Deep Learning-Powered Imaging Tool Measures Retinal Disease

LONG BEACH, CA – APRIL 01: LONG BEACH, CALIF. USA –Tony Urbiha has her eyes checked by Dr. Stanley Carson in his Long Beach, Calif. office on June 8, 2011. Urbiha suffers from macular degeneration. Just a few years ago, people of all ages who had the degenerative eye disease were told that they would eventually go blind. Then came a breakthrough, a treatment that seems to halt and sometimes even slightly reverse the condition in one kind of macular degeneration. The treatments include frequent shots in the eyeball to prevents the progression of the disease. Patients come at regular intervals, and special cameras show the condition of the retina and whether a shot is required at that visit. (Photo by Jeff Gritchen/Digital First Media/Orange County Register via Getty Images)
The technology can provide clinically useful measures of retinal diseases that would otherwise be difficult or not feasible to obtain.

A deep learning-powered segmentation tool can provide clinically useful measures of retinal diseases that would otherwise be difficult or not feasible to obtain, according to research results published in JAMA Ophthalmology. 

Optical coherence tomography (OCT) interpretation can be challenging in complex macular diseases — particularly when different clinical interpretations may be equally valid. In this study, researchers assessed the clinical applicability of a deep learning model to provide evaluations of these disease states.

The research included 5 datasets, split between an initial pilot study (set 1) and the main study (set 2), with 1 eye per patient selected. The first set — inclusive of 1 dataset — includes 15 OCT scans from patients with new presentations of severe age-related macular degeneration (AMD). The second set — inclusive of 4 datasets — included a total of 164 OCT scans that had not been previously used to train the models. 

For the first set, 2 specialist optometrist graders manually segmented all B-scans of each OCT, looking for 3 pathological features: intraretinal fluid, subretinal fluid, and pigment epithelial detachment. These segmentations were adjudicated independently by 2 senior ophthalmologists. In the second set, subsets were segmented by 2 certified graders, then adjudicated by 1 senior ophthalmologist. Scans from those with DME were segmented only for intraretinal and subretinal fluid, while the AMD scans were segmented for all 3 features as well as subretinal hyperreflective material. 

In total, 173 scans were included for analysis (AMD n=107; DME n=66). Grading to a mean of 50 and 7 hours per scan for sets 1 and 2, respectively. For the model, mean time was less than 10 seconds per scan. 

In the qualitative analysis, investigators found that the model was either better at or comparable with at least 1 expert in 73% of scans, and worse in 27% of scans (95% CI, 66% to 79% and 21% to 34%, respectively). When considering both slight and moderate differences, the model was either better or comparable to at least 1 expert in 86% and worse in 14% of scans (95% CI, 80% to 90% and 10% to 20%). In cases where the model ranked the highest, there was typically a slight difference from the next expert grade (range, 40% to 75%). 

Most specialists gave positive Likert ratings of 4 or 5 to 49% of the model segmentations and to 65% of the expert segmentations (225 of 346, with 2 per scan); this increased to 78% and 89%, respectively. 

In quantifying analysis, the model segmented higher volumes of intraretinal fluid (mean difference, -0.16 mm3), comparable volumes of subretinal fluid and subretinal hyperreflective material (mean differences, 0.01 and 0.002 mm3), and lower volumes of pigment epithelial detachment. Both linear and logarithmic limits of agreement were wider for model-grader than intergrader agreement. 

The anatomical and functional features of retinal pathologies each have individual roles in the prognosis. But features such as intraretinal fluid can be difficult to delineate when the OCT images are of suboptimal resolution or contrast, which can occur when retinal thickening —a common occurrence in DME — reduces clarity of the cystoid boundaries. Using deep learning, physicians can, potentially, automate the analysis of these features.

Study limitations include the challenges fundamental to quantitative validation of deep learning-based segmentation tools, the morphological complexity of severe pathology, and the sheer scale of the OCT segmentation task. 

“Segmentations were acceptable to specialists in most cases, and qualitative evaluation provided valuable insights in addition to more traditional analysis,” according to the report. “This tool has already advanced the understanding of the anatomical characteristics in AMD and in the future may provide novel quantitative end points for clinical trials to enable in-depth analysis.” 

“Automated segmentation systems offer the potential to transform clinical workflows,” the research says. 

Disclosure: Multiple study authors declared affiliations with the biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures.  


Wilson M, Chopra R, Wilson MZ, et al. Validation and clinical applicability of whole-volume automated segmentation of optical coherence tomography in retinal disease using deep learning. JAMA Ophthalmol. Published online July 8, 2021. doi:10.1001/jamaophthalmol.2021.2273