Research map / Label-efficient and robust learning / Domain adaptation and generalization

Sparse Bayesian Deep Learning for Cross Domain Medical Image Reconstruction

AAAI 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-domain few-shot with transformers · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a lightweight sparse Bayesian deep learning method combining fixed-form variational Bayes uncertainty priors with SGLD-based dynamic prior optimization for cross-domain medical image reconstruction (LDCT denoising and MRI artifact removal).

Problem

Deep reconstruction networks trained on one source medical dataset generalize poorly to target datasets from other hospitals or scanners, and existing cross-domain methods are parameter- and time-heavy.

Approach

Applies fixed-form variational Bayes to obtain pixel-wise uncertainty priors from source-domain degradation, then integrates these priors into SGLD sampling so the weight prior is dynamically updated with unlabeled target-domain data at test time, balancing generalization and computational cost.

Key results

• On Phantom→LDCTPD denoising, the method reaches 32.81 PSNR / 0.955 SSIM versus 31.79 PSNR for DGUNet (the strongest baseline in Table 1).
• On ADNI→MRNet and MRNet→ADNI artifact removal, the method reaches 28.85 and 29.45 PSNR, outperforming DGUNet (26.42 and 26.49).
• Table 5 reports the method has 2.65M MACs and 13.32M parameters versus 22.30M parameters for DGUNet, with shorter per-epoch training time (27.19 vs 46.44) than DGUNet.

Measured effects reported in the paper

Surprising findings

• Bayesian DIP and MAP produce noticeable artifacts, and the paper reports that Bayesian optimization broadens the overfitting window for SGLD and MCD but early stopping is still needed.

Limitations

• Efficiency comparison of training time is less favorable than DIP and CycleGAN in Table 5 (training time 27.19 vs 11.46 for DIP), and the abstract's lightweight claim is relative.
• Hardware and compute setup are not specified, and Table 4 results vary across datasets with some baseline (FFVB+MCD) close to the proposed method on LoDoPaB.

Experimental setup

Datasets: Phantom AAPM LoDoPaB-CT LDCT-and-Projection-data MRNet ADNI

Benchmarks: AAPM LoDoPaB-CT LDCT-and-Projection-data MRNet ADNI

Metrics: PSNR SSIM VIF IFC computational cost (time, params, MACs, memory)

Baselines: DIP CycleGAN Du et al. DPIR DGUNet DeblurGAN MCD MAP Bayesian DIP SGLD

Builds on: Bayes by backprop stochastic gradient langevin dynamics deep image prior DGUNet CPCE

Closest papers by meaning

• Sparse Bayesian Generative Modeling for Compressive Sensing (NeurIPS 2024 · similarity 0.85)
• Training Bayesian Neural Networks with Sparse Subspace Variational Inference (ICLR 2024 · similarity 0.85)
• SSC-VAE: Structured Sparse Coding Based Variational Autoencoder for Detail Preserved Image Reconstruction (AAAI 2025 · similarity 0.85)
• Efficient Noise Calculation in Deep Learning-based MRI Reconstructions (ICML 2025 · similarity 0.84)
• QN-Mixer: A Quasi-Newton MLP-Mixer Model for Sparse-View CT Reconstruction (CVPR 2024 · similarity 0.84)
• Stochastic Deep Restoration Priors for Imaging Inverse Problems (ICML 2025 · similarity 0.84)