Robust AMD Stage Grading with Exclusively OCTA Modality Leveraging 3D Volume

Abstract

Age-related Macular Degeneration (AMD) is a degenerative eye disease that causes central vision loss. Optical Coherence Tomography Angiography (OCTA) is an emerging imaging modality that aids in the diagnosis of AMD by displaying the pathogenic vessels in the subretinal space. In this paper, we investigate the effectiveness of OCTA from the view of deep classifiers. To the best of our knowledge, this is the first study that solely uses OCTA for AMD stage grading. By developing a 2D classifier based on OCTA projections, we identify that segmentation errors in retinal layers significantly affect the accuracy of classification. To address this issue, we propose analyzing 3D OCTA volumes directly using a 2D convolutional neural network trained with additional projection supervision. Our experimental results show that we achieve over 80% accuracy on a four-stage grading task on both error-free and error-prone test sets, which is significantly higher than 60%, the accuracy of human experts. This demonstrates that OCTA provides sufficient information for AMD stage grading and the proposed 3D volume analyzer is more robust when dealing with OCTA data with segmentation errors.

Figure 1.

Comparison between fundus and OCTA w.r.t. AMD stages. As shown by the blue arrows, all AMD stages exhibit drusens and it is difficult to differentiate each stage based on the pattern of drusens. For instance, in the provided example, the early stage (dry) displays clearer drusens than the progressive stage (active). In contrast, OCTA allows for a distinction between dry and normal stages using the hollows in CC projection, and between active and dry stages with the presence of CNV in avascular projection. It is still an ongoing challenge to tell active stage from remission for human experts, yet this paper demonstrates it is achievable with the proposed deep classifiers in both 2D and 3D cases.

Figure 2.

The proposed network structures for (a) 2D projections, (b) 3D volumes and (c) 3D volumes with 2D projection supervision. The layers in blue have pretrained weights while those in red are trained from scratch.

Figure 3.

Illustration of the interrelationships among OCT and OCTA raw volume, B-scans, and OCTA projection. A single B-scan is a cross-sectional slice of the 3D volume with a specific y-axis value. Retinal slab masks are derived from retinal layer segmentation in each B-scan of 3D OCT volume. OCTA projections are generated by summing up the motion responses in selected OCTA slabs followed by quality enhancement.

Figure 4.

Examples of avascular projection w/o and w/ manual layer segmentation error correction by human experts. Layer segmentation errors lead to incorrect vascular networks, missing vessels and noise in OCTA projections, which complicates the classification for both ophthalmologists and neural networks.

Figure 5.

Confusion matrix comparison between our proposed method and human experts on different test sets. Ours-3D, outperforms human experts in accurately distinguishing between the ‘active’ and ‘remission’ categories. Also, as indicated by the smaller performance drop observed in the ‘dry’ category, our method demonstrates greater robustness to layer segmentation errors.