16:00 | | Introduction |
16:12 | 1365.
| Simultaneous self-supervised reconstruction and denoising for low SNR, sub-sampled training data with Robust SSDU Charles Millard1 and Mark Chiew1,2,3 1Wellcome Centre for Integrative Neuroimaging, University of Oxford, Oxford, United Kingdom, 2Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada, 3Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada Keywords: Image Reconstruction, Image Reconstruction, Deep learning, Self-supervised Motivation: For low SNR training data, such as from low-field scanners, sub-sampled images reconstructed via deep learning can be susceptible to errors due to measurement noise. Goal(s): To evaluate the performance of the proposed Robust Self-Supervised Learning via Data Undersampling (Robust SSDU), which removes corruptions due to aliasing and measurement noise in an entirely self-supervised manner. Approach: On the fastMRI dataset and low-field dataset M4Raw, Robust SSDU was compared with a number of benchmarks including supervised training. Results: Robust SSDU exhibited a substantially higher fidelity image restoration than standard SSDU and sharper reconstructions than competing methods that remove measurement noise. Impact: This study demonstrates that high quality image reconstruction with deep learning is achievable when only sub-sampled, low SNR data is available for training. The proposed method could particularly impact the diagnostic potential of images acquired from low field scanners. |
16:24 | 1366.
| Explicit network noise amplification penalty in loss function for k-space interpolation networks through fast backpropagation Istvan Homolya1, Peter Dawood2, Jannik Stebani3, and Martin Blaimer3 1Molecular and Cellular Imaging, Comprehensive Heart Failure Center, University Hospital Würzburg, Würzburg, Germany, 2Department of Physics, University of Würzburg, Würzburg, Germany, 3Magnetic Resonance and X-ray Imaging Department, Fraunhofer Institute for Integrated Circuits IIS, Division Development Center X-Ray Technology, Würzburg, Germany Keywords: Image Reconstruction, Machine Learning/Artificial Intelligence Motivation: GRAPPA and RAKI optimize purely for data consistency, completely lacking physics-driven or model-based loss terms. Goal(s): Recurrently feed noise amplification information into k-space interpolation networks by penalizing the online computed g-factor. Approach: JAX-implemented GRAPPA and RAKI g-factors were estimated online in each training iteration and incorporated into the optimization as an inherent network noise amplification penalty. Results: Networks including g-factor loss outperformed implementations optimizing only for the data consistency term. Inclusion of g-factor loss terms manifested Tikhonov regularization-like effects on image noise distribution, as revealed by difference maps to the fully sampled gold standard. Impact: Incorporating the penalty of inherent noise amplification into k-space interpolation networks reduces reconstruction noise levels compared to implementation that optimize only for data consistency. G-factor-informed reconstructions manifest Tikhonov regularization-like effects, as revealed by noise distribution on difference maps. |
16:36 | 1367.
| Image space formalism of k-space interpolation networks for analytical expression of noise characteristics Peter Dawood1, Felix Breuer2, István Homolya3, Peter Michael Jakob1, and Martin Blaimer2 1Experimental Physics 5, University of Würzburg, Würzburg, Germany, 2Magnetic Resonance and X-ray Imaging Department, Fraunhofer Institute for Integrated Circuits IIS, Division Development Center X-Ray Technology, Würzburg, Germany, 3Molecular and Cellular Imaging, Comprehensive Heart Failure Center, University Hospital Würzburg, Würzburg, Germany Keywords: Machine Learning/Artificial Intelligence, Parallel Imaging, complex-valued convolutional neural networks, RAKI, GRAPPA, ReLU Motivation: Robust Artificial Neural Networks for k-space Interpolation (RAKI) exhibit superior image reconstructions compared to traditional Parallel Imaging. It is crucial to thoroughly characterize RAKI to gain insights into its functionality and stimulate further enhancements. Goal(s): Exploring how k-space interpolation with convolutional neural networks can be transformed into image domain to obtain an analytical description of noise characteristics. Approach: The nonlinear activation in k-space is expressed as elementwise multiplication. This can be transformed into convolution in image space. Results: The proposed image space formalism yields image reconstructions quasi-equivalent to k-space interpolation. The analytical expression of noise characteristics is in correspondence with Monte Carlo simulations. Impact: We propose an image space
formalism for k-space interpolation with convolutional neural networks. This
enables an analytical expression of the noise characteristics, analogous to
g-factor maps in traditional parallel imaging methods. |
16:48 | 1368.
