Multiscale Feature Fusion using Attention Mechanisms in Multiclass Skin Lesion Classification: A Comprehensive Study

Skin lesions represent a diverse spectrum of skin abnormalities, while skin diseases encompass conditions affecting the skin's structure and function.Among these, skin cancer emerges as a critical concern, driven by uncontrolled abnormal cell growth.Early detection of lesions is pivotal, yet manual approaches are prone to human error and misdiagnosis.This study introduces a convolutional neural network (CNN)-based solution, the Sequential Triple Atten-tionNet (STA), to automate lesion detection while ensuring precise identification of affected regions within dermoscopic images and addressing challenges related to dominant image classes.The proposed STA architecture integrates features from pre-trained CNN models (e.g., ResNet-50, ResNet-101, DenseNet-169, and DenseNet-201) with three attention mechanisms, emphasizing critical features for accurate classification.The architecture, augmented by a singlelayer ensemble (SLE), demonstrates enhanced performance and adaptability.Empirical evaluation on the HAM10000 and ISIC-2018 datasets yield remarkable accuracies of 98.78% and 98.40%, respectively, outperforming existing methods.These findings highlight the potential of the proposed approach for advancing skin lesion classification and enhancing proactive measures for the prevention of skin cancer. CCS Concepts Computing methodologies Object recognition; Interest point and salient region detections.