Abstract
Purpose :
Microperimetry (MP) is a valuable endpoint in clinical trials for various retinal diseases. Deep learning models that can predict MP from optical coherence tomography (OCT) B-scans may be beneficial given the laborious effort of MP. Transfer learning involves using pre-trained weights on a different but related problem where proficiency in one task may generalize to another. We sought to evaluate a deep learning model pre-trained on OCTs of macular telangiectasia type 2 (MacTel) and fine-tuned on patients with non neovascular age-related macular degeneration (AMD).
Methods :
The AMD dataset includes 322 eyes of 161 patients enrolled in the Laser Intervention in the Early Stages of AMD (LEAD) trial at the Centre for Eye Research Australia from 2013 to 2018, partitioned into 110,450 samples for training, 52,858 for validation, and 41,719 for testing. Each sample consists of B-scan window slices and MP sensitivities, for which infrared fundus photos were registered by aligning segmented blood vessels. Three VGGNet models were evaluated: 1) trained on MacTel data (VGG-Mac), 2) trained on AMD data (VGG-AMD), 3) pre-trained with MacTel data and fine-tuned with AMD data (VGG-FT). Primary outcome is the mean absolute error (MAE) of predicted vs. observed MP sensitivity.
Results :
All models were evaluated against the same AMD validation set. The fine-tuned model VGG-FT achieved lowest MAE of 2.88 dB (95% CI 2.86, 2.91). VGG-AMD achieved MAE of 3.01 dB (2.99, 3.04), and VGG-Mac achieved MAE of 3.44 dB (3.41, 3.48). Continuous predictions by VGG-FT across B-scan levels were then aggregated to create 2D en-face maps (Figure 2).
Conclusions :
The model pretrained on MacTel then fine-tuned on AMD data outperformed the model trained only on AMD data, suggesting that transfer learning between different retinal diseases may increase domain specific performance. Deep learning models to estimate retinal sensitivities using OCTs may be a valuable endpoint in following patients with AMD for clinical trials.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.