| Noise-Robust Reconstruction for Accelerated MRI using Contrastive Learning Seonghyuk Kim1, Sung-Hong Park1, and HyunWook Park1 1KAIST, Daejeon, Korea, Republic of Keywords: Image Reconstruction, Image Reconstruction, Noise-robust method Motivation: Deep learning-based accelerated MRI reconstruction methods have shown outstanding performance but do not consider noise. Corruption due to noise may lead to wrong diagnosis in clinical practices. Goal(s): Propose a noise-robust reconstruction method, which reconstructs noise-free full-sampled images from noisy undersampled data. Approach: A noise-robust reconstruction method is proposed using contrastive learning framework consisting of two stages. The first stage extracts feature representations related to the noise level, which is used in the second stage to reconstruct alias-free image. Results: Experiment results show that the proposed method provides robust reconstruction with limited training data, yielding superior image reconstruction compared to other reconstruction methods. Impact: The encoder
trained in the first stage extracts representation features that contain
content-invariant noise level information. Therefore, the trained encoder can
be applied to other downstream tasks with limited amount of training data. |
17:00 | 1369.
| A Positive and Negative Learning based Image Decomposition Network for Phase Unwrapping and Background Removal Lijun Bao1 and Zijun Zhao1 1Department of Electronic Science, Xiamen University, Xiamen, China Keywords: Quantitative Imaging, Quantitative Imaging, phase processing, background removal, deep learning, image decomposition, phase unwrapping Motivation: Phase images contain important information useful in many fields. However, the phase data is often wrapped into a specific range, while background or noise signal in imaging scene may bring significant interference. Goal(s): To obtain the exact information, phase images need an accurate processing that includes the unwrapping and the background removal. Approach: In this paper, we propose a positive and negative learning based image decomposition network (PNnet) to accomplish the phase processing by a single network. Results: Experimental results demonstrate that PNnet can achieve excellent performance and efficient generalization, even for complex wrapping and inhomogeneous background. Impact: Except magnitude
images, phase data in MRI also contain important information that is useful in
many fields and scenarios. This work proposed a SOTA method for phase
processing with high accuracy and excellent performance. |
17:12 | 1370.
| A Phase-Injected Complex Forward-Distortion Approach for Deep Unsupervised Correction of Susceptibility Artifacts in EPI Abdallah Zaid Alkilani1,2, Tolga Çukur1,2,3, and Emine Ulku Saritas1,2 1Department of Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey, 2National Magnetic Resonance Research Center (UMRAM), Bilkent University, Ankara, Turkey, 3Neuroscience Graduate Program, Bilkent University, Ankara, Turkey Keywords: Artifacts, Artifacts, susceptibility artifacts, echo planar imaging, reversed phase-encoding, deep learning, unsupervised learning Motivation: Classical susceptibility-artifact correction methods are impractical in clinical settings given their computational burden. Goal(s): Fast and effective correction of susceptibility artifacts in EPI via physics-driven unsupervised deep learning by utilizing phase-injected complex-valued forward-distortion.
Approach: Previous methods apply distortion correction on magnitude images, potentially yielding suboptimal performance near regions of signal dropout/pileup. We propose a novel model, compFD-Net, employing phase-injected complex forward-distortion that leverages a predicted phase image, additionally to the magnitude image and displacement field estimates, for improved capture of signal dropout/pileup artifacts in EPI images. Results: The proposed model boosts susceptibility-artifact correction performance, notably improving predicted image and field quality. Impact: Robust emulation of signal-dropout/pileup via the complex forward-distortion formulation boosts reliability in unsupervised artifact correction. compFD-Net facilitates rapid and performant correction of susceptibility artifacts in EPI, with possible impact in time-sensitive applications in clinical settings. |
17:24 | 1371.
