Abstract
Purpose :
State-of-the-art retinal OCT analysis often uses deep learning approaches trained on annotated data, which requires tedious expert tracings in large quantities. To minimize expert effort while maintaining high performance, a novel approach DeepLogismosRetina is reported as a general tool for accelerated model construction and accurate layer segmentation across devices, demonstrated here on choroidal layer segmentation.
Methods :
Segmentation methods are sensitive to OCT data origin. When a method developed for OCT device A is applied to data from device B, results may be imperfect. Our open-access DeepLogismosRetina offers intuitive, efficient, and algorithmic adjudication approach (Just-Enough Interactions or JEI) to quickly modify such imperfect results and create large-enough ground truth sets for additional training, thus achieving acceptable performance on device B. The underlying method in DeepLogismosRetina is based on a hybrid, inherently 3D strategy of deep learning and graph optimization (Fig. 1).
Results :
Tested on choroidal segmentation, the ground truth on 30 Spectralis (device B) images was created by applying CirrusSeg tool developed for Cirrus OCT (device A) and performing expert JEI to adjudicate resulting inaccuracies. The Spectralis-specific deep learning model was then trained on these 30 images. For performance evaluation, we used 22 eyes of 11 subjects that were imaged by both Cirrus and Spectralis OCTs. Choroidal thickness was underestimated by CirrusSeg on Cirrus images due to their lack of choroidal texture detail while the DeepLogismosRetina trained model achieved accurate results on Spectralis images. Comparing layer thicknesses from Cirrus and Spectralis images within a 3mm-radius fovea-centered circle showed improved performance of DeepLogismosRetina approach (Fig. 2).
Conclusions :
Our open-access DeepLogismosRetina application combines the strengths of deep learning and graph optimization to produce accurate results with guaranteed-correct layer topology. It utilizes the availability of imperfect results to streamline ground truth generation for training deep learning models with improved performance across OCT devices and disease populations. While demonstrated on choroidal segmentation, the reported DeepLogismosRetina approach is applicable to any retinal layered structure and deep learning network.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.