Attention Enhanced Inception-V3: A Multi-Scale Feature Fusion Network for Skin Lesion Detection with Explainable Artificial Intelligence

Skin lesions, although often benign, can pose significant risks if they develop into cancer, underscoring the critical importance of early and accurate detection for effective treatment and cancer prevention. However, traditional research approaches utilizing transfer learning (TL) techniques often fall short due to inadequate fine-tuning, posing challenges in extracting both shallow and deep features. To address this gap, this study introduces a novel approach by incorporating a feature fusion of diverse attention mechanisms, including Soft Attention (SA), Channel Attention (CA), and Squeeze and Excitation, into a finely tuned Inception-V3 architecture. This integration results in the development of Attention Enhanced Inception-V3 (AEIV3), which is capable of dynamically focusing on relevant features to enhance discriminative feature extraction, consequently leading to improved classification accuracy. Leveraging the PAD UFES 20 dataset comprising patient data and smartphone-captured clinical images, rigorous pre-processing techniques are employed to ensure balanced class distributions, thus facilitating robust model evaluation. Experimental results showcase the superior performance of AEIV3 compared to existing algorithms, yielding a remarkable accuracy score of 99.58%. Moreover, to provide insights into the decision-making process of the proposed model, Gradient Class Activation Map (GradCAM) visualization is utilized, highlighting the regions responsible for decision-making. Overall, we believe that our work will significantly contribute to the early detection of skin lesions, ultimately aiding in the advancement of research in this critical domain.