Abstract
Purpose :
Deep learning based optical coherence tomography (OCT) model needs continuous updates in practical deployment to learn new diseases and improve performance. However, continuous conventional training suffers from catastrophic forgetting of old classes and violates patient privacy by storing data. We performed a translational study to test the hypothesis that non-storage privacy-preserving continual learning (CL) OCT model can continuously learn and maintain learned classes performance for retinal pathologies such as drusen, diabetic macula edema (DME), choroidal neovascularization (CNV).
Methods :
We tested 11 CL algorithms among three main CL methodologies, such as regularization, expansion and replay, on the public OCT dataset. This dataset contains over 84,495 OCT images, including normal (31.1%), drusen (10.2%), CNV (44%), DME (13.4%). There are 250 testing images (1.1%) for each class. Because most CL algorithms are developed from non-medical dataset CIFAR10, it is included as the benchmark. It consists of 60,000 images in 10 daily classes, such as cat and truck. There are 5,000 training images (83.3%) and 1,000 testing images (16.6%) per class. Convolutional neural networks (CNN), specifically EfficientNetB4, and ResNet18, were used. In OCT training, CNV and normal were trained in the first task, DME was trained in the second task and drusen was trained in the third task. For CIFAR10 dataset, 4 classes were trained in the first task, and 3 classes were trained in the second and the third task. After the last task, all classes accuracies were measured. The conventional deep learning model was sequentially trained, and the centralized model of all classes was trained together as control.
Results :
Among all CL OCT models, the expansion based Efficient Feature Transformations (EFT) algorithm shows the best result with 51.72%, 78.64% and 60.47% accuracy respectively. While the conventional deep learning model showed a completely dropped performance on all previous tasks. Similarly, EFT algorithm performs best on CIFAR10 dataset.
Conclusions :
Our study is consistent with our hypothesis that privacy-preserving CL application in OCT shows better old classes performance compared to the conventional machine learning model. Further improvement in CL accuracy would be beneficial for future OCT AI model deployment in primary care settings
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.