![SonomaEye-full-color-vector[2].png](https://static.wixstatic.com/media/bed00b_da20b9cc704044b0b048ac8ef9190470~mv2_d_3024_2195_s_2.png/v1/fill/w_177,h_122,al_c,q_85,usm_0.66_1.00_0.01,enc_avif,quality_auto/SonomaEye-full-color-vector%5B2%5D.png)
Utilizing Machine Learning to Distinguish Uveitic Macular Edema from Diabetic Macular Edema Using Optical Coherence Tomography
Presenter:
Paul S Zhou, MD
Authors:
Paul Zhou, MD¹, Taylor Crook¹, Pooya Khosravi¹˒², Olivia L. Lee, MD¹
1. Department of Ophthalmology, School of Medicine, University of California, Irvine, Irvine, CA, USA
2. Department of Computer Science, Donald Bren School of Information and Computer Sciences, University of California, Irvine, Irvine, CA, USA
Affiliation:
Purpose: To develop and validate a machine learning model to differentiate cystoid macular edema (CME) secondary to diabetic retinopathy from uveitic macular edema (UME) using optical coherence tomography (OCT) images.
Methods: Patients with CME were identified using ICD-10 codes from the electronic medical record between February 2018 and May 2025 at a quaternary academic referral eye hospital. Eyes were excluded for recent intraocular surgery; concurrent diabetic macular edema (DME) and active uveitis; central serous chorioretinopathy; vitelliform dystrophy; age-related macular degeneration; or other inherited or degenerative maculopathies. Diagnoses were confirmed by comprehensive chart review, and all images underwent manual grading.
The final dataset included 925 OCT volumes (full macular cube and 21-line raster protocols) from 94 eyes of 48 patients, yielding individual 2D B-scans labeled as DME (n = 99), UME (n = 292), or no edema (n = 534). An additional 534 B-scans from contralateral eyes without macular edema served as controls. A FastViT-SA12 model integrating convolutional neural networks and vision transformer components was trained and validated using a patient-stratified 80/20 train–validation split.
Results: The model achieved an overall accuracy of 68.0% with an area under the receiver operating characteristic curve (AUC) of 83.4%. For UME, sensitivity was 50.2%, specificity 76.3%, precision 0.35, recall 0.50, and F1-score 0.41. For DME, sensitivity was 26.3%, specificity 97.0%, precision 0.67, recall 0.26, and F1-score 0.38. Normal OCT B-scans demonstrated sensitivity of 86.8%, specificity 73.1%, precision 0.84, recall 0.87, and F1-score 0.85.
Conclusions: Differentiating UME from DME based on OCT morphology alone remains challenging. This pilot study demonstrates that a FastViT-SA12–based model shows promising discriminatory capability among UME, DME, and normal OCT B-scans. Machine learning approaches may assist in identifying subtle imaging features and improving diagnostic accuracy and efficiency. Future work will expand sample size to improve model validity and evaluate performance in previously excluded eyes with CME in the setting of both diabetes and uveitis.