• Using Deep Learning to Evaluate Macular Thickening

    By Lynda Seminara
    Selected By: Deepak P. Edward, MD

    Journal Highlights

    Investigative Ophthalmology & Visual Science

    Download PDF

    Arcadu et al. set out to determine whether deep learning could be used to predict optical coherence tomogra­phy–equivalent quantitative measures of diabetic macular thickening (MT), using color fundus photographs. They found that it could, and they suggested that, when used in this manner, deep learning models could significantly benefit teleophthalmology initiatives.

    For this study, the authors obtained data from the phase 3 RIDE and RISE studies of diabetic macular edema (DME); nearly 18,000 color fundus im­ages were included. Deep learning with a transfer-learning cascade was applied to the photographs to predict time-domain optical coherence tomography (TD-OCT)–equivalent MT measures, including central subfield thickness (CST) and central foveal thickness (CFT). Two conventional TD-OCT cutoff points—250 μm and 400 μm—were used to identify abnormal MT. A deep learning regression model was created to quantify actual CST and CFT measurements from the fundus photo­graphs. Four models of deep convolu­tional neural networks were analyzed (two each for CST and CFT).

    The best deep learning model was able to predict CST ≥250 μm and CFT ≥250 μm, with area under the curve (AUC) of 0.97 and 0.91, respectively. For CST and CFT predictions of ≥400 μm, AUC of the best model was 0.94 and 0.96, respectively. The best neural network regression model to quantify CST and CFT had an R2 of 0.74 and 0.54, respectively. The models were less accurate when images were of poor quality or if laser scars were present.

    The researchers cautioned that their findings may not be generalizable to the overall population of patients with diabetes. In addition, it’s possible that the deep learning model is not truly detecting macular thickening but rather retinal phenotypes. Although abnormal thickening does correlate with such phenotypes, the authors affirmed that the deep learning model can detect abnormal MT regardless of diabetic retinopathy severity or the presence of hard exudates. More research is needed to validate such models with real-world data.

    The original article can be found here.