Development and evaluation of an R library to align macula and optic disc centered optical coherence tomography (OCT) scans

Abstract

Purpose : For many eye diseases, macula centered and optic disc centered OCT scans are separately obtained at the same visit for the same eye. To better detect or interpret clinical findings or retinal layer thickness maps, it is helpful to align them to a combined image, as illustrated in Fig.1. Here, we use OCT scans from a population-based study to develop a R library for this purpose and evaluate it on a separate clinical data set.

Methods : Scanning laser ophthalmoscopy fundus images of accompanying macula and optic disc centered Spectralis SD-OCT scans were extracted from all eyes of the baseline measurements of the age and sex stratified, population-based Leipzig Research Centre for Civilization Diseases - LIFE Adult study (training dataset) and from the most recent exam at the Mass. Eye and Ear glaucoma service (evaluation dataset). Affine image registration functions were developed in R, applied to the training set, and checked for failures by visual inspection. A random forest model was trained on the statistics of the affine transformations to predict failures, based on which a parameter failureAlert (between 0 and 1 for lowest/highest failure probabilities, respectively) was added as an output of the registration function. The resulting R library was applied to the evaluation dataset. The ability of the library to autonomously detect its own failures was analyzed by relating failureAlert to true failures.

Results : Training/evaluation datasets consisted of 18,047/3,570 eyes from 9,116/2,035 patients. 3,351 out of the 3,570 image pairs (93.9%) of the evaluation set were correctly aligned. Fig.2A shows the distribution of failureAlert for true successes and failures. The interquartile ranges (blue boxes) of both categories do not overlap. To quantify the usability of failureAlert to detect failures, a logistic regression model of failureAlert was fitted to true outcomes (Fig.2B). The area under the curve was 0.91.

Conclusions : Our newly developed R library correctly aligned 93.9% of macula/optic disc centered OCT scan pairs of a clinical evaluation dataset. Our software contains a failureAlert parameter (the higher, the more likely a registration failure), which may help to decide if a registration should be visually checked. failureAlert could discriminate well between true successes and failures (area under the ROC curve: 0.91).

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Fig.1

 

Fig.2