An Efficient Attention-Guided Xception Framework for Robust Multiclass Skin Lesion Classification

Abstract

Skin cancer is one of the most common and potentially deadly diseases worldwide. Early and accurate diagnosis of skin cancer is very important for better outcomes for the patient and a reduction in mortality rates. However, the visual similarity of different skin lesion categories and the class imbalance are major challenges for automated skin lesion classification. This study proposes an Attention-Guided Xception framework for robust multiclass skin lesion classification. The proposed framework leverages transfer learning based on the Xception architecture for discriminative dermoscopic feature extraction, and an attention mechanism to enhance the representation of clinically relevant lesion regions and suppress irrelevant background information. The model was trained on the ISIC skin lesion dataset, with melanoma samples removed during dataset refinement to create an eight-class non-melanoma classification framework. Data preprocessing and class balancing techniques were used in order to improve the stability of training and the classification performance. Accuracy, precision, recall, F1-score and confusion matrix analysis were used to evaluate the proposed framework. The experimental results show that the proposed Attention-Guided Xception model achieved a classification accuracy of 93 % and outperformed CNN, VGG16, EfficientNet-B2 and Xception architectures. The results demonstrate the benefit of attention-guided feature learning in improving lesion discrimination and reducing inter-class ambiguity. The proposed framework provides an accurate and computationally efficient way to automate skin lesion classification, and it demonstrates its capability to support computer-aided dermatological diagnosis and early disease detection.