| Revisiting outer volume subtraction with deep-learning tools for highly-accelerated real-time cine CMR Merve Gulle1, Peter Kellman2, and Mehmet Akcakaya3 1University of Minnesota, Saint Paul, MN, United States, 2National Heart-Lung and Blood Institute, Bethesda, MD, United States, 3University of Minnesota, Minneapolis, MN, United States Keywords: Image Reconstruction, Cardiovascular, real-time, cardiac cine, heart, outer volume subtraction Motivation: Real-time cine CMR provides a free-breathing ECG-free approach for heart function assessment. Nevertheless, commercially available real-time cine CMR methods without temporal regularization have limited acceleration and spatio-temporal resolutions. Goal(s): Use deep learning (DL) to remove extra-cardiac volume that aliases into the heart and improve acceleration rates for real-time cine CMR using only spatial regularization. Approach: We characterize pseudo-periodic ghosting artifacts arising from cardiac motion in time-interleaved sequences, then use DL to detect and remove them. This is followed by self-supervised physics-driven DL reconstruction. Results: Proposed technique effectively estimates and removes background signal, leading to substantial image quality improvement. Impact: We characterize and use deep learning (DL) to estimate pseudo-periodic ghosting artifacts arising from cardiac motion in time-interleaved real-time cine sequences. Background removal followed by physics-driven DL reconstruction substantially improves reconstruction at nominal R=8 for higher spatio-temporal resolution acquisitions. |
17:36 | 1372.
| 3D Free-Breathing Ungated Spiral bSSFP Functional Cardiac Imaging Using a Deep Image Prior Jesse Ian Hamilton1,2, Gastao Lima da Cruz1, and Nicole Seiberlich1,2 1Radiology, University of Michigan, Ann Arbor, MI, United States, 2Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States Keywords: Machine Learning/Artificial Intelligence, Cardiovascular, deep learning; spiral; real-time CMR Motivation: Real-time imaging methods are useful for patients with limited breathhold capacity or arrhythmias, but are typically limited to 2D scans that prevent evaluation of wall motion in 3D over the heart. Goal(s): The goal of this project is to develop a technique for 3D real-time (free-breathing ungated) cine imaging. Approach: The proposed method combines a highly undersampled 3D stack-of-spirals trajectory with a deep image prior reconstruction, which does not require ground truth training data. Results: Real-time 3D imaging is demonstrated in healthy subjects with temporal resolutions of 36ms per volume at 1.5T and 58ms per volume at 0.55T. Impact: Real-time 3D imaging could enable streamlined cardiac MRI exams, with whole-heart 3D cine images obtained in 10s without breathholds or gating. This technique may also simplify quantification compared to 2D real-time methods, since motion is synchronized over all partitions. |
17:48 | 1373.
| Locally Adaptive Low Rank Regularization with Collaborative Data Selection for Arterial Spin Labeling MRI Denoising Hangfan Liu1, Bo Li1, Yiran Li1, John A Detre2, and Ze Wang1 1University of Maryland School of Medicine, Baltimore, MD, United States, 2Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, United States Keywords: Sparse & Low-Rank Models, Arterial spin labelling, Denoising, MRI Motivation: Address the challenge of low SNR in arterial spin labeling (ASL) MRI that hinders its clinical and research potential. Goal(s): Develop an advanced ASL denoising algorithm that enhances image quality and overcomes limitations in ASL due to low SNR. Approach: Propose a Locally Adaptive low rank regularization with Collaborative data Selection (LACS) scheme that utilizes the structural characteristics of ASL images for collaborative data selection to improve low-rank modeling. The proposed low-rank regularization fundamentally performs locally adaptive PCA without explicit training. Results: Using a single ASL image pair, LACS significantly outperformed state-of-the-art MRI denoising methods and the standard pipeline. Impact: The proposed scheme has the potential to benefit
researchers, clinicians, and patients by setting a new benchmark for ASL MRI
denoising. It opens doors to exploring ASL's full clinical potential and offers
opportunities for innovative research. |