Abstract
Purpose :
Deep learning is increasingly used for automating tasks in medicine, such as diagnosis of ophthalmic images. However, most previous approaches to this problem have required the use of time-intensive manual labeling of images. We used self-supervised learning to train a network to extract features from a large number of unlabeled optical coherence tomography (OCT) images, and then used this feature extractor as the backbone of a classification system trained on a small set of labeled images.
Methods :
A total of 94,847 OCT volumes from 16,710 patients obtained from University of Washington were used. These were split into an unsupervised training set of 88,834 volumes (15,716 patients), a training set of 4490 volumes (795 patients), and a validation set of 1523 volumes (199 patients). Diagnostic labels were used for the training and validation sets, but not the unsupervised training set. Unsupervised training was done using the DINO framework. The resulting models produced feature vectors from B-scans. The models were then frozen and used to extract features from the labeled training set. Linear regressors were fit on the extracted features to predict age and logMAR, and a k-nearest neighbor (kNN) classifier was fit to diagnose age-related macular degeneration (AMD), diabetic eye disease, and glaucoma (Figure 1).
Results :
The OCT-trained network had an R2 of 0.61 for age and 0.32 for LogMAR. Its Average Precisions (AP) for AMD, diabetic eye disease, and glaucoma were 0.87, 0.70, and 0.56, respectively. The age R2 and all three diagnostic scores were higher than for the ImageNet-trained model, and the diagnostic scores for both models were significantly higher than the baseline prevalence (Figure 2).
Conclusions :
Here we leverage unlabeled images to perform feature extraction with deep learning and then use a small set of labeled images to train simple algorithms using these features. This approach may be useful for training deep learning models with few labeled examples. Pretraining on ophthalmic images without labels can improve model performance on labeled datasets.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.