Deep-Learning Model Can Identify Trabecular Meshwork in Real-Time Gonioscopy

Researchers have developed a convolutional neural network (CNN) model that can accurately identify the trabecular meshwork in gonioscopy videos in real-time. The model may enhance surgical training, automated screenings, and intraoperative guidance, according to the findings published in Ophthalmology Glaucoma.

It can be difficult to identify iridocorneal structures via gonioscopy accurately. In fact, attendees of the laser trabeculoplasty course conducted during the American Academy of Ophthalmology’s annual meeting consistently rate identifying the trabecular meshwork as one of the most challenging tasks, according to the study. Unfortunately, errors in identifying the structure can lead to adverse outcomes, such as trabeculo-Descemet membrane rupture or treatment failure.

CNNs are deep learning systems modeled on the pattern of neuronal connectivity in the animal visual cortex. Using a hierarchy of trainable layers of convolution filters, the network can determine the relative contribution of each layer and assign a weight to it.

This cross-sectional study examined 21 adults (mean age 47 years, 13 men, 9 women) with open-angle glaucoma and no history of minimally invasive glaucoma surgery. A total of 443 gonioscopic images were taken from different clock hours of both eyes.

The majority of the participants (68%) had glaucoma, and 32% did not. The level of pigmentation of the trabecular meshwork in each image varied from 0 to 4 (lighter to darker) with 5% being grade 0 (n=21), 26% being grade 1 (n=107), 44% being grade 2 (n=180), 17% being grade 3 (n=70), and 7% being grade 4 (n=30).

According to the study, the best CNN model produced test set predictions with a median deviation of 0.8% of the video frame’s height (15.25 µm) from the human experts’ annotations. This error is less than the average vertical height of the trabecular meshwork. The worst test frame prediction offered by the model had an average deviation of 4% of the frame height (76.28 µm), which researchers still consider a successful prediction. When challenged with unseen images, the CNN model scored more than 2 standard deviations above the mean performance of the surveyed general ophthalmologists.

Other findings include that the model:

• Performed as well as glaucoma specialists but outperformed the majority of ophthalmology residents, fellows, cornea specialists, and comprehensive ophthalmologists involved with the study.
• Performed well, even in intraoperative videos that were significantly different than their training data.
• Was trained with a relatively small dataset to identify the trabecular meshwork via gonioscopy accurately.
• Was efficient and could be applied in real-time.

“We anticipate that the proposed neural network model can have applications in surgical training, automated screenings, and intraoperative guidance,” investigators report. “As the model is more widely deployed, the diversity of the existing dataset will be enriched, thus creating a positive feedback loop to further improve model accuracy and generalizability.”

Limitations of this study include its dataset, which only included patients with open-angle glaucoma and without prior angle surgeries. It also only included data from patients who were East Asian or White.

Reference

Lin KY, Urban G, Yang MC, et al. Identification of the trabecular meshwork under gonioscopic view in real time using deep learning. Ophthal Glaucoma. Published online November 16, 2021. doi:10.1016/j.ogla.2021.11.003.