358 — Clinically applicable deep learning for diagnosis and referral in retinal disease

OCT is a powerful technology that enables comprehensive, deep, 3D imaging of retinal tissue noninvasively. It’s a powerful tool for neurological disease diagnosis [#14 #38 #65] and for better understanding the anatomy and physiology of the eye [#94 #177]. But — much like in the field of connectomics — image segmentation is one of the greatest time-costs and challenges. This task often requires human intervention and proofreading, which is why automation of this process is so enticing. At APL, colleagues of mine recently published a deep-learning approach to OCT segmentation. This paper, by DeepMind, performs training on almost 15,000 retinal scans and it too achieves better-than-human accuracy when segmenting the retinal scans in 3D. In particular, it seems that 3D networks dramatically (and unsurprisingly) outperform 2D nets.

The net also performed at expert level at diagnosis and referral based upon these images. Because these two processes were separated into two phases, the authors explain that the results are more trustworthy, as they depend less on end-to-end black-box results.