Development of an open source application for melanoma image detection

Details

Supervisors

Jodogne, Sébastien

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enSkin melanoma is one of the most aggressive forms of skin cancer, with early detection being critical to patient prognosis. Despite the growing availability of AI-based diagnostic tools, the majority of high-performance dermatological systems remain proprietary, lacking accessible open-source alternatives that integrate both a specialized deep learning model and a ready-to-use interface. This thesis presents the full development of an open-source melanoma detection system, from model design to clinical web application. The proposed deep learning model follows a multimodal feature-level fusion architecture, combining a visual feature extractor based on EfficientNet-B6 with a Multi-Layer Perceptron that encodes clinical metadata. The model is trained on the ISIC 2020 dataset and addresses the inherent class imbalance through the use of Focal Loss. Data augmentation and Stratified K-Fold cross-validation are employed to improve generalization and robustness. The trained model achieves a mean AUC of 0.910 across cross-validation folds, with a recall of 0.807, meeting the clinical imperative to minimize false negatives. Regarding the deployment of the model, the PyTorch network is exported to ONNX format with preprocessing logic encapsulated directly within the computation graph. A mobile-oriented web application enables fully client-side inference via the ONNX Runtime Web library and WebAssembly, guaranteeing that no patient data is transmitted over the network. Experimental evaluation confirms that browser-based inference remains sufficiently responsive for real-world use. The resulting system is therefore practically viable for end-users on their mobile devices. This work delivers a complete, open-source and ready-to-use melanoma detection system.