A Hybrid Attention-Guided Fusion Network with Grad-CAM for MPox Skin Lesion Classification

The rapid identification of MPox in skin lesion images is crucial to restrain its spread. Current diagnostic systems often struggle to distinguish between diseases with similar visual features. Disease like chickenpox, measles, cowpox, HFMD, and healthy skin are quite similar. These increases the likelihood of misdiagnosis. This research gap needs more effective methods to classify MPox among other similar diseases accurately. In response, we propose an Attention-Guided Fusion Network (AGFN) that combines a Convolutional Neural Network (CNN) backbone for feature extraction with a multi-head attention block, enhancing representation through contextual fusion. Our model incorporates ResNet50 for strong feature extraction alongside a multi-head attention mechanism that emphasizes critical patterns, facilitating accurate classification among six categories, including MPox, chickenpox, measles, cowpox, HFMD, and healthy skin. Additionally, the interpretability of our model is significantly enhanced by the integration of Grad-CAM (Gradient-weighted Class Activation Mapping), which generates class-specific activation maps to visually highlight the key regions in skin lesion images driving the model's decisions. Our proposed model, trained and tested on a dataset comprising 755 images across these categories, achieved an accuracy of 99.81%, precision of 99.92%, recall of 99.67%, and F1 score of 99.79% on the test dataset, underscoring its high diagnostic reliability. The proposed AGFN model demonstrates the potential for real-world application in clinical settings, where accurate differentiation of skin lesions can enhance disease control and improve patient outcomes. Our approach promises to fill a significant research gap by providing a highly accurate, adaptable detection tool for MPox and similar conditions, supporting faster and more precise diagnoses in healthcare systems worldwide.