Abstract
Purpose :
Various OCT devices are implemented in multicenter clinical trials and retina clinics. However, their differences still pose a challenge in unifying the performance of fluid volume quantification. This study aimed to evaluate the accuracy and reliability of device-specific deep learning models for automated measuring fluid volume in patients with neovascular age-related macular degeneration (nAMD) using two different devices (Spectralis and Triton) and to compare the results with human measurements.
Methods :
Intraretinal fluid (IRF) and subretinal fluid (SRF) volume measurements from 29 patients with nAMD were obtained on device-paired data (Spectralis and Triton) from expert readers via manual annotation and segmentation using MDR-certified device-specific deep learning models: RetInSight TCFM and HEFM. The models had been trained on unpaired manual annotations using different architectures (ensemble of encoder–decoder semantic segmentation networks); however, the validation performance was similar. Bland-Altman analysis was used to investigate the agreement of automated measurements between devices and manual measurements between devices. The limit of agreement was determined via bootstrapping since the data was not normally distributed. Outliers were identified and investigated based on the differences between devices in human measurements but were not excluded.
Results :
The mean bias for AI-based measurement of IRF volume between devices was -8.08 nl (95% CI: -17.72 to -0.56) Human measurement between devices exhibited a mean bias of -17.1 nl (95% CI: -38.48 to -2.66). AI-based agreement on SRF volume showed a mean bias of -7.8 nl (95% CI: -16.31 to -1.88) and the human measurement a mean bias of -9.28 nl (95% CI: -17.78 to -2.65) (Figs. 1 and 2).
Conclusions :
We observed less mean bias and a higher degree of agreement between AI-based fluid volume measurements for Spectralis and Triton than for human measurements. We observed a consistent bias for higher fluid measurements with the Spectralis device (human and AI), indicating that the visible fluid in the Triton device underestimated the actual fluid volume.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.