Abstract
Purpose :
Although intraocular anti-vascular endothelial growth factor (VEGF) injection is now considered first-line treatment for diabetic macula edema (DME), 30-50% of patients poorly respond to the treatment, resulting in poor compliance and placing heavy finical burden on patients. We developed a deep-learning (DL) system for predicting anti-VEGF treatment response for eyes with DME using optical coherence tomography (OCT) images, aimed to personalize treatment for patients with DME.
Methods :
Training dataset from an OCT device (Spectralis, Heidelberg Engineering, Germany) were collected from patients with center involved-DME (CI-DME) who received 3 injections within 6 months from 3 hospitals in Hong Kong. We extracted macular volumetric scans prior to receive anti-VEGF treatment, obtained with high-resolution 6.3mm × 6.3mm (25 B-scans) and high-speed 6.5mm × 4.9mm (19 B-scans) scanning protocol. A volumetric scan with good response was defined according to the DRCR.Net protocol-defined thresholds: ≥1-line gain on the Early Treatment Diabetic Retinopathy Study letter score or Snellen VA chart and >10% reduction in central subfield thickness. Label of a given volumetric scan was applied to its corresponding B-scans. Treatment response-related features were manually segmented in approximately two representative B-scans of each volumetric scan to train the system to predict robust outcome. A convolutional neural network-based B-scan classifier was established and trained with the volumetric labels. By feeding all B-scans of a given volumetric scan sequentially as input, we took the maximum prediction value among all B-scans as the volume-level result (good or poor response, Fig.1).
Results :
949 OCT volumes (23,356 B-scans) from 679 eyes of 531 patients were included for training (60%), validation (20%) and testing (20%). The DL system achieved dice scores of 0.689, 0.880, and 0.856 for segmentation of sub-retinal fluid, outer-retinal defects, and disorganization of the retinal inner layers. The system achieved AUROC of 0.731 for prediction good or poor response at volume-level.
Conclusions :
The DL system could predict anti-VEGF treatment response with good performance. It may optimize treatment modality for an individual’s condition of DME for better compliance and less financial burden.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.