Deep learning to identify diffuse retinal thickening (DRT) on optical coherence tomography (OCT)

Abstract

Purpose : DRT is difficult to assess on OCT, even by expert imaging graders. Given its relevance as an imaging biomarker in diabetic macular edema (DME) and neovascular age-related macular degeneration (nAMD), precise identification of DRT holds potential for personalizing treatment paradigms. Automatic DRT identification on OCT may enable further scientific exploration of this biomarker. We trained and evaluated convolutional neural networks (CNNs) to detect DRT presence in OCT images at b-scan level.

Methods : Our dataset consists of 5133 b-scan images of 276 eyes from patients with DME and nAMD who participated in the clinical trials NCT02484690, NCT02699450, NCT02510794, NCT03038880 and NCT04597918 and the ACDC research collaboration. Graders classified each image into one of 4 categories of DRT: positively present, possibly present, positively absent, ungradable (due to poor image quality). Grading was performed by 4 graders in 90% of the images, and by ≥2 graders in 98%. Due to the rarity of “possibly present” gradings (17 out of 19993), we merged the first 2 categories (DRT-positive). We used the 784 images of 39 patients for validation and the remaining for model selection and training.
For the validation set, the category chosen by the majority of graders was treated as the ground truth resulting in 490 images graded as DRT-negative and 293 as DRT-positive. The single image graded by majority as ungradable was omitted. The graders were in good agreement with the aggregate gradings (Table 1). The hyperparameters of a CNN for this binary classification task (InceptionV3, ImageNet initialization) were selected with a cross-validation nested in the training set.

Results : Over 10 training repetitions, the CNNs classified the images of the validation set with an average area under the receiver operator characteristic (AUROC) of 99.2% (0.4% standard deviation [SD]). In the DME validation subset (424 of 783 images), the AUROC was 98.5% (0.6% SD). Table 2 shows the values of more metrics.

Conclusions : Presence of DRT was detected by CNN models with a performance very similar to that of expert graders. The DL-based automatic identification of DRT can be useful for use in the research or clinical settings and might enable future research on the clinical relevance of DRT.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Table 1. TPR and FPR of the gradings for 7 graders against the aggregate ones

 

Table 2. Performance of the classifiers