Please use this identifier to cite or link to this item:
|Title:||A machine learning-enabled mobile app for glaucoma detection||Authors:||Toshiko, Seki Jennifer||Keywords:||Engineering::Electrical and electronic engineering::Electronic systems::Signal processing||Issue Date:||2022||Publisher:||Nanyang Technological University||Source:||Toshiko, S. J. (2022). A machine learning-enabled mobile app for glaucoma detection. Master's thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/161747||Abstract:||While some illnesses can be diagnosed based on simple metrics, afflictions like glaucoma tend to rely on doctor subjectivity for a diagnosis. Studies have shown machine learning algorithms, which remove this subjectivity, can actually outperform doctors in correct glaucoma diagnoses. In addition, screenings done at clinics can be expensive and time-consuming. This dissertation proposes a machine learning-enabled Android mobile app called Glaucoma AI for glaucoma detection. The classification algorithm used in the mobile app was created by Yuan Liu in . The Attention-Guided Stereo Ensemble Network (AGSE-Net) consists of Convolutional Neural Network (CNN) and Attention branches. The network was modified to classify non-stereo fundus images of the retina and integrated into the app. Users can classify images taken within the app – with the use of a smartphone fundus photography attachment – or images selected from the mobile device photo gallery. When tested on the RIM-ONE DL data set, the app was able to classify images with 84.54% accuracy, 91.27% specificity, and 72.29% sensitivity. The app is all-in-one in that it does not require resources outside the mobile device to run and only requires Internet connection during installation. The Glaucoma AI app uses approximately 2.17 GB of device internal storage. During a typical run, the peak CPU usage is 87% and peak memory usage is 0.7 GB on the Samsung Galaxy Tab S7. Glaucoma AI is the only glaucoma screening app that has a simple, easy-to-use interface, only requires computational resources within the mobile device, was trained and tested with various data sets to show realistic results, and has well-documented implementation and testing details.||URI:||https://hdl.handle.net/10356/161747||Schools:||School of Electrical and Electronic Engineering||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||EEE Theses|
Updated on Oct 2, 2023
Updated on Oct 2, 2023
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.