Automated Psoriasis Detection using Deep Learning

Abstract

Psoriasis is a chronic, noncontagious skin condition that cannot be cured but its early detection can help prevent serious life-threatening complications. The high visual similarity between normal skin and psoriasis has made the detection of psoriasis a very complex task. Moreover, it can be confused with different skin abnormalities like eczema, tinea corporis, lichen planus, pityriasis, dandruff, seborrheic dermatitis. Image processing using deep learning has proven better than other approaches in this context because of its automatic feature extractions with intelligent decisions and less chances of distorted features. In this paper, automated detection of psoriasis using deep learning has been proposed. To obtain good results for a small dataset transfer learning mechanism is used in which pre-trained deep learning models are applied on a dataset to obtain the required results. Firstly, different transfer learning models are applied on our data to work on the best-obtained accuracy. Among them, ResNeXt gave the best output for an appropriate accuracy to detect psoriasis from healthy skin as well as other skin diseases. Secondly, we have worked on facilitating the development of an automated system that classifies psoriasis, lichen planus, eczema, seborrheic dermatitis, pityriasis, normal skin, and tinea corporis diseases by applying and improving the final layers of the pre-trained model. We have obtained an accuracy of 94% on test images with 2 classifiers and output to show if the input image is classified as psoriasis or not. Finally, we have also applied the classifier on 3 classes; normal skin, psoriasis, and other skin diseases, and obtained good results