Abstract
Purpose :
Few methods have been proposed to predict the morphology of the macula after full-thickness macular hole (FTMH) surgery. In this multicenter retrospective observational study, we implement a conditional variational autoencoder (CVAE) that learns preoperative and postoperative three-dimensional optical coherence tomography (OCT) data of FTMH patients and propose a predictive model capable of estimating the macular morphology as well as its thickness.
Methods :
A 6 x 6 mm2 volume scan OCT dataset of 125 successfully closed FTMHs was enrolled from two tertiary hospitals. After extracting retinal pigment epithelial layers as point clouds, we derived a rigid transformation through least-squares fitting and registered preoperative OCT data. As far away from the fovea, different condition vectors were set at intervals of 500 μm to the horizontal OCT image of each volumetric data. Randomly selected and augmented 96,000 training pairs (from 80 sets of volumetric data) and 4,000 validation pairs of OCT images were used to train the model (Fig. 1). The F1 score of the retinal region, multi-scale structural similarity (SSIM), foveal height (FH), mean foveal thickness in the central 1 mm (MFT), mean parafoveal thickness profile (MPTP, a 2.5 mm region centered on the fovea), and subfoveal ellipsoid zone (EZ) attenuation were assessed to measure the performance of the model using independent test sets.
Results :
Of 25 test sets, the mean F1 score between the generative and actual postoperative OCT images of the retinal region was 0.92 ± 0.03. Multi-scale SSIM was 0.74 ± 0.03. Predictive versus postoperative FH (228.0 vs. 237.7 μm, p = 0.53) and MFT (MFT; 277.2 vs. 287.4 μm, p = 0.29) were similar. The root-mean-squared error between predictive and postoperative MPTP was under 10 pixels. The accuracy of prediction for subfoveal EZ attenuation was 76.0%.
Conclusions :
The proposed CVAE model can predict macular features and morphology after FTMH surgery, even in small datasets. By presenting realistic and convincing postoperative OCT images, the models could assist patients and ophthalmologists in planning surgery for FTMH.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.