State-of-the-art image reconstruction often relies on complex, abundantly parameterized deep architectures. We propose an alternative: a data-driven reconstruction method inspired by the classic Tikhonov regularization. Our approach iteratively refines intermediate reconstructions by solving a sequence of quadratic problems. These updates have two key components: (i) learned filters to extract salient image features; and (ii) an attention mechanism that locally adjusts the penalty of the filter responses. Our method matches leading plug-and-play and learned regularizer approaches in performance while offering interpretability, robustness, and convergent behavior. In effect, we bridge traditional regularization and deep learning with a principled reconstruction approach.

Abstract. Image restoration tasks often admit a wide range of solutions that are equally consistent with the observed data. Quantifying this uncertainty is crucial for the robust interpretation of restored images and their reliable use in science and decision-making. We consider the classification of images reconstructed from noisy and corrupted measurements, with special attention to the quantification of uncertainty in the delivered classification results. We address this problem by constructing a Bayesian statistical approach that combines learning-based image priors and image classifiers with explicit image observation models specified during inference. Following the manifold hypothesis, we assume that the image prior is supported on a submanifold of the ambient space—which we learn from uncorrupted training data using a variational autoencoder—and use as a classifier a support vector machine operating in this low-dimensional representation. The observation model is incorporated during inference time through its likelihood function. Bayesian computation is then efficiently performed by leveraging variants of the unadjusted Langevin algorithm that operate directly on the submanifold and are robust to multimodality. This results in a robust image classification method that provides uncertainty estimates that are provably well-posed, derived from Bayesian decision theory rigorously and transparently, and which incorporate physical and instrumental aspects of the data acquisition process through Bayes’ theorem. We demonstrate the effectiveness of the proposed approach through experiments on the MNIST and CelebA datasets, where we achieve accurate uncertainty estimates, as measured by the expected calibration error, even in challenging image restoration problems with significant inherent uncertainty.

In cardiac cine magnetic resonance imaging (MRI), the heart is repeatedly imaged at numerous time points during the cardiac cycle. Frequently, the temporal evolution of a certain region of interest such as the ventricles or the atria is highly relevant for clinical diagnosis. In this paper, we devise a novel approach that allows for an automatized propagation of an arbitrary region of interest (ROI) along the cardiac cycle from respective annotated ROIs provided by medical experts at two different points in time, most frequently at the end-systolic (ES) and the end-diastolic (ED) cardiac phases. At its core, a 3D TV-$\boldsymbol {L^{1}}$-based optical flow algorithm computes the apparent motion of consecutive MRI images in forward and backward directions. Subsequently, the given terminal annotated masks are propagated by this bidirectional optical flow in 3D, which results, however, in improper initial estimates of the segmentation masks due to numerical inaccuracies. These initially propagated segmentation masks are then refined by a 3D U-Net-based convolutional neural network (CNN), which was trained to enforce consistency with the forward and backward warped masks using a novel loss function. Moreover, a penalization term in the loss function controls large deviations from the initial segmentation masks. This method is benchmarked both on a new dataset with annotated single ventricles containing patients with severe heart diseases and on a publicly available dataset with different annotated ROIs. We emphasize that our novel loss function enables fine-tuning the CNN on a single patient, thereby yielding state-of-the-art results along the complete cardiac cycle.

Video denoising is a fundamental problem in numerous computer vision applications. State-of-the-art attention-based denoising methods typically yield good results, but require vast amounts of GPU memory and usually suffer from very long computation times. Especially in the field of restoring digitized high-resolution historic films, these techniques are not applicable in practice. To overcome these issues, we introduce a lightweight video denoising network that combines efficient axial-coronal-sagittal (ACS) convolutions with a novel shifted window attention formulation (ASwin), which is based on the memory-efficient aggregation of self- and cross-attention across video frames. We numerically validate the performance and efficiency of our approach on synthetic Gaussian noise. Moreover, we train our network as a general-purpose blind denoising model for real-world videos, using a realistic noise synthesis pipeline to generate clean-noisy video pairs. A user study and non-reference quality assessment prove that our method outperforms the state-of-the-art on real-world historic videos in terms of denoising performance and temporal consistency.

