.69% precision, 90.05% recall, and a 90.73% F1-score. Compared to baseline models and existing literature, it demonstrated a 10.8% improvement in accuracy over ResNet50 and a 3.3% improvement over the best-performing vision transformer (Swinv2-Base). This innovative design establishes a new benchmark in skin cancer detection, offering accurate, scalable, and generalizable predictions to support early diagnosis and improved clinical outcomes.