Meta-Learning-Based Classification of Melanoma Types in Dermoscopy Images

Melanoma is the most dangerous skin cancer that affects a large number of patients every year; however, if melanoma is diagnosed at the initial stages the probability of treatment and survival of the patient increases to a great extent. Since there is no access to a dermatologist in all places, and on the other hand, even less experienced physicians face problems in diagnosing melanoma, the existence of computer-aided diagnosis systems that can accurately diagnose skin cancers, especially melanoma is necessary. Despite extensive research in this field, most of which has been conducted on the binary classification of skin lesions into melanoma and non-melanoma, very little researches have focused on the classification of melanoma types due to the scarcity of labeled datasets for different kinds of melanoma. Deep learning algorithms require large datasets for training, making them ineffective for diagnosing rare diseases or their subtypes due to insufficient data. Therefore, the purpose of this research is to utilize new algorithms of artificial intelligence, such as meta-learning that is a sub-field of deep learning. This method is capable of classifying images even with a limited number of images, and this method is used to classify the types of melanoma. Therefore, in this research, various meta-learning algorithms, such as Prototypical Networks, Matching Networks, Relation Networks, and MAML, have been used to classify melanoma types, which are suitable for classifying images even with a limited number. Given that the meta-learning paradigm comprises two phases—meta-training and meta-testing—we utilized distinct datasets for each phase. The HAM10000 dataset was used for meta-training, while the melanoma-types dataset was employed for meta-testing. We also employed various data augmentation techniques, such as random cropping, color jittering, and random horizontal flipping, to enhance the model’s performance and achieve robust results. Out of all the algorithms, the Prototypical Networks achieved the highest accuracy, reaching approximately $74.5 \%$ using only ten images. The innovation of this research lies in the classification of melanoma types using meta-learning algorithms, a technique that, to our knowledge, has not been previously employed in any other research.