Deep learning has recently been intensively studied in the context of image compressive sensing (CS) to discover and represent complicated image structures. These approaches, however, either suffer from nonflexibility for an arbitrary sampling ratio or lack an explicit deep-learned regularization term. This paper aims to solve the CS reconstruction problem by combining the deep-learned regularization term and proximal operator. We first introduce a regularization term using a carefully designed residual-regressive net, which can measure the distance between a corrupted image and a clean image set and accurately identify to which subspace the corrupted image belongs. We then address a proximal operator with a tailored dilated residual channel attention net, which enables the learned proximal operator to map the distorted image into the clean image set. We adopt an adaptive proximal selection strategy to embed the network into the loop of the CS image reconstruction algorithm. Moreover, a self-ensemble strategy is presented to improve CS recovery performance. We further utilize state evolution to analyze the effectiveness of the designed networks. Extensive experiments also demonstrate that our method can yield superior accurate reconstruction (PSNR gain over 1 dB) compared to other competing approaches while achieving the current state-of-the-art image CS reconstruction performance. The test code is available at https://github.com/zjut-gwl/CSDRCANet.

The retinogeniculate visual pathway (RGVP) plays an important role in the visual system. Diffusion MRI-based tractography has been successfully used to identify RGVP. However, challenges of RGVP tractography remain because of its highly curved path and intricate anatomical environment. One of the key challenges is the large false-positive fibers generated from RGVP tractography that requires the labor costs to hand-draw ROIs for fiber filtering. Therefore, we presented a pipeline to enable automated RGVP identification in dMRI tractography. First, we generated a tractography-based RGVP atlas. Herein, the multi-fiber unscented Kalman filter tractography was performed using high-resolution data from 50 subjects. Then, we transformed the 50 tractography cases into a common space and implemented data-driven fiber clustering to group the neighboring fibers with similar trajectories into one cluster. Two experienced anatomists were responsible for RGVP annotation in the tractography atlas. Second, the high-dimensional RGVP atlas was applied to identify subject-specific RGVP in testing datasets and two patients with different scanning parameters. Experimental results showed that our automatic identification results have ideal colocalization with expert manual identification in terms of hausdorff distance, fiber distance, and visualization. Therefore, the proposed method provides an efficient tool for analyzing large-scale datasets in vision-related neuroscience research.

Video snapshot compressive imaging (SCI) samples 3D high-speed video frames with temporally varying spatial modulation and compresses them into a single 2D measurement, and the SCI reconstruction algorithm aims to recover the original high-speed frames from the measurement. However, conventional video SCI systems encounter challenges when raising the frame rate of the reconstructed video. To achieve higher frame rates, the modulation mask's rate must be increased, which in turn leads to an increase in the associated hardware expenses. In this paper, we propose a deep unfolding-based reconstruction framework with optical flow for video SCI. The framework recovers both observed and unobserved frames from measurements, resulting in increased video frame rate. To estimate the optical flow, we transform the video features of the network into optical flow features during the iteration. Then, we design a deep denoiser and an optical flow-based motion regularizer combined with the voxels of coarse reconstructed frames to update the observed and unobserved frames. To improve the performance, we employ group convolution in the network and fuse the optical flow information from different phases to reduce the information loss. We further extend the proposed deep unfolding framework to the reconstruction of color SCI videos. Extensive experiments on benchmark data and real data prove that our proposed method has state-of-the-art reconstruction performance and visual effects

Objective: This study aims to accurately extract the visual pathway (VP) from multimodal MR images while minimizing reliance on extensive labeled data and enhancing extraction performance. Method: We propose a novel approach that incorporates a Modality-Relevant Feature Extraction Module (MRFEM) to effectively extract essential features from T1-weighted and fractional anisotropy (FA) images. Additionally, we implement a mean-teacher model integrated with dual uncertainty-aware ambiguity identification (DUAI) to enhance the reliability of the VP extraction process. Results: Experiments conducted on the Human Connectome Project (HCP) and Multi-Shell Diffusion MRI (MDM) datasets demonstrate that our method reduces annotation efforts by at least one-third compared to fully supervised techniques while achieving superior extraction performance over six state-of-the-art semi-supervised methods. Conclusion: The proposed label-efficient approach alleviates the burdens of manual annotation and enhances the accuracy of multimodal MRI-based VP extraction. Significance: This work contributes to the field of medical imaging by facilitating more efficient and accurate visual pathway extraction, thereby improving the analysis and understanding of complex brain structures with reduced reliance on expert annotation.