Automated Machine Learning Pipeline for Predicting Retinal Sensitivity from Optical Coherence Tomography in Macular Telangiectasia Type 2

Abstract

Purpose : Optical coherence tomography (OCT) imaging and microperimetry are key modalities that allow the assessment of retinal structure and function loss associated with macular telangiectasia type 2 (Mactel) such as foveal detachment, IS/OS disruption, retinal hyperreflectivity, intraretinal pigment migration, and overall retinal sensitivity. Using OCT and microperimetry data, we developed a fully automated machine learning pipeline to explore the structure and function correlation in Mactel and predict retinal sensitivity from OCT data alone.

Methods : Data obtained from the Mactel Study included baseline and multiple-follow up (1, 3, 12, 18, 24, and 36 months) OCTs and microperimetry results registered on the MAIA infrared scanning laser ophthalmoscope images of study participants. We first built a deep learning model for segmentation of blood vessels in OCT infrared (IR) and MAIA images by training U-Net on Digital Retinal Images for Vessel Extraction database. Using the retinal vessels as landmarks, the OCT IR images with B-scans were mapped and superimposed to the MAIA images with microperimetry points (Figure1). Using these data, we trained a deep learning model that generates a dense retinal sensitivity map.

Results : Baseline and follow-up data from 103 study patients (ages 43-80) were used. A total of 241,961 microperimetry sensitivities were mapped onto OCT B-scans. Mean absolute error (MAE) between predicted and observed retinal sensitivity was 3.12 dB (95% CI: 3.07 to 3.16 dB). Our model was tested on another dataset with a total of 2,499 microperimetry sensitivities that were manually mapped onto OCT B-scans from 63 eyes of 38 patients and achieved MAE of 3.51 dB (95% CI: 3.22 to 3.80 dB).

Conclusions : We present a new automated machine learning pipeline that generates high resolution en face maps of predicted retinal sensitivities from OCTs, and have applied it to a large-scale dataset by introducing registration algorithm that registers two sets of images with different modalities. Our pipeline can be used to monitor structural and functional disease progression in MacTel (Figure2) and might be applicable for other macular diseases.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Schematic of machine learning pipeline.

 

Observed microperimetry (1st, 3rd) and predicted sensitivity maps (2nd, 4th). The predicted functional loss (red) increased over time, in keeping with a slowly progressive disorder.