Ogul Uner, MD, of Oregon Health & Science University, presented a study at AAO 2025 that used artificial intelligence to assess OCT scans and determine whether patients had vitreoretinal lymphoma.
Ogul Uner, MD:
Hello. My name is Dr. Ogul Uner. I’m a current uveitis fellow at KCI Institute at Oregon Health & Science University. I’m very excited to talk about the talk that I gave at AAO 2025 titled “Differentiating Vitreoretinal Lymphoma and Non-Infectious Uveitis: An OCT-Based Artificial Intelligence Model.” As we know, vitreoretinal lymphoma is most commonly a diffuse large B-cell lymphoma and it can present as pseudo-uveitis in about 75% to 90% of cases. Patients get diagnosed very late as a result with a mean time to diagnosis from symptom onset of 15 months. Hence, we need better tools to identify these patients earlier. An OCT can be a useful tool, but a lot of features on OCT can overlap with a lot of uveitic disorders. Hence, our question whether artificial intelligence can help differentiate the 2 entities using OCT. Our aim was to train, validate, and test a deep learning model to differentiate the 2 entities and we hypothesized that the training model would be able to better characterize it compared to chance alone.
To accomplish this, we took spectral domain OCT macula scans of biopsy-proven vitreoretinal lymphoma patients, intermediate posterior and panuveitis patients, and normal scans and we only included patients who had clinically active disease above the age of 45 and excluded patients with infectious uveitis and other macular pathologies. Our main outcome measures were the performance metrics based on sensitivity, specificity, and area under the receiving operator characteristic curve. We had about 43 vitreoretinal lymphoma eyes and 51 uveitis eyes, with over 2,000 B-scans that we fed into our deep learning model; 70% used to train, 15% to test, and 15% to validate the model. How we developed this was a custom convolutional neural network where we put spectral domain OCT images through a variety of different deep learning columns and produced an image based on gradient-weighted class activation mapping. This is a unique tool that we use that allows us to highlight different areas of the OCT that the AI is using to contribute to the model’s decision making. That was unique for us, because we wanted to see why is the model highlighting it as uveitis or lymphoma, so we can see that directly in real time.
The model was quite successful. We had about 80% to 90% area under the receiver operating characteristic curve, depending on the models that we used. When we analyzed it as a binary model, so lymphoma or not, the model had moderate specificity and sensitivity and overall, performed reasonably well. Big things that the model took into account were vitreoretinal interface abnormalities like epiretinal membranes, intraretinal cells and scleral and choroidal thickness changes were more indicative of uveitis and RPE changes like excrescences, subretinal or sub-RPE deposits and other changes in the outer retina were more identified as lymphoma. This is really exciting, because we need better tools to diagnose these patients earlier and this proof of concept study shows that artificial intelligence can help in differentiating the 2 entities, not necessarily to replace the diagnostic biopsy, which is the gold standard right now, but to clue the clinician into the diagnosis earlier so we can find, diagnose and treat these patients to ultimately provide better outcomes. Thank you.
Related Content
