Deep learning-based segmentation uncertainty improves the correlation between RNFL structure and visual function

Abstract

Purpose : The uncertainty quantification of segmentation results is critical for understanding the reliability of the segmentation model. The purpose of this study is to investigate the effect of deep learning-based segmentation with uncertainty measurement in the relationship between RNFL thickness and visual field mean deviation (MD).

Methods : Optical coherence tomography (OCT) scans were acquired from both eyes on 634 glaucoma patients, 404 glaucoma suspects, and 49 healthy controls using commercial OCT device (Cirrus HD-OCT, 200x200 Optic Disc Cubes; Zeiss, Dublin, CA). All subjects had visual field (VF) tests at each visit (Humphrey VF, SITA 24-2 test; Zeiss). A segmentation model was trained using Bayesian deep learning for voxel-wise segmentation of RNFL layer in OCT volume and compute the voxel-wise uncertainty of the segmentation output. The higher uncertainty denotes the unreliability of the segmentation and vice versa and it allows the determination of erroneous segmentation at test time. Uncertainty-guided global mean of the RNLF thickness (RNFL-U_mean) was then computed by discarding the voxels with erroneous segmentation labels with higher uncertainty during the thickness computation. Also, the global mean of the RNLF thickness (RNFL_mean) was computed without taking uncertainty into account. Pearson correlation coefficient between RNFL-U_mean and MD was computed and compared with the Pearson correlation coefficient between RNFL_mean and MD.

Results : The proposed RNFL-U_mean gave stronger correlation with MD than RNFL_mean. The Pearson correlation coefficients were (0.67 (RNFL-U_mean ) vs 0.63 (RNFL_mean); p<0.001) for glaucoma subjects, (0.56 vs 0.53 ;p=0.01) for glaucoma suspects and (0.08 vs 0.01; p=0.21) for normal subjects.

Conclusions : The proposed uncertainty-guided computation of RNFL thickness showed improved correlation with the visual field MD. This demonstrates that segmentation uncertainty can be used to reduce the effect of inaccurate segmentation in computing the RNFL thickness. This also shows that uncertainty-guided computation of RNFL thickness is a better predictor of visual function than the normal RNFL thickness computed without using uncertainty.

This is a 2021 ARVO Annual Meeting abstract.