SegNet3+: An Attention‐Guided ResNet‐U‐Net 3Plus Framework for Accurate and Explainable Skin Lesion Segmentation in Dermoscopic Images

ABSTRACT Skin cancer is one of the most prevalent and rapidly increasing malignancies in the world and the timely diagnosis of the disease plays a crucial role in improving patient outcomes. Lesion segmentation in dermoscopic imaging is not a simple procedure due to the variability in lesion size, shape and colour, low‐contrast edges and imaging artefacts. It needs to be segmented for diagnosis and treatment. The new deep learning model presented in this paper, SegNet3+, aims to improve skin lesion segmentation. Its primary aim is to enhance the precision of the segmentation process by addressing problems of localization and boundary demarcation, as well as issues of local and global contextual information. SegNet3+ employs state‐of‐the‐art deep learning algorithms, including U‐Net 3Plus, convolutional block attention module (CBAM) and ResNet, to facilitate skin lesion segmentation. For feature extraction, the model uses a ResNet to capture intricate, discriminative features from dermoscopic images effectively. The U‐Net 3Plus framework takes segmentation a step further by using multi‐scale decoding to provide contextual information at different scales and by enhancing boundary detection and lesion representation. There is also the CBAM, which can be employed to reduce the model's bias toward non‐relevant regions of the image by attending to clinically meaningful regions with channel and spatial attention. The binary cross‐entropy (BCE) and focal loss are combined to train the model to address class imbalance and improve segmentation precision. This is achieved through extensive data augmentation to improve generalisation, minimise overfitting and reduce the model's sensitivity to variations in dermoscopic images. The popular international skin imaging collaboration (ISIC) 2018 benchmark dataset, to which SegNet3+ has been trained, is for skin lesion segmentation. The model's lesion detection rate was quite high (96.13). The intersection over union (IoU) of 83.78 implies that the rate of agreement between the segmentation that is predicted and the ground truth is high. Also, the Dice coefficient and F1‐score of 91.17 indicate that the model is effective at controlling false positives (FP) and false negatives (FN). The model achieved 89.42% sensitivity and 98.06% specificity, indicating it is sensitive to lesions while minimizing FP. These results show that SegNet3+ outperforms other state‐of‐the‐art segmentation models, indicating that it is a strong and reliable model for diagnosing skin cancer. SegNet3+ is a highly accurate and interpretable skin lesion segmentation network. Ablation studies can also be used to improve segmentation quality by combining CBAM and ResNet encoders. Moreover, the gradient‐weighted class activation mapping (Grad‐CAM) explainability analysis indicates that SegNet3+ focuses on clinically relevant areas of the lesion, thereby enhancing the model's interpretability and reliability. The above characteristics make SegNet3+ a useful tool for computer‐aided dermatological diagnosis.