Validation Study for Corneal Microlayer Tomography Automatic Segmentation Algorithm

Abstract

Purpose : To validate a proposed automatic segmentation algorithm (AuS) against trained manual operators (TMO) for corneal microlayer optical coherence tomography (CMLT) registered images of normal corneas.

Methods : A high resolution optical coherence tomography machine (HD-OCT; Envisu R2210, Bioptigen, Buffalo Grove, IL, USA) was used to obtain CMLT images. Fifteen eyes were chosen randomly from 66 normal eyes. Images were registered using custom-made registration method. Automatic segmentation was done by adaptively thresholding the image columns then rows to keep high intensity pixels. Then, median filter is applied. Corneal epithelium (EP) and endothelium (EN) were estimated using RANSAC and polynomial fitting. The original image was then flattened using EP and EN and projected to find relative locations of other layers. Finally, other layers were estimated as perturbed replicas of EP and EN at their relative locations from EP and EN. Two masked trained manual operators segmented each image twice. The intra-operator error and inter-operator error were calculated for the AuS and the TMO. The time needed for segmentation was compared. A masked cornea specialist randomly graded the segmentation accuracy for each layer, in the central, and the peripheral parts of the segmented images for both the AuS and the TMO on a scale of 5 where 5 is “Excellent” and 1 is “poor”.

Results : The mean intra-operator error in pixels for the AuS was significantly less across all layers as compared to the TMO (0.87±0.81 vs. 2.45±0.51 and 2.2±0.49; P=0.05, Fig. 1). There was a significant inter-operator error in pixels between the two TMO (3.84±1.93) and between each one of them and the AuS (3.56±0.80 and 4.44±2.20). The mean time of the AuS in seconds was significantly less compared to the TMO (0.25±0.07 vs. 159.10± 71.72 and 154.25±51.74; P=0.05). There was no statistically significant difference in the subjective accuracy grading score between the AuS and the TMO (4.88±0.10 vs. 4.94±0.09 and 4.99±0.03; P=0.001; Fig. 2).

Conclusions : The proposed AuS has less intra-operator errors and comparable inter-operator error when compared to the manual segmentations of the TMO. The AuS is significantly faster than the manual segmentation. The AuS has comparable accuracy to the manual segmentation done by the TMO across all layers of the cornea of normal eyes.

This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.

 

Fig. 1 Intra-operator Error Comparison

 

Fig. 2 Subjective Test Accuracy Comparison