Automated 2D-3D quantitative analysis of corneal graft detachment post DSAEK based on AS-OCT images

In current ophthalmological practices, assessment of graft condition post Descemet's stripping automated endothelial keratoplasty (DSAEK) is performed qualitatively using few (four) anterior-segment optical coherence tomography (AS-OCT) radial B-scans. From those scans, clinicians need to mentally synthesize the graft in 3D, and estimate its overall condition. In contrast, quantitative representation of 360° thickness profile would facilitate better visualization of graft condition, and hence medical decision making. Consequently, clinicians seek to quantify potential detachments in 3D from the aforementioned sparse B-scans. Against this backdrop, aiming to assist doctors in making an accurate postoperative assessment, we attempted at 3D visualization and quantification of donor lenticule separation (DLS) using only four AS-OCT radial B-scans.

We developed an automated methodology to model and quantify DLS in 3D using only four AS-OCT B-scans. Firstly, we demonstrated that detachment can be viewed as a tubular vessel-like structure and hence can be detected using Hessian matrix analysis. Secondly, a two-stage interpolation was employed for determining the 3D profile of the detachment. Finally, various clinically significant parameters including type of separation (communicative and non-communicative), volume and 360° thickness profile of the detachment, thickness (central and peripheral) of the recipient cornea and donor graft were also obtained. Accuracy of the proposed algorithm was substantiated via thorough statistical analysis, specifically, vis-á-vis intra- and inter-observer repeatability using Dice coefficient (DC).

On twenty seven eyes of 27 patients (male and female), four radial OCT B-scans with 45° angular separation taken per eye, proposed algorithm found that donor graft detached completely in 3 eyes and detached one side (communicative) in remaining 24 eyes which is in agreement with expert's opinion. Quantitatively, proposed algorithm achieves a mean DC of 81.35% with manual reference which is close to the corresponding inter-observer repeatability value of 86.77%. Volume estimation of the detachment indicates that 11 eyes had 0-25%, 9 had 25-50%, 5 had 50-75% and 2 had 75-100% detachment.

Accuracy of the proposed methodology was corroborated vis-à-vis observer delineation. This model of image analysis may aid in prognostication of graft outcome in patients with graft detachment after DSAEK.