Novel deep learning tool shows promise in classifying DME

Approach shows potential as a promising second-line screening tool for patients with diabetes.

Fangyao Tang, MD, from the Department of Ophthalmology and Visual Sciences, The Chinese University of Hong Kong, Hong Kong, and colleagues reported the “excellent performance” of a deep learning system1 that they developed to classify diabetic macular edema (DME) using images obtained from 3 commercially available optical coherence tomography (OCT) systems, Cirrus (Zeiss), Spectralis (Heidelberg Engineering), and Triton (Topcon) OCTs.

Underscoring the potential of such a tool, the investigators emphasized that DME is the primary cause of visual loss among individuals with diabetes.

Dr. Tang reported that his team trained and validated 2 versions (using three-dimensional [3D] volume scans and 2D B-scans, respectively) of a multitask convolution neural network (CNN) to classify the DME as center-involved or non-center-involved DME and no DME.

For both versions of the CNNs, he described, they used the residual network (ResNet) as the backbone. For the 3D CNN, they used a 3D version of the ResNet-34 with the last fully connected layer removed as the feature extraction module.

The training and validation processes involved use of 73,746 OCT images. External testing was performed using 26,981 images across 7 independent data sets from Singapore, Hong Kong, the U.S., China, and Australia, he reported.

Deep learning analysis

During the classification of the presence/absence of DME, the investigators reported that this new system achieved area under the receiver operating characteristic curves (AUROCs) of 0.937 (95% confidence interval [CI], 0.920–0.954), 0.958 (0.930–0.977), and 0.965 (0.948–0.977), respectively, for the primary data set obtained using the Cirrus, Spectralis, and Triton OCTs, in addition to AUROCs >0.906 for the external data sets.

When classifying the center-involved and non-center-involved DME subgroups, the respective AUROCs for the 3 OCT systems were 0.968 (0.940–0.995), 0.951 (0.898–0.982), and 0.975 (0.947–0.991) for the primary data set and >0.894 for the external data sets.

Based on these results, the investigators concluded, “We demonstrated excellent performance with a [deep learning] system for the automated classification of DME, highlighting its potential as a promising second-line screening tool for patients with diabetes, which may potentially create a more effective triaging mechanism to eye clinics.”

1. Tang F, Wang X, Ran A-R, et al. A multitask deep-learning system to classify diabetic macular edema for different optical coherence tomography devices: a multicenter analysis. Diabetes Care. 2021;44: 2078-2088;