Abstract
Diffusion Bridges (DBs) have emerged as highly efficient generative models for image reconstruction, significantly reducing the iterative steps required by standard Denoising Diffusion Models (DDMs). However, their reliance on clean, ground-truth images for training limits their applicability in domains like medical imaging where such data is scarce. This paper introduces SelfDB, the first self-supervised learning algorithm for Diffusion Bridges. By employing a hierarchical sub-sampling strategy, SelfDB enables the training of high-performance reconstruction networks using only noisy and incomplete measurements. Experimental results on the fastMRI dataset demonstrate that SelfDB outperforms existing self-supervised baselines in both quantitative metrics and perceptual quality while maintaining the speed advantages of the Diffusion Bridge framework.
Introduction
In biomedical imaging, particularly Compressed Sensing MRI (CS-MRI), the central challenge is reconstructing high-fidelity images from undersampled and noisy sensor data. While deep learning and diffusion models have set new benchmarks for these inverse problems, standard DDMs are computationally expensive due to the large number of iterative steps needed to map Gaussian noise to the image manifold.
Diffusion Bridges (DBs) address this efficiency gap by directly interpolating between the measurement distribution and the target image distribution. This approach allows for high-quality reconstruction in as few as 2 to 8 steps. Despite their promise, DBs traditionally require supervised training with paired datasets of noisy measurements and perfect ground-truth references. In many clinical scenarios, obtaining fully sampled ground-truth images is impractical or impossible. While self-supervised techniques like "Ambient Diffusion" exist for DDMs, adapting them to the DB framework has remained an open challenge.
Methodology: The SelfDB Framework
The core innovation of SelfDB lies in its approach to self-supervision within the bridge framework. A naive application of self-supervised methods—where one simply further degrades the available measurements—fails for DBs. This failure occurs because DBs require a meaningful transition between distinct data distributions; interpolating between two nearly identical sub-sampled states causes the model to collapse into a static inverse problem solver.
Hierarchy of Degradation
To overcome this, SelfDB introduces a hierarchy of sub-sampling operators:
- Intermediate Operator ($\overline{M}$): A moderate reduction in sampling.
- Severe Operator ($M'$): A more drastic reduction in sampling.
During training, the forward diffusion process is defined as a transition from the intermediately sub-sampled state to the severely sub-sampled state. The neural network is then optimized to reverse this process. Crucially, the original available measurements ($y$) serve as the training target. By incrementally incorporating information from the intermediate state during the reverse process, the model learns to solve the inverse problem progressively, effectively mimicking a true diffusion process without ever seeing a clean reference image.
Experimental Setup
The authors validated SelfDB using the fastMRI dataset, specifically focusing on T2-weighted brain MRIs from approximately 1,000 subjects. To simulate real-world CS-MRI conditions, fully sampled images were retrospectively sub-sampled.
- Initial Measurements ($y$): 25% sampling rate.
- Intermediate State ($\overline{M}$): 16.6% sampling rate.
- Severe State ($M'$): 12.5% sampling rate.
The fully sampled images were reserved strictly as a hidden ground truth for final metric evaluation (NRMSE, SSIM, and LPIPS).
Key Results
SelfDB was benchmarked against two primary baselines: a naive DB adaptation (Ambient-DB) and a conditional version of Ambient Diffusion (C-Ambient-DDM).
As shown in the figure above, SelfDB achieves superior visual clarity and quantitative performance. Specifically, SelfDB reached an NRMSE of 0.108 and an LPIPS of 0.036, significantly outperforming the DDM-based baseline (LPIPS 0.056). Visually, SelfDB produces sharper textures and more accurate structural details (indicated by the red arrows) compared to the blurred outputs of Ambient-DB or the slightly noisier results of C-Ambient-DDM.
Furthermore, the study highlights the Perception-Distortion Trade-off. By adjusting the number of inference steps at test time, users can prioritize either pixel-wise accuracy (fewer steps) or visual sharpness (more steps), providing a flexible tool for radiologists.
Conclusion
SelfDB represents a significant step forward in making efficient generative models accessible for medical imaging. By removing the requirement for ground-truth training data, it opens the door for high-speed, high-quality MRI reconstruction in settings where only undersampled data is available. Future work may extend this hierarchical degradation strategy to non-Cartesian sampling trajectories and other imaging modalities such as CT or ultrasound.
Summary updated