JILDYA-Net: An Efficient Lightweight Multi-Class Classification Architecture for Skin Lesions

Skin lesion classification has become increasingly important yet challenging due to the time physicians spend manually analyzing very similar lesions. While traditional deep learning methods have historically offered dependable automated support in lesion detection, thus improving patient care, newer lightweight architectures bring distinct advantages like decreased computational requirements and quicker training, making them better suited for mobile devices, microcontrollers, and embedded systems. The present paper proposes JILDYA-Net: A lightweight method designed for mobile applications and embedded systems, enabling accurate, rapid, and consistent diagnosis of skin lesions from dermoscopic images. The proposed approach aims to improve the analysis of dermoscopic images containing diverse features through two main components. First, a novel convolutional attention component called attention-based structural feature enhancement is introduced to enhance skin lesion features. Then, an encoder-based FNet enables faster processing and lower memory usage, which is especially beneficial for longer input lengths via Fourier Transforms. Additionally, an external attention module refines learned representations and emphasizes relevant features, accelerating convergence and improving model performance and stability during training. Furthermore, augmentation techniques are employed to address class imbalance sensitivity, generating additional data and reducing overfitting. Overall, the goal is to achieve optimal performance with a simple model that trains quickly. Regarding the evaluation metrics, we employ accuracy, sensitivity, specificity, and AUC. Our approach displays a competitive performance, validated through experiments on augmented and balanced versions of the HAM10000 and ISIC-2019 datasets, as compared with state-of-the-art methods. It demonstrates superior performance in accuracy, sensitivity, and specificity relative to competing methods.