Human Skin Fungal Diseases Classification Using Deep Learning Technique

Abstract

The skin is our body's outermost layer. It has a lot of different functions. Since it is the outer covering of the whole body part that acts as a barrier, protecting our body from harmful things in the outside world such as sun rays, germs, and toxic substances. Skin plays a great role in body temperature regulation. Several risks affect the skin from those the common cause of skin disorder are bacteria, viruses, fungus. It takes a long time to identify a disease based on manual feature extractions or symptoms, and it requires a lot of expertise to do so correctly. There have been previous studies on the diagnosis, detection, and classification of skin diseases. However, in the previous work tinea pedies, and tinea corpories are not identified especially for black skin color. In this thesis, we develop a model that uses CNN to classify skin fungal diseases such as tinea pedies, tinea capitis, tinea corpories, and tinea unguium. The images are then classified as tinea pedies, tinea capitis, tinea corpories, or tinea unguium by Softmax. We have collected 407 skin fungal lesion images from patients at Dr. Gerbi medium clinic of Jimma, and JUMC using the smartphone camera (Techno pop 2 power, Techno Spark4, SamsungA20). After collecting datasets, Image Preprocessing, Image augmentation techniques are applied to increase the performance of the human skin disease classification model. In this study, we have done image preprocessing (image size normalization, RGB to Grayscale conversion). We have normalized the images to three image sizes which are 120 x120, 150X150, and 224x224. From the total augmented 1069 images, 80% (727) for training, 10% (164) validation, and the remaining 10% (178) for testing. We have registered the overall performance accuracy of 83% using our CNN-based HSFDCModel after the model is evaluated. The accuracy achieved 79%, 69% for MobileNetV2 and ResNet50 respectively. This implies the developed model is better than the MobileNetV2 and ResNet50 pre-trained CNN Models for our dataset