Purchase this article with an account.
Carol Yim-lui Cheung, Fangyao Tang, Anran Ran, Gavin Siew Wei Tan, Daniel SW Ting, Haoyu Chen, Hongjie Ma, Shibo Tang, Theodore Leng, Schahrouz Kakavand, Suria S Mannil, Robert Chang, Gerald Liew, Bamini Gopinath, Tien Y Wong, Xi Wang; A Multi-Task Deep-Learning System to Classify Diabetic Macular Edema for Different Optical Coherence Tomography Devices: A Multi-Center Analysis. Invest. Ophthalmol. Vis. Sci. 2021;62(8):1033.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Diabetic macular edema (DME) is the primary cause of irreversible vision loss among individuals with diabetes mellitus (DM). We developed, validated, and tested a deep-learning (DL) system for classifying DME and for retinal abnormalities other than DME simultaneously using images from three common commercially available optical coherence tomography (OCT) devices.
We trained and validated two versions of a multi-task network, one based on three-dimensional (3D) ResNet-34 and another based on two-dimensional (2D) ResNet-18, to classify DME (center-involved DME [CI-DME], non-CI-DME, or absence of DME) and other retinal abnormalities (their presence or absence) using 3D volume-scans and 2D B-scans respectively. A total of 73,746 OCT images, representing 2,444 eyes from 1,238 subjects with DM, were used for training and primary validation. These images include 3,788 3D volume-scans from Cirrus OCT, 30,515 2D B-scans from Spectralis OCT, and 39,443 2D B-scans from Triton OCT. External testing was performed using 3,218 volume-scans from Cirrus OCT, 18,295 B-scans from Spectralis OCT, and 5,468 B-scans from Triton OCT across seven independent datasets under different clinical settings.
In classifying the presence or absence of DME, the DL system achieved AUROCs of 0.937 (95% CI 0.920–0.954), 0.958 (0.930–0.977), and 0.965 (0.948–0.977) for primary datasets obtained from Cirrus, Spectralis, and Triton OCTs respectively, in addition to AUROCs greater than 0.906 for the external datasets. For the further classification of the CI-DME and non-CI-DME subgroups, the AUROCs were 0.968 (0.940–0.995), 0.951 (0.898–0.982), and 0.975 (0.947–0.991) for the primary datasets and greater than 0.894 for the external datasets. In classifying the presence or absence of other retinal abnormalities, the AUROCs were 0.948 (0.930–0.963) 0.949 (0.901–0.996), and 0.938 (0.915–0.960) among images obtained from the Cirrus, Spectralis, and Triton OCTs, respectively, in primary datasets. The performance in external datasets remained excellent, with the ranges for AUROCs being 0.901–0.969.
We demonstrated excellent performance with a DL system for the automated classification of DME, highlighting its potential as a promising second-line screening tool for patients with DM, which will create a more effective triaging mechanism to tertiary eye clinics.
This is a 2021 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only