Deep implicit functions (DIFs) effectively represent shapes by using a neural network to map 3D spatial coordinates to scalar values that encode the shape's geometry, but it is difficult to establish correspondences between shapes directly, limiting their use in medical image registration. The recently presented deformation field-based methods achieve implicit templates learning via template field learning with DIFs and deformation field learning, establishing shape correspondence through deformation fields. Although these approaches enable joint learning of shape representation and shape correspondence, the decoupled optimization for template field and deformation field, caused by the absence of deformation annotations lead to a relatively accurate template field but an underoptimized deformation field. In this paper, we propose a novel implicit template learning framework via a shared hybrid diffeomorphic flow (SHDF), which enables shared optimization for deformation and template, contributing to better deformations and shape representation. Specifically, we formulate the signed distance function (SDF, a type of DIFs) as a one-dimensional (1D) integral, unifying dimensions to match the form used in solving ordinary differential equation (ODE) for deformation field learning. Then, SDF in 1D integral form is integrated seamlessly into the deformation field learning. Using a recurrent learning strategy, we frame shape representations and deformations as solving different initial value problems of the same ODE. We also introduce a global smoothness regularization to handle local optima due to limited outside-of-shape data. Experiments on medical datasets show that SHDF outperforms state-of-the-art methods in shape representation and registration.

Deep learning models for semi-supervised medical image segmentation have achieved unprecedented performance for a wide range of tasks. Despite their high accuracy, these models may however yield predictions that are considered anatomically impossible by clinicians. Moreover, incorporating complex anatomical constraints into standard deep learning frameworks remains challenging due to their non-differentiable nature. To address these limitations, we propose a Constrained Adversarial Training (CAT) method that learns how to produce anatomically plausible segmentations. Unlike approaches focusing solely on accuracy measures like Dice, our method considers complex anatomical constraints like connectivity, convexity, and symmetry which cannot be easily modeled in a loss function. The problem of non-differentiable constraints is solved using a Reinforce algorithm which enables to obtain a gradient for violated constraints. To generate constraint-violating examples on the fly, and thereby obtain useful gradients, our method adopts an adversarial training strategy which modifies training images to maximize the constraint loss, and then updates the network to be robust to these adversarial examples. The proposed method offers a generic and efficient way to add complex segmentation constraints on top of any segmentation network. Experiments on synthetic data and four clinically-relevant datasets demonstrate the effectiveness of our method in terms of segmentation accuracy and anatomical plausibility.

We present an unsupervised domain adaptation method for image segmentation which aligns high-order statistics, computed for the source and target domains, encoding domain-invariant spatial relationships between segmentation classes. Our method first estimates the joint distribution of predictions for pairs of pixels whose relative position corresponds to a given spatial displacement. Domain adaptation is then achieved by aligning the joint distributions of source and target images, computed for a set of displacements. Two enhancements of this method are proposed. The first one uses an efficient multi-scale strategy that enables capturing long-range relationships in the statistics. The second one extends the joint distribution alignment loss to features in intermediate layers of the network by computing their cross-correlation. We test our method on the task of unpaired multi-modal cardiac segmentation using the Multi-Modality Whole Heart Segmentation Challenge dataset and prostate segmentation task where images from two datasets are taken as data in different domains. Our results show the advantages of our method compared to recent approaches for cross-domain image segmentation. Code is available at https://github.com/WangPing521/Domain_adaptation_shape_prior.

Many real-life tasks such as military reconnaissance and traffic monitoring require high-quality images. However, images acquired in foggy or hazy weather pose obstacles to the implementation of these real-life tasks; consequently, image dehazing is an important research problem. To meet the requirements of practical applications, a single image dehazing algorithm has to be able to effectively process real-world hazy images with high computational efficiency. In this article, we present a fast and robust semi-supervised dehazing algorithm named SADnet for practical applications. SADnet utilizes both synthetic datasets and natural hazy images for training, so it has good generalizability for real-world hazy images. Furthermore, considering the uneven distribution of haze in the atmospheric environment, a Channel-Spatial Self-Attention (CSSA) mechanism is presented to enhance the representational power of the proposed SADnet. Extensive experimental results demonstrate that the presented approach achieves good dehazing performances and competitive running times compared with other state-of-the-art image dehazing algorithms.

The contrastive pre-training of a recognition model on a large dataset of unlabeled data often boosts the model’s performance on downstream tasks like image classification. However, in domains such as medical imaging, collecting unlabeled data can be challenging and expensive. In this work, we consider the task of medical image segmentation and adapt contrastive learning with meta-label annotations to scenarios where no additional unlabeled data is available. Meta-labels, such as the location of a 2D slice in a 3D MRI scan, often come for free during the acquisition process. We use these meta-labels to pre-train the image encoder, as well as in a semi-supervised learning step that leverages a reduced set of annotated data. A self-paced learning strategy exploiting the weak annotations is proposed to furtherhelp the learning process and discriminate useful labels from noise. Results on five medical image segmentation datasets show that our approach: i) highly boosts the performance of a model trained on a few scans, ii) outperforms previous contrastive and semi-supervised approaches, and iii) reaches close to the performance of a model trained on the full data.

To address the hazy image degradation problem, we introduce a single image numerical iterative dehazing method based on local physical features. The method involves three components: region division based on haze density, local atmospheric light estimation and transmission map estimation, and recovery of hazy image scene radiance by using an iterative algorithm. Because of the nonuniform haze density within an image, we first employ the affinity propagation (AP) clustering algorithm to divide a hazy image into different haze density regions. Second, to reflect the difference in atmospheric light among regions and avoid the generation of halo artifacts in recovered images, we estimate the local atmospheric light in each region to replace the global atmospheric light and then estimate the transmission via a dark channel prior. Finally, an iterative dehazing algorithm, which can be used to not only further optimize local atmospheric light and transmission but also remove haze completely, is developed based on a physical model. Experimental results illustrate that our method can effectively improve the quality of a foggy image without sacrificing color fidelity and can retain image details sufficiently.

In this paper, a novel single image super-resolution (SR) method based on progressive-iterative approximation is proposed. To preserve textures and clear edges, the image SR reconstruction is treated as an image progressive-iterative fitting procedure and achieved by iterative interpolation. Due to different features in different regions, we first employ the nonsubsampled contourlet transform (NSCT) to divide the image into smooth regions, texture regions, and edges. Then, a hybrid interpolation scheme based on curves and surfaces is proposed, which differs from the traditional surface interpolation methods. Specifically, smooth regions are interpolated by the non-uniform rational basis spline (NURBS) surface geometric iteration. To retain textures, control points are increased, and the progressive-iterative approximation of the NURBS surface is employed to interpolate the texture regions. By considering edges in an image as curve segments that are connected by pixels with dramatic changes, we use NURBS curve progressive-iterative approximation to interpolate the edges, which sharpens the edges and can maintain the image edge structure without jaggy and block artifacts. The experimental results demonstrate that the proposed method significantly outperforms the state-of-the-art methods in terms of both subjective and objective measures.