Evaluation of a UNet Convolutional Neural Network (CNN) for Automatic Measurements of Ellipsoid Zone (EZ) Area and Photoreceptor Outer Segment (POS) Volume in X-Linked Retinitis Pigmentosa (xlRP)

Abstract

Purpose : Manual segmentation of OCT volume scans is time-consuming and costly. CNN models may provide new tools to facilitate this task. Previously we reported that a UNet was effective and efficient for segmenting retinal layers of SD-OCT B-scan images of patients with RP (Wang et al., ARVO 2020). Here we evaluated this model for automatic measurements of EZ area and POS volume from volume scans in xlRP.

Methods : UNet was implemented in MATLAB and trained with 480 midline B-scan images obtained from 220 patients with RP and 20 control subjects. The test included 38 high-speed 9mm 31-line macular volume scans from a separate group of 38 patients with xlRP. All volume scans showed a POS transition zone within the scan window. The Spectralis segmentation of volume scans was corrected manually for EZ and apex RPE to serve as a reference. UNet was used for segmentation of B-scan images in a volume scan to obtain EZ and RPE boundary lines that defined POS. The 3-D POS map was reconstructed by interpolating the discrete 2-D POS layers from 31 B-scans over the grid of scan area. EZ area was measured by multiplying the area of a single grid pixel by the number of pixels having measurable POS. The POS volume was the sum of the products of the grid pixel area and the POS length at the pixel. Bland-Altman and correlation analyses were conducted to compare EZ area and POS volume measured by UNet to the reference.

Results : For EZ area (range: 0.17 to 27.8 mm²), Bland-Altman analysis revealed a mean±SE difference of -1.55±0.25 mm² and CoR of 2.98 mm² between UNet and the reference. EZ area estimated by UNet was highly correlated with the reference (r=0.98; slope=0.85). For POS volume (range: 0.002 to 0.65 mm³), Bland-Altman analysis showed a close agreement between UNet and the reference (mean±SE difference of 0.0004±0.004 mm³ and CoR of 0.05 mm³). POS volume measured by UNet was also highly correlated with the reference (r=0.99; slope = 1.05).

Conclusions : While both EZ area and POS volume determined by UNet had a similar correlation with the reference, UNet tended to underestimate EZ area but had a closer agreement with the reference in measuring POS volume. The deep machine learning method may provide an effective tool for studying the relationship between disease progression and POS volume and EZ area changes in RP.

This is a 2021 ARVO Annual Meeting abstract.