Efficient deep learning for skin lesion segmentation using convolutional block attention mechanism

Abstract

Skin cancer, particularly melanoma, poses a significant worldwide medical challenge as a result of growing incidence rates and limited awareness until advanced stages. This issue is exacerbated in developing countries, where a shortage of dermatologists hinders timely diagnosis. The subtle color variations in affected skin regions, which often resemble normal skin tones, underscore the critical need for early detection systems. To address this challenge, this paper proposes a hybrid deep learning segmentation model for automated skin lesion segmentation, aimed at improving early diagnosis of skin cancer. The proposed model leverages the HAM10000 dataset and integrates EfficientNetB7 as the encoder, UNet++ as the decoder, and Convolutional Block Attention Modules (CBAM) within skip connections. By incorporating CBAM, the model achieves a Dice coefficient of 0.9468 and an Intersection over Union (IoU) score of 0.9001, reflecting a 1–2% performance improvement compared to the model without CBAM. This noise-robust augmentation strategy enhances model generalization, ensuring robust performance on both noisy and normal dermatoscopic images. To further validate its effectiveness, the model was trained on additional datasets, including ISIC2016, ISIC2017, PH2, and MSK, and evaluated on their respective validation and test splits. The model demonstrated superior generalization and robustness, achieving a Dice coefficient of 0.9301 on the ISIC2016 dataset. The proposed model also achieved superior performance on the other datasets as well. These lightweight models establish a strong foundation for automated skin cancer diagnosis. Future work will focus on refining preprocessing techniques and incorporating multimodal data to further enhance model performance, with the ultimate goal of revolutionizing dermatological care and improving patient outcomes.