Skin cancer detection using late fusion of pretrained models

Skin cancer is one of the most common and fatal malignancies worldwide, and reliable, early diagnosis systems are required to provide effective clinical intervention. Nevertheless, even with recent progress in deep learning, current models tend to overfit and lack generalizability and robustness across various lesion types. This paper overcomes these challenges by presenting a late-fusion ensemble of pre-trained convolutional neural networks (CNNs) integrated with Diverse Convolution Networks (DCNs) to build a complementary and discriminative skin cancer detection framework from dermoscopic images. The study utilized a curated dataset of 10,600 dermoscopic images from the ISIC Archive, comprising 9600 images for training and 1000 images for testing, encompassing both malignant and benign lesion classes. The proposed late-fusion model combines the predictions of several pre-trained architectures, leveraging their strengths while mitigating bias. Experimental results show excellent performance with an accuracy of 99.7%, an F1-score of 99.75%, a precision of 99.12%, and a recall of 99.34% on the test set, and an accuracy of 99.1% and an F1-score exceeding state-of-the-art models of 99.8% on the melanoma dataset. The conclusions validate the robustness and diagnostic accuracy of the framework in complex clinical situations. Future work will focus on model explainability, integration with clinical decision support systems, and validation using larger, multi-source datasets to enhance real-world applicability.