Continual Active Learning for OCT Scans

 Collaborators: Zoe Fowler, Kiran Kokilepersaud, Mohit Prabhushankar, Ghassan Alregib

Goal: Retinal diseases manifest themselves as aberrant biomarkers in Optical Coherence Tomography (OCT) scans. Analyzing these scans is commonly used to detect eye diseases, such as DR and DME, and patients with these diseases receive treatment at every visit. The goal of this project is to create a continual active-learning framework that mimics a clinical setup by accommodating the addition of new visit data at every training round by sampling the most informative OCT scans for labeling in a time-series set of scans.

Challenges: The first challenge is due to the domain shift. Because patients are receiving treatment at every medical visit, there is a slight amount of domain shift at each visit’s eye scans. Over time, this domain shift is so drastic that a model trained with data from later visits fails to perform well on data from earlier visits, a phenomenon referred to as regression or catastrophic forgetting. The second challenge is understanding how to sample the most informative OCT scans for training the model.

Our Work: Previous work has focused on different query strategies for active learning using OCT scans. In particular, [1] improved generalization performance for OCT disease classification by partitioning the training pool by patient identity and sampling based upon this partition, demonstrating how different query strategies can influence model results. Such an approach can be taken for this project, where the query strategies investigated can include using standard query strategies (like random, entropy, etc.) and/or incorporating clinical biomarkers into the query strategy. In addition, our method must combat regression, and the first step in this process is to understand and model the domain shift across visits. Afterwards, we will use a metric, such as the negative fliprate (NFR), to evaluate how much the model is forgetting.

References:

  1. Y. Logan, R. Benkert, A. Mustafa, G. Kwon, G. AlRegib, "Patient Aware Active Learning for Fine-Grained OCT Classification," IEEE International Conference on Image Processing (ICIP), Bordeaux, France, Oct. 16-19 2022. [PDF][Code]