Investigators hope to develop a novel artificial intelligence model.
Reviewed by Austin Pereira, MD, MEng.
The development of a deep learning model to help conduct segmentation of swept-source optical coherence tomography (OCT) images to measure the volume of macular holes was the focus of a presentation in the Lions’ Lair competition, a retina research proposals competition held during the 2021 Canadian Ophthalmological Society Annual Meeting and Exhibition.
“The premise of this is to possibly augment the International Vitreomacular Traction Study [IVTS] group’s classification of macular holes from 2013,” said Austin Pereira, MD, MEng, a second-year ophthalmology resident at the University of Toronto, department of ophthalmology and vision sciences, in Toronto, Canada. Pereira made a pitch for $35,000 in hopes to fund the development of a novel artificial intelligence (AI) model.
“The IVTS upgraded the initial classification macular holes by taking an OCT-dependent and anatomical classification of macular holes,” Pereira said. “They revolutionized macular hole classification through the minimum linear diameter of the neurosensory retinal deficit through OCT imaging, as well as adding the presence of vitreomacular traction into the equation.”
Macular hole classification through IVTS is crucial for surgical planning and prognostic considerations for patients, Pereira explained.
“The IVTS group took anatomical classification as the main etiology for macular holes,” Pereira said.
Pereira noted that small macular holes measure less than 250 µm, medium-sized macular holes measure between 251 µm and 400 µm, and large macular holes measure more than 400 µm.
However, vitreomacular traction can occur in areas other than the macula, according to Pereira. Moreover, the retina is 3-dimensional tissue.
“Looking at the retina in a 2-dimensional manner is doing a disservice not only to surgeons but to our patients,” Pereira said.
Indeed, using a measure like minimum linear diameter is 1-dimensional, yet the retina is a 3-dimensional structure, Pereira added.
A 2014 study that highlighted this mismatch looked at the impact of OCT scan pattern and density on the detection of full-thickness macular holes. The study by Tamer Mahmoud et al, compared the use of 1-dimensional horizontal raster line scan with 3-dimensional radial raster line scan, which looks at 360° to detect macular holes.
With 3-dimensional viewing, vitreomacular traction was easier to identify, according to Pereira. The study concluded that high density radial scanning is superior to standard raster volume scanning in detecting small full-thickness macular holes, he noted.
Moreover, research has shown high inter- and intra-user variability within same eye measurements for minimum linear diameter macular holes.
“This changes the prognostic considerations and surgical planning, bringing a patient at times from a small macular hole classification to a large deficit, and vice versa,” according to Pereira.
With minimum linear diameter, there is a tremendous amount of OCT segmentation dependence, he added.
“With [minimum linear diameter] being our main biomarker for surgical planning for macular hole surgery, our biomarkers need to be as accurate, precise, and reproducible as possible for our patients,” Pereira explained.
Pereira, under the supervision of the primary investigator, Netan Choudhry, MD, FRCSC, DABO, employed a deep learning model to obtain automated measurements of macular hole volume, which he described as a robust, 3-dimensional biomarker for macular hole size that can assist in surgical planning and prognostic counseling for patients with full-thickness macular holes.
The pilot study involved 25 patients from the Vitreous Retina Macula Specialists of Toronto (VRMTO) practice, who had full-thickness macular holes, where 3-dimensional OCT images of the holes were obtained preoperatively, 1 month after surgery, and 1 year after surgery.
A single vitreoretinal surgeon, Choudhry, medical director at VRMTO, manually assessed the minimum linear diameter and macular hole volume. A convolutional neural network was developed to assess volume in an automated fashion.
“This will ultimately provide the volume of the macular hole, or detect the presence of a macular hole,” Pereira noted. “The main outcome measure was the correlation between the human grader [assessment of the full-thickness macular hole volume], and the automated volume by our deep learning model.”
Investigators also looked to see if there were correlations between volume and change in vision in comparison with minimum linear diameter, according to Pereira.
“The correlation between automated macular hole volume, as determined by our deep learning model, and the manual macular hole volume, as determined by our vitreoretinal surgery, was 0.94,” he said. “There was a poor correlation between minimum linear diameter and change in vision from the preoperative time point to 1 year postoperatively. There was a significantly higher correlation using our automated macular hole volume.”
The deep learning model needs to be validated before being implemented in daily surgical practice, Pereira noted.
“This would give vitreoretinal surgeons another biomarker in their arsenal to help plan surgery and to improve prognostic outcomes for patients,” he said. “The only way to do that is through large, multicenter analysis. We need to demonstrate to surgeons that this will help their clinical acumen. We have to demonstrate it is more efficacious than a 1-dimensional measurement.”
Increasing the amount of OCT scans in the training set will help reveal what the macular hole volume truly is, according to Pereira.
Future directions may see AI provide information, such as how much peeling of the internal limiting membrane is needed, and employ data from known past surgeries, Pereira added. The variability in minimum linear diameter was the driving force in the project.
“We want to get away from 1-dimensional measures and more robust grading of the size [of the macular hole],” he said. “The volume is the starting point.”
The funding would be used to extract the OCT scans from the participating sites, segment them, and run the AI model, Pereira added.