A New Approach using Deep Learning and Reinforcement Learning in HealthCare: Skin Cancer Classification


  • Dahdouh Yousra Faculty of Sciences and Techniques, LIST Laboratory FSTT UAE Tangier, Morocco https://orcid.org/0000-0001-7147-6637
  • Anouar Boudhir Abdelhakim Faculty of Sciences and Techniques, LIST Laboratory FSTT UAE Tangier, Morocco
  • Ben Ahmed Mohamed Faculty of Sciences and Techniques, LIST Laboratory FSTT UAE Tangier, Morocco




Deep Learning, CNN, Reinforcement Learning, Classification, Skin Cancer, Deep Q_Learning, Dermoscopy Image, Segmentation


Nowadays, skin cancer is one of the most important problems faced by the world, due especially to the rapid development of skin cells and excessive exposure to UV rays. Therefore, early detection at an early stage employing advanced automated systems based on AI algorithms plays a major job in order to effectively identifying and detecting the disease, reducing patient health and financial burdens, and stopping its spread in the skin. In this context, several early skin cancer detection approaches and models have been presented throughout the last few decades to improve the rate of skin cancer detection using dermoscopic images. This work proposed a model that can help dermatologists to know and detect skin cancer in just a few seconds. This model combined the merits of two major artificial intelligence algorithms: Deep Learning and Reinforcement Learning following the great success we achieved in the classification and recognition of images and especially in the medical sector. This research included four main steps. Firstly, the pre-processing techniques were applied to improve the accuracy, quality, and consistency of a dataset. The input dermoscopic images were obtained from the HAM10000 database. Then, the watershed algorithm was used for the segmentation process performed to extract the affected area. After that, the deep convolutional neural network (CNN) was utilized to classify the skin cancer into seven types: actinic keratosis, basal cell carcinoma, benign keratosis, dermatofibroma melanocytic nevi, melanoma vascular skin lesions. Finally, in regards to the reinforcement learning part, the Deep Q_Learning algorithm was utilized to train and retrain our model until we found the best result. The accuracy metric was utilized to evaluate the efficacy and performance of the proposed method, which achieved a high accuracy of 80%. Furthermore, the experimental results demonstrate how reinforcement learning can be effectively combined with deep learning for skin cancer classification tasks.






Original Scientific Papers