Various problems in computer vision and medical imaging can be cast as inverse problems. A frequent method for solving inverse problems is the variational approach, which amounts to minimizing an energy composed of a data fidelity term and a regularizer. Classically, handcrafted regularizers are used, which are commonly outperformed by state-of-the-art deep learning approaches. In this work, we combine the variational formulation of inverse problems with deep learning by introducing the data-driven general-purpose total deep variation regularizer. In its core, a convolutional neural network extracts local features on multiple scales and in successive blocks. This combination allows for a rigorous mathematical analysis including an optimal control formulation of the training problem in a mean-field setting and a stability analysis with respect to the initial values and the parameters of the regularizer. In addition, we experimentally verify the robustness against adversarial attacks and numerically derive upper bounds for the generalization error. Finally, we achieve state-of-the-art results for several imaging tasks.

Recent deep learning approaches focus on improving quantitative scores of dedicated benchmarks, and therefore only reduce the observation-related (aleatoric) uncertainty. However, the model-immanent (epistemic) uncertainty is less frequently systematically analyzed. In this work, we introduce a Bayesian variational framework to quantify the epistemic uncertainty. To this end, we solve the linear inverse problem of undersampled MRI reconstruction in a variational setting. The associated energy functional is composed of a data fidelity term and the total deep variation (TDV) as a learned parametric regularizer. To estimate the epistemic uncertainty we draw the parameters of the TDV regularizer from a multivariate Gaussian distribution, whose mean and covariance matrix are learned in a stochastic optimal control problem. In several numerical experiments, we demonstrate that our approach yields competitive results for undersampled MRI reconstruction. Moreover, we can accurately quantify the pixelwise epistemic uncertainty, which can serve radiologists as an additional resource to visualize reconstruction reliability.

We propose a novel learning-based framework for image reconstruction particularly designed for training without ground truth data, which has three major building blocks: energy-based learning, a patch-based Wasserstein loss functional, and shared prior learning. In energy-based learning, the parameters of an energy functional composed of a learned data fidelity term and a data-driven regularizer are computed in a mean-field optimal control problem. In the absence of ground truth data, we change the loss functional to a patch-based Wasserstein functional, in which local statistics of the output images are compared to uncorrupted reference patches. Finally, in shared prior learning, both aforementioned optimal control problems are optimized simultaneously with shared learned parameters of the regularizer to further enhance unsupervised image reconstruction. We derive several time discretization schemes of the gradient flow and verify their consistency in terms of Mosco convergence. In numerous numerical experiments, we demonstrate that the proposed method generates state-of-the-art results for various image reconstruction applications---even if no ground truth images are available for training.

Diverse inverse problems in imaging can be cast as variational problems composed of a task-specific data fidelity term and a regularization term. In this paper, we propose a novel learnable general-purpose regularizer exploiting recent architectural design patterns from deep learning. We cast the learning problem as a discrete sampled optimal control problem, for which we derive the adjoint state equations and an optimality condition. By exploiting the variational structure of our approach, we perform a sensitivity analysis with respect to the learned parameters obtained from different training datasets. Moreover, we carry out a nonlinear eigenfunction analysis, which reveals interesting properties of the learned regularizer. We show state-of-the-art performance for classical image restoration and medical image reconstruction problems.

In this work, we propose a learning-based variational network (VN) approach for reconstruction of low-dose 3D computed tomography data. We focus on two methods to decrease the radiation dose: (1) x-ray tube current reduction, which reduces the signal-to-noise ratio, and (2) x-ray beam interruption, which undersamples data and results in images with aliasing artifacts. While the learned VN denoises the current-reduced images in the first case, it reconstructs the undersampled data in the second case. Different VNs for denoising and reconstruction are trained on a single clinical 3D abdominal data set. The VNs are compared against state-of-the-art model-based denoising and sparse reconstruction techniques on a different clinical abdominal 3D data set with 4-fold dose reduction. Our results suggest that the proposed VNs enable higher radiation dose reductions and/or increase the image quality for a given dose.