International Journal on Advanced Science, Engineering and Information Technology, Vol. 13 (2023) No. 4, pages: 1472-1479, DOI:10.18517/ijaseit.13.4.18290

Mobile Skin Disease Classification using MobileNetV2 and NASNetMobile

Idhawati Hestiningsih, Afandi Nur Aziz Thohari, - Kurnianingsih, Nur Diyana Kamarudin


The people of Indonesia often suffer from three diseases: tinea versicolor, ringworm, and scabies. Most Indonesians cannot distinguish the type of skin disease they suffer from because some have the same characteristics and patterns. Therefore, this study built an M-Health application to predict skin diseases. The prediction process uses a deep learning model deployed in the smartphone. The challenge of this study is the limited number of datasets because the data used is personal and requires permission from the patient or hospital. Therefore, transfer learning is used to overcome these data limitations. The transfer learning process in this study uses two pre-trained models, MobileNetV2 and NASNetMobile. To obtain a model with high accuracy, performed modifications to the architecture MobileNetV2 and NASNetMobile. The test results showed that MobileNetV2 performs best using a learning rate of 0.0005 and activation function ELU. While NASNetMobile produces the best performance using a learning rate of 0.0001 and activation function ReLU6. The test results using data from gallery smartphones show that NASNetMobile has an accuracy of 91.6%, while MobileNetV2 has an accuracy of 88.9%. They were testing using the camera in real-time, which resulted in accuracy that was not as accurate as if using data from the gallery. Accuracy using a smartphone camera shows an accuracy of 75% when using NASNetMobile and 72.2% when using MobileNetV2. Accuracy will increase when using a flashlight while capturing objects. Based on the testing results using a flashlight, the NASNetMobile accuracy value increased to 80.5%, while MobileNetV2 accuracy changed to 77.8%.


M-Health; skin disease; transfer learning; MobileNetV2; NASNetMobile

Viewed: 253 times (since abstract online)

cite this paper     download