Adaptive Weighted Ensemble of CNN Models with Focal Loss and Grad-CAM Explainability for Skin Lesion Classification

Abstract Skin cancer is one of the oldest and most widespread forms of cancer globally, and discovering it early is crucial for improving patient outcomes. Traditional diagnosis heavily depends on experience, which may be time-consuming and human-scheduled. Recent developments in algorithms for deep learning, particularly convolutional neural networks (CNNs), provide promising platforms for independent image analysis, but individual models are prone to fail with unstable predictions and weak generalizability. In this research, we have worked with a weighted ensemble transfer learning framework that integrates ResNet34, ResNet50 and EfficientNet-B0 to enhance the utilization of the HAM10000 dataset to sort skin lesions. Each model is individually finetuned. And Ensemble weights are assigned according to validation performance. That ensures a stronger network and a stronger influence on the final prediction at the end. Additionally, Grad-CAM visualisation helps to show where lesions are located. It also confirms that the ensemble focuses on the area that matters clinically. Our tests show that the weighted ensemble method outperforms single CNNs in accuracy and F1 score. It also gives a better understanding and is more dependable. These results show that pairing adaptive weighting with visual details produces a more reliable and clear way of automatically detecting skin cancer for clinics.