DRSAL: Deep Reinforcement Skin Cancer Diagnosis with Active Learning Technique

Malignancy is the most dangerous disease that causes higher fatality rate among humans. Albeit, different malignancy found in recent times, skin cancer is considered as highly contagious and affects human life in a slower pace. Current development of skin imaging modality such as Dermoscopy, clinicians can able to diagnose the disease quickly. With less availability of experts around the world, automating the disease diagnosis helps to start the treatment earlier. Advancement of deep learning techniques helps this automation procedure for disease diagnosis. Main problem behind this algorithm to get good accuracy, is that it needs more amount of labelled input data. Labelling huge input data is a time-consuming process and the chance for getting experts in medical field to label the input data is very less. We propose a novel algorithm which can train themselves with less data available and predicts and annotate the remaining dataset available with less intervention of human experts. This is possible with the improvement of deep reinforcement learning algorithm which smartly analyses the given data and diagnose the disease by providing a smaller number of datasets with active learning technique. Multi-Agent in the reinforcement algorithm helps to learn the important features from the given class and based on the similarity it aligns the input data of individual class accordingly. Each agent shares the gained information to other agent in order to improvise the learning method in a quicker way and start labelling or predicting the remaining unknown dataset. The proposed method Deep Reinforcement Skin Cancer Diagnosis with Active Learning Technique (DRSAL) achieves higher accuracy of 94.2%in diagnosing the disease compared with other popular deep learning and active learning algorithms.