Retinal OCT Synthesis with Denoising Diffusion Probabilistic Models for Layer Segmentation

TL;DR

An image synthesis method that utilizes denoising diffusion probabilistic models (DDPMs) to automatically generate retinal optical coherence tomography (OCT) images is proposed and it is discovered that a layer segmentation model trained solely with synthesized images can achieve comparable results to a model trained exclusively with real images.

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.

Figures (6)

• Figure 1: Generation Pipeline. The upper workflow illustrates the denoising diffusion probabilistic model (DDPM) ho2020denoising training process using the real retinal circumpapillary OCT images. The lower pipeline illustrates the data synthesis from the sketches with the trained DDPM. The dashed arrows from right to left denote forward diffusion processes (adding noise) and the solid arrows from left to right denote reverse diffusion processes (denoising).
• Figure 1: Ablation of Blurring and Perturbation with $t_{start}$=300. The average and the best total Dice scores among 5 networks are listed.
• Figure 2: Histograms of a real image (blue) and a sketch (orange) w.r.t. the timestep $t$ in the forward diffusion process.
• Figure 3: A layer ground-truth label image is shown in (a). The corresponding sketch image is shown in (b). Synthetic OCT images generated using this sketch image are illustrated in (c-f) based on DDPMs starting with different timesteps $t_{start}$.
• Figure 4: Total Dice ($y$) to $t_{start}$ ($x$) with 5 networks using B and P.