OCT Image Analysis

Optical Coherence Tomography (OCT) is an optical imaging modality that performs high-resolution, cross-sectional tomographic imaging of the eye in real-time. The OCT technology is used as a diagnostic tool designed to assist ophthalmologists in identifying retinal diseases, such as Sickle Cell Retinopathy (SCR) and Diabetic Macular Edema (DME). This work focuses on analyzing the OCT scans to:

  1. Perform retinal layer segmentation
  2. Detect and track retinal thinning due to SCR

Retinal Layer Segmentation (RLS)

OCT examination produces a sequence of retinal cross-section images or B-scans. A retinal diagnostic study usually involves examining the distance between the retinal layers using B-scans. Therefore, detecting the retinal layers is crucial. This work focuses on developing a deep learning-based contour detection system to produce accurate retinal layer segments. Our encoder-decoder-based deep learning network - RLUnet- specializes in performing multi-class contour detection. We have trained and tested the network in the largest benchmarked OCT dataset for RLS. Our evaluation shows a Mean Absolute Distance (MAE) of 1.97±0.35 μM, thus surpassing the current state-of-the-art.

Fig: Left - ground-truth RLS, center - original b-scans, right - RLS from RLUNet

Sickle Cell Retinopathy (SCR) Detection

SCR is a condition of inner retinal thinning observed in patients with Sickle Cell Disease. Ophthalmologists diagnose SCR by studying retinal images obtained from Optical Coherence Tomography (OCT) exam. The need for manual diagnosis and the lack of a large scale SCR dataset has been a topic of concern to anyone interested in studying the disease. In this research, we aim to develop an effective deep learning-based system that detects and tracks the changes in retinal thickness due to SCR. We collaborate with Nemours Children Hospital and renowned opthalmologist Dr Jing Jin to collect and annotate high-quality OCT scans of children with SCD. We approach the problem as an object detection task and therefore annotate the instances of SCR and Fovea with a bounding box. We calculate Mean Average Precision (MAP) based on Intersection over Union (IoU) for evaluation between the annotated and predicted instances. We prepared a dataset containing 3906 B-scans from 33 SCD patients (14 male, 21 female, mean age 12.99±4.56 years) to train and evaluate a YOLOv5 model as a preliminary study. The trained model could correctly identify 85% and 98% of the total SCR and fovea instances respectively. Using this result as a baseline, we aim to develop a more sophisticated system capable of tracking SCR-related retinal thinning progression over time.

Fig: B-scans with SCR and fovea instances enclosed within bounding boxes