Enhancing Retinal Disease Classification from OCTA Images via Active Learning Techniques

TL;DR

This paper tackles retinal disease classification from OCTA images under the constraint of limited labeled data. It introduces an active learning framework with both subject-based and instance-based sampling to selectively label informative OCTA images, using an Inception-V3 backbone and uncertainty metrics to guide sample selection. Empirical results on the OCTA500 dataset show that active learning significantly outperforms traditional class-balancing baselines, with Ratio sampling achieving the highest accuracy and F1 scores, and ablations highlight the importance of sampling strategy and probability calibration. The findings demonstrate that active learning can enable robust OCTA-based disease detection in imbalanced data settings and suggest future work on diversification and explainability of imaging biomarkers.

Abstract

Eye diseases are common in older Americans and can lead to decreased vision and blindness. Recent advancements in imaging technologies allow clinicians to capture high-quality images of the retinal blood vessels via Optical Coherence Tomography Angiography (OCTA), which contain vital information for diagnosing these diseases and expediting preventative measures. OCTA provides detailed vascular imaging as compared to the solely structural information obtained by common OCT imaging. Although there have been considerable studies on OCT imaging, there have been limited to no studies exploring the role of artificial intelligence (AI) and machine learning (ML) approaches for predictive modeling with OCTA images. In this paper, we explore the use of deep learning to identify eye disease in OCTA images. However, due to the lack of labeled data, the straightforward application of deep learning doesn't necessarily yield good generalization. To this end, we utilize active learning to select the most valuable subset of data to train our model. We demonstrate that active learning subset selection greatly outperforms other strategies, such as inverse frequency class weighting, random undersampling, and oversampling, by up to 49% in F1 evaluation.

Figures (2)

• Figure 1: Comparison of OCT and OCTA data for Normal, CNV, DR, and AMD eyes
• Figure 2: Example of a single active learning iteration, where $\mathcal{D}_{\{train, val\}}$ are the train and validation set, $s$ is the uncertainty score, and $k$ is the number of images to move at each iteration