Three-dimensional coronary magnetic resonance angiography (CMRA) requires reconstruction algorithms that can significantly suppress the artifacts encountered in heavily undersampled acquisitions. While unrolling-based deep reconstruction methods have achieved state-of-the-art performance on 2D image reconstruction, their application in 3D reconstruction is hindered by the large amount of memory required to train an unrolled network. In this study, we propose a memoryefficient deep compressed sensing method that employs a sparsifying transform based on a pre-trained artifact estimation network. The artifact image estimated by a welltrained network is expected to be sparse when the input image is artifact-free and less sparse when the input image has artifacts. Thus, the artifact estimation network can be used as an inherent sparsifying transform. The proposed method, De-Aliasing Regularization-based Compressed Sensing (DARCS), was compared with a patch-based lowrank method, de-aliasing generative adversarial network (DAGAN), 3D model-based deep learning (MoDL), plug-and-play, and AI-assisted compressed sensing (AI-CS) in terms of 3D CMRA acceleration. The results demonstrate that DARCS surpasses the reconstruction quality of the comparison methods, by approximately 2 dB in peak signal-to-noise ratio (PSNR). Furthermore, the proposed method generalizes well to different undersampling rates, patterns, and noise levels, with a memory usage of only 63% of that needed by 3D MoDL. In conclusion, DARCS improves reconstruction quality for 3D CMRA with reduced memory burden.
15 October 2024 in IEEE Transactions on Biomedical Engineering
Objective: To propose a 3D nonrigid registration method that accurately estimates the 3D displacement field from artifact-corrupted Coronary Magnetic Resonance Angiography (CMRA) images. Methods: We developed a novel registration framework for registration of artifact-corrupted images based on a 3D U-Net initializer and a deep unrolling network. By leveraging a supervised learning framework with training labels estimated from fully-sampled images, the unrolling network learns a task-specific motion prior which reduces motion estimation biases caused by undersampling artifacts in the source images. We evaluated the proposed method, UNROLL, against an iterative Free-Form Deformation (FFD) registration method and a recently proposed Respiratory Motion Estimation network (RespME-net) for 6-fold (in-distribution) and 11-fold (out-of-distribution) accelerated CMRA. Results: Compared to the baseline methods, UNROLL improved both the accuracy of motion estimation and the quality of motion-compensated CMRA reconstruction at 6-fold acceleration. Furthermore, even at 11-fold acceleration, which was not included during training, UNROLL still generated more accurate displacement fields than the baseline methods. The computational time of UNROLL for the whole 3D volume was only 2 seconds. Conclusion: By incorporating a learned respiratory motion prior, the proposed method achieves highly accurate motion estimation despite the large acceleration. Significance: This work introduces a fast and accurate method to estimate the displacement field from low-quality source images. It has the potential to significantly improve the quality of motion-compensated reconstruction for highly accelerated 3D CMRA.
Existing works have extensively studied adversarial examples, which are minimal perturbations that can mislead the output of deep neural networks (DNNs) while remaining imperceptible to humans. However, in this work, we reveal the existence of a harmless perturbation space, in which perturbations drawn from this space, regardless of their magnitudes, leave the network output unchanged when applied to inputs. Essentially, the harmless perturbation space emerges from the usage of non-injective functions (linear or non-linear layers) within DNNs, enabling multiple distinct inputs to be mapped to the same output. For linear layers with input dimensions exceeding output dimensions, any linear combination of the orthogonal bases of the nullspace of the parameter consistently yields no change in their output. For non-linear layers, the harmless perturbation space may expand, depending on the properties of the layers and input samples. Inspired by this property of DNNs, we solve for a family of general perturbation spaces that are redundant for the DNN's decision, and can be used to hide sensitive data and serve as a means of model identification. Our work highlights the distinctive robustness of DNNs (i.e., consistency under large magnitude perturbations) in contrast to adversarial examples (vulnerability for small noises).
Diffusion is well known for its success in improving retrieval performance by exploiting the local structure of data distribution. Some recent works have focused on improving its efficiency by shifting the computing burden offline. However, we find that efficient offline diffusion handles continuously updating datasets with difficulty, which directly hinders its application in the real world. Unlike previous methods that apply diffusion to the entire gallery, we introduce an anchor graph to serve as an agent of the complete gallery graph. By doing that, we empower diffusion with the ability of retrieving newly added images at acceptable computational cost. We demonstrate that our proposed method is a good approximation of diffusion featuring fast online search speed and the ability of handling growing data. Moreover, experiments on benchmark datasets show that the proposed method outperforms the state of the art by a large margin with proper parameter settings.