A Hybrid Ensembled Deep Learning Framework for Explainable Skin Cancer Classification and Lesion Detection

Automated methods for dermoscopy image analysis support the detection of skin cancer in an early stage and reducing skin cancer-related deaths.Deep learning based methods have shown compelling results for the analysis of skin cancer images.In this paper, an enhanced ECRNet-based hybrid model is explored for the classification and detection of skin cancer and diverse skin anomalies.This model utilizes an ensemble of classification models, namely, ResNet50, ResNet101, MobileNetV2, Vision Transformer, Con-vNeXt, DeiT-Small, EL-DLOA, WavIntNet, Conformer, Xception, VGG16 and an Ensemble of ECRNet and other models.For lesion localization, the YOLO model family and Faster R-CNN architecture are examined.Experimental results demonstrate that the Hybrid-Ensemble approach outperforms other models with an accuracy of 97.2%, precision of 95.4%, recall of 93.7%, and F1 score of 94.5% while YOLOV26 achieved an mAP of 71.3% with a precision of 73.9%.Grad-CAM was used to improve the model interpretability by highlighting the image regions containing the lesions.A web application for image upload, automated prediction, and diagnostic visualization was developed using Flask and SQLite.