An EfficientNet-based mobile model for classifying eczema and acne

Abstract

Eczema is one of the diseases that attack the skin. A diagnosis of eczema is generally made through a medical check-up by a dermatologist. It is quite difficult for a patient to differentiate eczema and non-eczema disease that have similar symptoms. Recent computer vision studies have been conducted to help patients classify eczema and other skin diseases using photo images. In this study, we designed a deep learning-based mobile model that classifies eczema and acne, since both eczema and acne have visually similar redness symptoms. To develop the model, we take EfficientNet architecture families to select the best model. The result says that EfficientNet family architectures give an overall good result in terms of accuracy. The highest accuracy in both training and test that we get are 92.71% and 95.42%. Since we consider the inference time, we choose the model that has pretty high accuracy but has pretty low latency on inference. The chosen model has been translated into a mobile optimized model through TFLite and has been implemented in an Android mobile application. The inference time of the classification is about 0.04 seconds.