Development of mobile application for detection and grading of diabetic retinopathy

• Main Authors: Kipli, Kuryati, Lee, Yee Hui, Tajudin, Nurul Mirza Afiqah, Sapawi, Rohana, Sahari, Siti Kudnie, Awang Mat, Dayang Azra, A. Jalil, M., Ray, Kanad, Kaiser, M. Shamim, Mahmud, Mufti
• Format: Book Section
• Published: Springer Science and Business Media Deutschland GmbH 2022
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://eprints.utm.my/id/eprint/101088/; http://dx.doi.org/10.1007/978-981-16-8826-3_29

The key to preventing blindness caused by diabetic retinopathy (DR) is regular screening and early recognition during its early stages. Currently, DR grading is done manually by ophthalmologists and trained graders where the process is time-consuming. Therefore, this paper aims to develop a mobile app that can provide DR detection and grading without a professional or doctor. The patients will be referred to ophthalmologists if further evaluations are required. This research builds an image classification within a mobile application by using deep learning techniques which utilized the Google AI technologies: Google TensorFlow and Google Cloud Platform (Cloud AutoML and Cloud storage). Image classification is performed in two layers which involve DR detection and grading. A total of 12,062 fundus images are chosen from the dataset collected and undergo image preprocessing. The preprocessed images are used to train the model in TensorFlow and Cloud AutoML, respectively. The model will be implemented into the mobile application after being trained with high accuracy. The final test accuracy for the MobileNet pretrained model is 82.9%, while averaging precision for the model of Cloud AutoML is 75%. Further research is required to improve the stability of this algorithm and mobile app for real clinical environment settings.