[Paper Review] Diffusion Adversarial Representation Learning for Self-supervised Vessel Segmentation

Boah Kim, Yujin Oh+1 more|arXiv (Cornell University)|Sep 29, 2022

Retinal Imaging and Analysis30 citations

TL;DR

DARL combines a diffusion module with an adversarial generator to learn vessel representations and produce segmentation masks in a self-supervised manner without ground-truth labels, achieving state-of-the-art results on multiple datasets.

ABSTRACT

Vessel segmentation in medical images is one of the important tasks in the diagnosis of vascular diseases and therapy planning. Although learning-based segmentation approaches have been extensively studied, a large amount of ground-truth labels are required in supervised methods and confusing background structures make neural networks hard to segment vessels in an unsupervised manner. To address this, here we introduce a novel diffusion adversarial representation learning (DARL) model that leverages a denoising diffusion probabilistic model with adversarial learning, and apply it to vessel segmentation. In particular, for self-supervised vessel segmentation, DARL learns the background signal using a diffusion module, which lets a generation module effectively provide vessel representations. Also, by adversarial learning based on the proposed switchable spatially-adaptive denormalization, our model estimates synthetic fake vessel images as well as vessel segmentation masks, which further makes the model capture vessel-relevant semantic information. Once the proposed model is trained, the model generates segmentation masks in a single step and can be applied to general vascular structure segmentation of coronary angiography and retinal images. Experimental results on various datasets show that our method significantly outperforms existing unsupervised and self-supervised vessel segmentation methods.

Motivation & Objective

• Motivate vessel segmentation in medical images with limited or no labeled data.
• Introduce a diffusion-adversarial representation learning framework (DARL) to learn vessel semantics.
• Enable simultaneous generation of synthetic angiograms and vessel segmentation masks via switchable SPADE layers.
• Demonstrate robustness to noise and generalization across retinal and x-ray angiography datasets.

Proposed method

• Integrate a diffusion module that estimates latent background signals from noisy inputs.
• Use a generation module with switchable SPADE layers to produce vessel masks and synthetic angiograms.
• Employ two discriminators to distinguish real vs. fake segmentation masks and angiograms.
• Train end-to-end with diffusion loss, adversarial losses, and cycle-consistency loss.
• Inference performs a single-step segmentation using latent features from the diffusion module (non-iterative).

Experimental results

Research questions

• RQ1 Can a non-iterative diffusion-based model learn vessel-relevant representations without labels for segmentation?
• RQ2 Does adversarial generation with switchable SPADE layers improve segmentation and synthetic angiogram quality jointly?
• RQ3 How well does the method generalize across datasets and modalities (angiography and retinal images)?
• RQ4 Is the method robust to noise and varying imaging conditions commonly seen in medical vessels?
• RQ5 What is the contribution of each component (diffusion module, S-SPADE, cyclic loss) to overall performance?

Key findings

• DARL outperforms unsupervised and self-supervised baselines on X-ray coronary angiography and retinal datasets.
• The method achieves higher Dice, IoU, and precision scores across external XCA tests and cross-modality retinal datasets.
• Ablation shows diffusion module, S-SPADE, and cyclic loss each contribute to performance gains.
• DARL exhibits robustness to Gaussian noise and maintains reasonable segmentation under varied noise levels.
• Latent features from the diffusion path emphasize vessel structures, supporting effective segmentation without labels.