Tri-Attention boosted scalable EfficientNet for skin lesion classification via Triple-Stage Gain Ratioed Averaging

Skin lesions appear in different abnormal forms, and skin diseases often disrupt both the structure and functionality of the skin. Early diagnosis and automated prediction are crucial, particularly since skin cancer is characterized by uncontrolled cellular proliferation. Despite progress in this area, a major research challenge lies in accurately locating the dominant regions of dispersion after lesion detection. To address this, we present an innovative framework based on Transfer Learning, incorporating multiple variants of Customized EfficientNet. Our approach integrates EfficientNet versions with three distinct attention mechanisms—Channel Attention, Soft Attention, and Squeeze-Excitation Attention—leading to the development of a scalable EfficientNet architecture. To further enhance performance, we introduce a specialized Triple-Stage Ensemble Learning strategy, termed Gain Ratioed Averaging (GRA), which effectively tackles the problem of optimal weight allocation for individual models within the ensemble. The central aim of GRA is to achieve efficient aggregation by determining the most suitable weights, and its extension, Triple-Stage GRA (TS-GRA), allows for broader application across three stages. Experimental results on the HAM10000 dataset validate our method, achieving a remarkable accuracy of 94.57% with TS-GRA, surpassing state-of-the-art approaches. Moreover, to improve interpretability and reveal the regions influencing the model’s decisions, we employ Gradient-weighted Class Activation Mapping (Grad-CAM). This visualization technique strengthens explainability by clearly highlighting the critical areas relevant to diagnosis.