[Paper Review] Towards Label-Free 3D Segmentation of Optical Coherence Tomography Images of the Optic Nerve Head Using Deep Learning

Sripad Krishna Devalla, Tan Hung Pham+17 more|arXiv (Cornell University)|Jan 1, 2020

Retinal Imaging and Analysis71 references4 citations

TL;DR

This paper proposes a label-free, device-agnostic deep learning framework for 3D segmentation of optic nerve head tissues in optical coherence tomography (OCT) images. By using a two-stage approach—first enhancing image quality across three OCT devices via a device-specific enhancer network, then applying a segmentation network (ONH-Net) trained on one device to accurately segment tissues on the other two—the method achieves high Dice scores (>0.92) without requiring manual segmentation for new devices, enabling broad clinical translation.

ABSTRACT

Recently proposed deep learning (DL) algorithms for the segmentation of optical coherence tomography (OCT) images to quantify the morphological changes to the optic nerve head (ONH) tissues during glaucoma have limited clinical adoption due to their device specific nature and the difficulty in preparing manual segmentations (training data). We propose a DL-based 3D segmentation framework that is easily translatable across OCT devices in a label-free manner (i.e. without the need to manually re-segment data for each device). Specifically, we developed 2 sets of DL networks: the 'enhancer' (enhance OCT image quality and harmonize image characteristics from 3 devices) and the 'ONH-Net' (3D segmentation of 6 ONH tissues). We found that only when the 'enhancer' was used to preprocess the OCT images, the 'ONH-Net' trained on any of the 3 devices successfully segmented ONH tissues from the other two unseen devices with high performance (Dice coefficients > 0.92). We demonstrate that is possible to automatically segment OCT images from new devices without ever needing manual segmentation data from them.

Motivation & Objective

• To address the clinical limitation of device-specific deep learning models in OCT-based glaucoma segmentation.
• To eliminate the need for manual segmentation annotations when deploying models on new OCT devices.
• To develop a framework that enables cross-device generalization of 3D segmentation without retraining with labeled data.
• To improve the clinical translatability of deep learning tools in glaucoma management across diverse OCT scanners.

Proposed method

• A two-stage deep learning framework is proposed: an image enhancer network followed by a 3D segmentation network (ONH-Net).
• The enhancer network is trained on unlabelled OCT volumes from three different OCT devices to harmonize image characteristics such as contrast and speckle noise.
• ONH-Net is trained exclusively on data from one device but leverages enhanced images from other devices to achieve cross-device segmentation.
• The enhancer ensures consistent image quality across devices, enabling ONH-Net to generalize without retraining on labeled data from new devices.
• The framework uses only 100 unlabelled volumes per new device to retrain the enhancer, avoiding the need for time-intensive manual segmentation.
• The approach is distinct from transfer learning, as the segmentation and enhancement networks learn independently, with the enhancer enabling generalization without sharing knowledge with ONH-Net.

Experimental results

Research questions

• RQ1 Can a deep learning-based 3D segmentation framework achieve high performance across multiple OCT devices without requiring manual segmentation for each device?
• RQ2 How effective is a device-specific image enhancer in enabling cross-device generalization of a 3D segmentation network?
• RQ3 Does the use of enhanced images from multiple OCT devices improve segmentation accuracy and robustness compared to raw images?
• RQ4 To what extent does the framework maintain segmentation consistency across different OCT scanners, especially for critical glaucoma-related structures like the lamina cribrosa?
• RQ5 Can this label-free approach reduce the time and cost of deploying deep learning models in clinical settings with heterogeneous OCT equipment?

Key findings

• The proposed framework achieved mean Dice coefficients above 0.92 for all six optic nerve head tissue layers across all three OCT devices, demonstrating high segmentation accuracy.
• ONH-Net trained on data from one device successfully segmented tissues from the other two devices with no significant performance difference (p > 0.05), confirming device independence.
• The image enhancer was critical for cross-device performance, as segmentation accuracy dropped significantly without it.
• The framework enabled automated segmentation on new OCT devices using only 100 unlabelled volumes per device, eliminating the need for manual annotations.
• Segmentation performance remained clinically reliable across devices, with no significant differences in Dice scores regardless of the training device.
• Despite minor variations in lamina cribrosa curvature and thickness across training devices, the overall segmentation quality was deemed clinically acceptable.