Retinal OCT Synthesis with Denoising Diffusion Probabilistic Models for Layer Segmentation

Yuli Wu, Weidong He, Dennis Eschweiler, Ningxin Dou, Zixin Fan, Shengli Mi, Peter Walter, Johannes Stegmaier

Abstract

Modern biomedical image analysis using deep learning often encounters the challenge of limited annotated data. To overcome this issue, deep generative models can be employed to synthesize realistic biomedical images. In this regard, we propose an image synthesis method that utilizes denoising diffusion probabilistic models (DDPMs) to automatically generate retinal optical coherence tomography (OCT) images. By providing rough layer sketches, the trained DDPMs can generate realistic circumpapillary OCT images. We further find that more accurate pseudo labels can be obtained through knowledge adaptation, which greatly benefits the segmentation task. Through this, we observe a consistent improvement in layer segmentation accuracy, which is validated using various neural networks. Furthermore, we have discovered that a layer segmentation model trained solely with synthesized images can achieve comparable results to a model trained exclusively with real images. These findings demonstrate the promising potential of DDPMs in reducing the need for manual annotations of retinal OCT images.

Introduction

In medical AI, high-quality annotated data is a critical but scarce resource. For retinal Optical Coherence Tomography (OCT)—a vital non-invasive imaging technique for diagnosing conditions like glaucoma—training deep learning models for layer segmentation requires thousands of expert-annotated images. This process is both expensive and time-consuming. While Generative Adversarial Networks (GANs) have been used to synthesize data, they often struggle with structural consistency, leading to misalignments between synthesized images and their corresponding ground-truth labels. This paper introduces a novel pipeline using Denoising Diffusion Probabilistic Models (DDPMs) to bridge this gap by generating high-fidelity, structurally accurate OCT scans from simple sketches.

Methodology

The DDPM Synthesis Pipeline

The proposed method leverages DDPMs to transform rough hand-drawn-style sketches into realistic OCT images. The process involves three main stages:

  1. Sketch Generation: Initial "sketches" of retinal layers are created based on the statistical distribution of layer thicknesses and intensities derived from a small real-world dataset (the GOALS Challenge dataset).
  2. Guided Synthesis: These sketches are fed into a DDPM. Rather than starting the denoising process from pure noise ($t=T$), the model utilizes a "short-cut" approach, beginning at an intermediate timestep ($t_{start}$). This preserves the global structure of the input sketch while allowing the model to generate realistic textures and biological details.
  3. Knowledge Adaptation: To ensure pixel-perfect alignment between the synthetic image and its labels, a "Teacher" model (a $U^2$-Net trained on real data) is used to generate refined pseudo-labels for the synthetic images. This "Teacher-Student" framework accounts for minor structural shifts that occur during the denoising process.
Overview of the DDPM-based synthesis pipeline, illustrating training on real OCT images and application using layer sketches.

Experimental Setup

The study utilized 100 circumpapillary OCT images from the MICCAI 2022 GOALS Challenge. The models targeted three specific layers: the Retinal Nerve Fiber Layer (RNFL), the Ganglion Cell-Inner Plexiform Layer (GCIPL), and the Choroid Layer (CL). Performance was evaluated across five neural network architectures, including U-Net, $U^2$-Net, and TransUNet.

Results and Discussion

The "Goldilocks" Zone for Synthesis

A key finding of this research is the critical role of the starting timestep ($t_{start}$). The researchers identified $t_{start}=300$ (out of 400 total steps) as the optimal balance. At lower values, the output remains too similar to the artificial sketch; at higher values, the model ignores the sketch's structural constraints, producing random anatomy.

Visual comparison of synthetic OCT images generated from sketches at different starting timesteps ($t_{start}$), demonstrating the balance between structural guidance and realistic texture.

Performance Gains and Robustness

Integrating 1,000 synthetic images with just 50 real images led to significant performance improvements across all tested architectures. For instance, the standard U-Net's Dice score improved from 84.70% to 91.25%. Notably, the model trained exclusively on synthetic data achieved results comparable to those trained on real, expert-annotated images. The synthetic data proved particularly effective for the Choroid Layer, where boundaries are often faint and difficult for standard models to detect.

Conclusion

This work demonstrates that DDPMs can effectively synthesize high-quality medical images that are structurally aligned with their labels. By utilizing a sketch-to-photo approach and knowledge adaptation, the need for massive human-annotated datasets can be significantly reduced. Future research will focus on incorporating pathological features to synthesize images of specific retinal diseases and exploring unsupervised domain adaptation across different OCT hardware manufacturers.

Summary updated

Abstract

Modern biomedical image analysis using deep learning often encounters the challenge of limited annotated data. To overcome this issue, deep generative models can be employed to synthesize realistic biomedical images. In this regard, we propose an image synthesis method that utilizes denoising diffusion probabilistic models (DDPMs) to automatically generate retinal optical coherence tomography (OCT) images. By providing rough layer sketches, the trained DDPMs can generate realistic circumpapillary OCT images. We further find that more accurate pseudo labels can be obtained through knowledge adaptation, which greatly benefits the segmentation task. Through this, we observe a consistent improvement in layer segmentation accuracy, which is validated using various neural networks. Furthermore, we have discovered that a layer segmentation model trained solely with synthesized images can achieve comparable results to a model trained exclusively with real images. These findings demonstrate the promising potential of DDPMs in reducing the need for manual annotations of retinal OCT images.
BibTeX
@inproceedings{wu2024retinal,
  author = {Yuli Wu and Weidong He and Dennis Eschweiler and Ningxin Dou and Zixin Fan and Shengli Mi and Peter Walter and Johannes Stegmaier},
  title = {{Retinal OCT Synthesis with Denoising Diffusion Probabilistic Models for Layer Segmentation}},
  booktitle = {ISBI},
  year = {2024},
  doi = {10.1109/isbi56570.2024.10635836},
}
Figure