Application of Semantic Segmentation Models to Quantify En-face Optical Coherence Tomography Imaging Abnormalities in Uveitis - 5297
Mon statut pour la session
Author’s Disclosure Block: Fan Ze Wang: none; Zejia Chen: none; Shuja Khalid: none; Efrem Mandelcorn: none; Tina Felfeli: none
Abstract Body
Purpose: The presence of ellipsoid zone abnormalities is associated with worse visual acuity in uveitis patients. This study aims to characterize out-of-the-box and fine-tuned performances of the Segment Anything Model (SAM) variant MedSAM for segmenting ellipsoid zone abnormalities detected on en-face optical coherence tomography (OCT) in patients with uveitis. Study Design: Retrospective cohort analysis. Methods: OCT images from 94 eyes (55 patients) with idiopathic intermediate, posterior, or panuveitis between 2010-2021 were obtained. En-face OCT imaging was performed to determine ellipsoid zone abnormalities. A total of 153 image-ground truth pairs were split into 80% for training (fine-tuning), 20% for testing. Using bounding-boxes as segmentation prompts, we compared the performance between an out-of-the box and a fine-tuned MedSAM. Dice Similarity Coefficients (DSCs) were calculated for both models. Linear correlation (Pearson correlation coefficient) between the segmentation pixel ratio (segmented pixel area divided by the total image area) and the logarithm of the minimum angle of resolution (logMAR) visual acuity was determined for each model. Results: Fine tuning of the MedSAM model demonstrated high accuracy in abnormality segmentation with a mean Dice coefficient of 0.85 (95% confidence interval 0.81, 0.88), compared to an out of the box model, which had a Dice coefficient of 0.61 (95% confidence interval 0.51, 0.71). A positive linear correlation between the segmentation pixel ratio and the logMAR visual acuity measured at the initial and final visit was observed to have a Pearson correlation coefficient of 0.613 with a r-squared value of 0.3758. Conclusions: These results highlight the potential of automated segmentation models for efficient and precise detection of subtle ellipsoid zone abnormalities, which may further enhance diagnostics and therapeutics for uveitis.