Learning spatially varying regularisation parameters of low regularity for image reconstruction
本文探讨了通过学习空间变化的正则化参数来改善图像重建中边缘和细节保留的问题，特别是结合深度神经网络的方法。
0 citationsRead paper
Selected work
Representative Papers
Augment to Augment: Diverse Augmentations Enable Competitive Ultra-Low-Field MRI Enhancement
Ultra-low-field MRI (ULF-MRI) suffers from inherently low signal-to-noise ratio, poor spatial resolution, and contrast distortion. Existing image-to-image translation methods are hindered by the scarcity of paired high-field/low-field 3D volumetric data. To address this, we propose a task-adaptive data augmentation framework that, using only 50 paired 3D volumes, integrates strong geometric and intensity augmentations alongside auxiliary supervision derived from high-field images—specifically structural consistency reconstruction—to enhance model generalizability. Our method significantly improves ULF-MRI image quality, achieving third place in brain mask SSIM and fourth place in overall test score on the ULF-EnC Challenge public leaderboard. The source code is publicly released, establishing a reproducible paradigm for few-shot medical image enhancement.
0 citationsRead paper
Estimating systematic errors in Bayesian inversion using transport maps
In Bayesian inverse problems, approximation errors in the measurement process model induce systematic bias, distorting the posterior distribution. This paper proposes a unified framework integrating transport maps with Bayesian model error modeling to jointly and adaptively estimate both the corrected posterior and the model error. The key innovation lies in constructing a differentiable transport map from a reference distribution to the bias-corrected posterior, while coupling it with a stochastic process representation of the model error to enable dynamic bias correction. The method is theoretically rigorous and computationally efficient. It is validated on two canonical indirect measurement problems, demonstrating significantly improved correction accuracy and sampling efficiency over conventional approaches. By explicitly accounting for structural model inadequacy, the framework effectively mitigates posterior distortion arising from model misspecification.
0 citationsRead paper
Plans for acceptance sampling by attributes when observations are destructive
ISO 2859-2 exhibits inaccurate estimation of residual lot quality risk—particularly for small lots—and inflated consumer’s risk in destructive testing due to non-recoverable samples. To address this, this paper proposes a novel Bayesian attribute acceptance sampling method. It innovatively treats the residual lot size as a fixed parameter and integrates a hypergeometric likelihood with a reference prior to construct a decision-theoretic framework that rigorously controls consumer’s risk. The resulting standardized sampling plans are compact and computationally efficient, substantially reducing actual consumer’s risk in small-lot scenarios. This approach extends the applicability of acceptance sampling beyond the limitations of conventional standards in destructive inspection settings. Its theoretical foundation—grounded in objective Bayesian inference—and empirical efficacy support its potential adoption into international standardization frameworks.
0 citationsRead paper
MRpro - open PyTorch-based MR reconstruction and processing package
Existing MR image reconstruction tools suffer from limited data operation consistency, inflexible algorithm design, poor deep learning integration, and low reproducibility. To address these limitations, we propose and open-source a modular, PyTorch-based framework for MR reconstruction and processing. The framework introduces a unified differentiable Fourier operator, an extensible phase map simulator, and a data-consistency layer supporting Cartesian, radial, spiral, and other sampling trajectories. It seamlessly integrates model-based optimization (e.g., proximal algorithms) with deep learning components—including learnable regularizers and composable network backbones—to enable motion correction, MR fingerprinting, and quantitative parameter mapping. Standardized data structures, public dataset interfaces, and plug-and-play operator design significantly improve reproducibility and collaborative development efficiency. We validate the framework’s generality and effectiveness across multiple quantitative imaging tasks.
0 citationsRead paper
Recent publications
Latest Papers
Learning spatially varying regularisation parameters of low regularity for image reconstruction
本文探讨了通过学习空间变化的正则化参数来改善图像重建中边缘和细节保留的问题，特别是结合深度神经网络的方法。
0 citationsRead paper
Augment to Augment: Diverse Augmentations Enable Competitive Ultra-Low-Field MRI Enhancement
Ultra-low-field MRI (ULF-MRI) suffers from inherently low signal-to-noise ratio, poor spatial resolution, and contrast distortion. Existing image-to-image translation methods are hindered by the scarcity of paired high-field/low-field 3D volumetric data. To address this, we propose a task-adaptive data augmentation framework that, using only 50 paired 3D volumes, integrates strong geometric and intensity augmentations alongside auxiliary supervision derived from high-field images—specifically structural consistency reconstruction—to enhance model generalizability. Our method significantly improves ULF-MRI image quality, achieving third place in brain mask SSIM and fourth place in overall test score on the ULF-EnC Challenge public leaderboard. The source code is publicly released, establishing a reproducible paradigm for few-shot medical image enhancement.
0 citationsRead paper
Estimating systematic errors in Bayesian inversion using transport maps
In Bayesian inverse problems, approximation errors in the measurement process model induce systematic bias, distorting the posterior distribution. This paper proposes a unified framework integrating transport maps with Bayesian model error modeling to jointly and adaptively estimate both the corrected posterior and the model error. The key innovation lies in constructing a differentiable transport map from a reference distribution to the bias-corrected posterior, while coupling it with a stochastic process representation of the model error to enable dynamic bias correction. The method is theoretically rigorous and computationally efficient. It is validated on two canonical indirect measurement problems, demonstrating significantly improved correction accuracy and sampling efficiency over conventional approaches. By explicitly accounting for structural model inadequacy, the framework effectively mitigates posterior distortion arising from model misspecification.
0 citationsRead paper
Plans for acceptance sampling by attributes when observations are destructive
ISO 2859-2 exhibits inaccurate estimation of residual lot quality risk—particularly for small lots—and inflated consumer’s risk in destructive testing due to non-recoverable samples. To address this, this paper proposes a novel Bayesian attribute acceptance sampling method. It innovatively treats the residual lot size as a fixed parameter and integrates a hypergeometric likelihood with a reference prior to construct a decision-theoretic framework that rigorously controls consumer’s risk. The resulting standardized sampling plans are compact and computationally efficient, substantially reducing actual consumer’s risk in small-lot scenarios. This approach extends the applicability of acceptance sampling beyond the limitations of conventional standards in destructive inspection settings. Its theoretical foundation—grounded in objective Bayesian inference—and empirical efficacy support its potential adoption into international standardization frameworks.
0 citationsRead paper
MRpro - open PyTorch-based MR reconstruction and processing package
Existing MR image reconstruction tools suffer from limited data operation consistency, inflexible algorithm design, poor deep learning integration, and low reproducibility. To address these limitations, we propose and open-source a modular, PyTorch-based framework for MR reconstruction and processing. The framework introduces a unified differentiable Fourier operator, an extensible phase map simulator, and a data-consistency layer supporting Cartesian, radial, spiral, and other sampling trajectories. It seamlessly integrates model-based optimization (e.g., proximal algorithms) with deep learning components—including learnable regularizers and composable network backbones—to enable motion correction, MR fingerprinting, and quantitative parameter mapping. Standardized data structures, public dataset interfaces, and plug-and-play operator design significantly improve reproducibility and collaborative development efficiency. We validate the framework’s generality and effectiveness across multiple quantitative imaging tasks.
0 citationsRead paper