AI-Enabled Cross-Platform OCT Segmentation

Abstract

Purpose : To develop an OCT segmentation algorithm capable of consistent performance across different OCT devices, addressing a significant interoperability challenge in ophthalmic diagnostics.

Methods : 505 OCT volume scans of normal eyes from either Heidelberg Spectralis (n=329) or Zeiss Cirrus (n=176) were converted into DICOM format, manually segmented using a semi automated algorithm by certified graders. Using a predefined 3D Unet Transformer architecture, a model was trained to segment the total retinal thickness between the internal limiting membrane and the retinal pigment epithelium layer using 382 (75.6%) images (Spectralis: 259; Cirrus: 123). The remaining 123 (24.4%) images (Spectralis: 70; Cirrus: 53) were used to test the algorithm based on the manual segmentation. Dice coefficient was used to measure the degree of overlap for the 2 layers of segmentations.

Results : Average (±SD) central subfield thickness (CST) for the manual segmentation was 255±29 µm while the average for the AI-predicted segmentation was 256±27 µm. The difference between the manual and AI-predicted segmentation within the CST was an average of 1.80±6.63µm with the dice coefficient average was 0.988. There was no significant difference when comparing the dice coefficient between devices (Spectralis: 0.992; Cirrus: 0.982). The AI-predicted segmentations done on the Zeiss Cirrus had greater variability in thickness compared with Spectralis (Spectralis: 2.70±2.17 µm; Cirrus: 0.63±9.72 µm). When comparing across the 6mm ring, the average retinal thickness for the manual segmentation was 261±15 µm with the AI-predicted retinal thickness having similar results of 262±15 µm.

Conclusions : This AI model provides a reliable device-agnostic segmentation of total retinal thickness for both Heidelberg Spectralis and Zeiss Cirrus OCT volume scans. Expanding this segmentation model to eyes with pathology can provide a universal and efficient method to segment OCT scans in diverse clinical settings.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.