Score-based generative models have demonstrated highly promising results for medical image reconstruction tasks in magnetic resonance imaging or computed tomography. However, their application to Positron Emission Tomography (PET) is still largely unexplored. PET involves a variety of challenges, including Poisson noise with high variance and wide dynamic range. To address these we propose several PET-specific adaptations score-based models. The proposed framework developed both 2D 3D PET....

Denoising diffusion models have emerged as the go-to generative framework for solving inverse problems in imaging. A critical concern regarding these is their performance on out-of-distribution tasks, which remains an under-explored challenge. Using a model dataset, realistic reconstructions can be generated, but with hallucinating image features that are uniquely present training dataset. To address this discrepancy and improve reconstruction accuracy, we introduce novel test-time...

Deep learning-based image reconstruction approaches have demonstrated impressive empirical performance in many imaging modalities. These usually require a large amount of high-quality paired training data, which is often not available medical imaging. To circumvent this issue we develop novel unsupervised knowledge-transfer paradigm for learned within Bayesian framework. The proposed approach learns network two phases. first phase trains with set ordered pairs comprising ground truth images...

Recent advances in reconstruction methods for inverse problems leverage powerful data-driven models, e.g., deep neural networks. These techniques have demonstrated state-of-the-art performances several imaging tasks, but they often do not provide uncertainty on the obtained reconstruction. In this work, we develop a scalable, data-driven, knowledge-aided computational framework to quantify model via Bayesian The approach builds on, and extends gradient descent, recently developed greedy...

Deep image prior (DIP) was recently introduced as an effective unsupervised approach for restoration tasks. DIP represents the to be recovered output of a deep convolutional neural network, and learns network's parameters such that matches corrupted observation. Despite its impressive reconstructive properties, is slow when compared supervisedly learned, or traditional reconstruction techniques. To address computational challenge, we bestow with two-stage learning paradigm: (i) perform...

Learned image reconstruction techniques using deep neural networks have recently gained popularity and delivered promising empirical results. However, most approaches focus on one single recovery for each observation, thus neglect information uncertainty. In this work, we develop a novel computational framework that approximates the posterior distribution of unknown at query observation. The proposed is very flexible: it handles implicit noise models priors, incorporates data formation...

Multi-fidelity Bayesian Optimization (MFBO) is a promising framework to speed up materials and molecular discovery as sources of information different accuracies are at hand increasing cost. Despite its potential use in chemical tasks, there lack systematic evaluation the many parameters playing role MFBO. In this work, we provide guidelines recommendations decide when MFBO experimental settings. We investigate methods applied molecules problems. First, test two families acquisition...

Existing deep-learning based tomographic image reconstruction methods do not provide accurate estimates of uncertainty, hindering their real-world deployment. This paper develops a method, termed as the linearised deep prior (DIP), to estimate uncertainty associated with reconstructions produced by DIP total variation regularisation (TV). Specifically, we endow conjugate Gaussian-linear model type error-bars computed from local linearisation neural network around its optimised parameters. To...

Image reconstruction methods based on deep neural networks have shown outstanding performance, equalling or exceeding the state-of-the-art results of conventional approaches, but often do not provide uncertainty information about reconstruction. In this work we propose a scalable and efficient framework to simultaneously quantify aleatoric epistemic uncertainties in learned iterative image We build Bayesian gradient descent method for quantifying uncertainty, incorporate heteroscedastic...

Denoising diffusion models have emerged as the go-to framework for solving inverse problems in imaging. A critical concern regarding these is their performance on out-of-distribution (OOD) tasks, which remains an under-explored challenge. Realistic reconstructions inconsistent with measured data can be generated, hallucinating image features that are uniquely present training dataset. To simultaneously enforce data-consistency and leverage data-driven priors, we introduce a novel sampling...

We investigate adaptive design based on a single sparse pilot scan for generating effective scanning strategies computed tomography reconstruction. propose novel approach using the linearised deep image prior. It allows incorporating information from measurements into angle selection criteria, while maintaining tractability of conjugate Gaussian-linear model. On synthetically generated dataset with preferential directions, DIP reducing number scans by up to 30% relative an equidistant baseline.

Recent advances in reconstruction methods for inverse problems leverage powerful data-driven models, e.g., deep neural networks. These techniques have demonstrated state-of-the-art performances several imaging tasks, but they often do not provide uncertainty on the obtained reconstruction. In this work, we develop a scalable, data-driven, knowledge-aided computational framework to quantify model via Bayesian The approach builds on, and extends gradient descent, recently developed greedy...

Generative modelling paradigms based on denoising diffusion processes have emerged as a leading candidate for conditional sampling in inverse problems. In many real-world applications, we often access to large, expensively trained unconditional models, which aim exploit improving sampling. Most recent approaches are motivated heuristically and lack unifying framework, obscuring connections between them. Further, they suffer from issues such being very sensitive hyperparameters, expensive...

Motivation: Reducing the number of parameters needed to represent and reconstruct parallel MRI measurements. Goal(s): Reconstruct measurements with coordinate-transformed Gaussian functions (blobs) where forward model is formulated directly. We term this MR-blob. Approach: MR-blob directly represents measurements; coil sensitivities are modelled as isotropic Gaussians image represented by blobs. Results: Noisy, undersampled simulations Shepp-Logan phantom reconstructed a pixelised image,...

The linearised Laplace method for estimating model uncertainty has received renewed attention in the Bayesian deep learning community. provides reliable error bars and admits a closed-form expression evidence, allowing scalable selection of hyperparameters. In this work, we examine assumptions behind method, particularly conjunction with selection. We show that these interact poorly some now-standard tools learning--stochastic approximation methods normalisation layers--and make...

Widespread adoption of deep learning in medical imaging has been hampered, part, due to a lack integration with clinically applicable software. In this work, we establish direct connection between an established PET reconstruction suite, SIRF, and PyTorch. This allows for advanced methodologies be deployed on clinical data unsupervised approach: the Deep Image Prior. Results show consistent quality metrics DIP comparison OSMAP.

Deep learning-based image reconstruction approaches have demonstrated impressive empirical performance in many imaging modalities. These usually require a large amount of high-quality paired training data, which is often not available medical imaging. To circumvent this issue we develop novel unsupervised knowledge-transfer paradigm for learned within Bayesian framework. The proposed approach learns network two phases. first phase trains with set ordered pairs comprising ground truth images...

Deep learning has been widely used for solving image reconstruction tasks but its deployability held back due to the shortage of high-quality training data. Unsupervised methods, such as deep prior (DIP), naturally fill this gap, bring a host new issues: susceptibility overfitting lack robust early stopping strategies and unstable convergence. We present novel approach tackle these issues by restricting DIP optimisation sparse linear subspace parameters, employing synergy dimensionality...

The deep image prior (DIP) is a well-established unsupervised learning method for reconstruction; yet it far from being flawless. DIP overfits to noise if not early stopped, or optimized via regularized objective. We build on the fine-tuning of pretrained DIP, by adopting novel strategy that restricts adaptation singular values. proposed SVD-DIP uses ad hoc convolutional layers whose parameters are decomposed value decomposition. Optimizing then solely consists in values, while keeping left...

Score-based generative models have demonstrated highly promising results for medical image reconstruction tasks in magnetic resonance imaging or computed tomography. However, their application to Positron Emission Tomography (PET) is still largely unexplored. PET involves a variety of challenges, including Poisson noise with high variance and wide dynamic range. To address these we propose several PET-specific adaptations score-based models. The proposed framework developed both 2D 3D PET....