Abstract
Early diagnosis is still the best method to face skin cancer. The diagnosis of skin lesions remains as a challenge for physicians and researchers. In the past few years, it has benefited from computer-aided diagnosis methods that successfully apply classic Machine Learning techniques and more recently Convolutional Neural Networks. This work is aimed at discovering architectures that best fuse clinical records and medical images for the diagnosis of skin lesions. As a result, a genetic algorithm is designed in order to select how to combine such information and the main details of the new architecture. The architecture is able to cope with multiple inputs and learn multiple outputs, proving flexibility by sharing network parameters, which implicitly mitigates the overfitting of the model. An extensive experimental study was conducted on the well-known ISIC2019 dataset, where the models were trained with a total of 72,106 images and meta-data, including the augmented images. The proposal outperformed the baseline state-of-the-art model while diagnosing from eight skin lesion categories. Furthermore, the discovered architecture achieved 85%, 94%, and 84% of recall score when diagnosing malignant lesions - melanoma, basal cell carcinoma, and squamous cell carcinoma, respectively. Finally, the results showed the suitability of the proposed genetic algorithm, which was able to automatically build a multimodal fusion architecture for the diagnosis of skin lesions.