QuadResBottleNeck: A lightweight bottleneck-based residual CNN for efficient multi-class skin lesion classification

The classification of skin lesions from dermoscopic images remains challenging due to significant intra-class variation, high inter-class similarity, class imbalance, and a limited number of annotated medical images. To address these challenges, this study introduces QuadResBottleNeck, a lightweight residual convolutional neural network that embeds bottleneck compression within a controlled four-stage residual network, maintaining discriminative feature learning while reducing the number of parameters and computational costs. The proposed model was tested on the ISIC 2019 dataset using patient-level stratified five-fold cross-validation and compared with a Vision Transformer (ViT) baseline and other CNN architectures, using the same data preprocessing and evaluation procedures. On ISIC 2019, QuadResBottleNeck achieved 98.2% end-to-end accuracy, 98.1% macro F1-score, and an AUC of 0.994, with significantly fewer parameters and lower computational cost than the ViT. Cross-dataset evaluation on D2 also showed strong generalization, achieving 96.4% accuracy and a macro F1-score of 95.9%. The computational analysis revealed a favorable balance between classification performance and resource usage, supporting the proposed architecture's application in scenarios with limited resources and edge assistance in clinical settings. The results show that competitive multi-class skin lesion classification can be achieved with well-designed lightweight residual architectures without the need for large-scale model architectures.