Multi-Coil MRI Reconstruction Challenge—Assessing Brain MRI Reconstruction Models and Their Generalizability to Varying Coil Configurations

Beauferris, Youssef; Teuwen, Jonas; Karkalousos, Dimitrios; Moriakov, Nikita; Caan, Matthan W.A.; Yiasemis, George; Rodrigues, Lívia; Lopes, Alexandre Gomes; Pedrini, Hélio; Rittner, Letícia; Dannecker, Maik; Studenyak, Viktor; Gröger, Fabian; Vyas, Devendra; Faghihroohi, Shahrooz; Jethi, Amrit Kumar; Raju, Jaya Chandra; Sivaprakasam, Mohanasankar; Lasby, Mike; Nogovitsyn, Nikita; Loos, Wallace; Frayne, Richard; Souza, Roberto

doi:10.3389/fnins.2022.919186

Cited by 19 publications

(8 citation statements)

References 55 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For the main experiments on downsampled data, it takes U-net+FBP approximately 10 s to reconstruct volumes for both geometries, while Uformer+FBP takes approximately 50 s. PDHG takes 14 min to reconstruct a small FOV volume and 18 min to reconstruct a large FOV volume 8 .…”

Section: Inference Times and Evaluation Metricsmentioning

confidence: 99%

“…In recent years, reconstruction methods based on deep learning have attracted a lot of interest in the community and demonstrated very promising results in public reconstruction challenges. For example, in the recent MRI reconstruction challenges 7,8 deep learning methods have strongly outperformed the classical baselines. Generally speaking, any medical image reconstruction task can be viewed as an abstract inverse problem for a suitable forward operator, and different approaches have been proposed in the literature for solving such problems with deep learning 9 .…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

End‐to‐end memory‐efficient reconstruction for cone beam CT

Moriakov,

Sonke,

Teuwen

2023

Medical Physics

Self Cite

View full text Add to dashboard Cite

BackgroundCone beam computed tomography (CBCT) plays an important role in many medical fields nowadays. Unfortunately, the potential of this imaging modality is hampered by lower image quality compared to the conventional CT, and producing accurate reconstructions remains challenging. A lot of recent research has been directed towards reconstruction methods relying on deep learning, which have shown great promise for various imaging modalities. However, practical application of deep learning to CBCT reconstruction is complicated by several issues, such as exceedingly high memory costs of deep learning methods when working with fully 3D data. Additionally, deep learning methods proposed in the literature are often trained and evaluated only on data from a specific region of interest, thus raising concerns about possible lack of generalization to other regions.PurposeIn this work, we aim to address these limitations and propose LIRE: a learned invertible primal‐dual iterative scheme for CBCT reconstruction.MethodsLIRE is a learned invertible primal‐dual iterative scheme for CBCT reconstruction, wherein we employ a U‐Net architecture in each primal block and a residual convolutional neural network (CNN) architecture in each dual block. Memory requirements of the network are substantially reduced while preserving its expressive power through a combination of invertible residual primal‐dual blocks and patch‐wise computations inside each of the blocks during both forward and backward pass. These techniques enable us to train on data with isotropic 2 mm voxel spacing, clinically‐relevant projection count and detector panel resolution on current hardware with 24 GB video random access memory (VRAM).ResultsTwo LIRE models for small and for large field‐of‐view (FoV) setting were trained and validated on a set of 260 + 22 thorax CT scans and tested using a set of 142 thorax CT scans plus an out‐of‐distribution dataset of 79 head and neck CT scans. For both settings, our method surpasses the classical methods and the deep learning baselines on both test sets. On the thorax CT set, our method achieves peak signal‐to‐noise ratio (PSNR) of 33.84 ± 2.28 for the small FoV setting and 35.14 ± 2.69 for the large FoV setting; U‐Net baseline achieves PSNR of 33.08 ± 1.75 and 34.29 ± 2.71 respectively. On the head and neck CT set, our method achieves PSNR of 39.35 ± 1.75 for the small FoV setting and 41.21 ± 1.41 for the large FoV setting; U‐Net baseline achieves PSNR of 33.08 ± 1.75 and 34.29 ± 2.71 respectively. Additionally, we demonstrate that LIRE can be finetuned to reconstruct high‐resolution CBCT data with the same geometry but 1 mm voxel spacing and higher detector panel resolution, where it outperforms the U‐Net baseline as well.ConclusionsLearned invertible primal‐dual schemes with additional memory optimizations can be trained to reconstruct CBCT volumes directly from the projection data with clinically‐relevant geometry and resolution. Such methods can offer better reconstruction quality and generalization compared to classical deep learning baselines.

show abstract

Section: Inference Times and Evaluation Metricsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

End‐to‐end memory‐efficient reconstruction for cone beam CT

Moriakov,

Sonke,

Teuwen

2023

Medical Physics

Self Cite

View full text Add to dashboard Cite

show abstract

“…For this challenge, reconstruction using one quarter of the typically acquired data was selected as the objective inspired by the results of magnetic resonance imaging (MRI) reconstruction challenges(Matthew J. Muckley et al, 2021; Youssef Beauferris et al, 2022) and the quality limitations of GABA-edited MRS data.…”

Section: Introductionmentioning

confidence: 99%

Advancing GABA-edited MRS Research through a Reconstruction Challenge

Berto,

Bugler,

Dias

et al. 2023

Preprint

Self Cite

View full text Add to dashboard Cite

Purpose To create a benchmark for the comparison of machine learning-based Gamma-Aminobutyric Acid (GABA)-edited Magnetic Resonance Spectroscopy (MRS) reconstruction models using one quarter of the transients typically acquired during a complete scan. Methods The Edited-MRS reconstruction challenge had three tracks with the purpose of evaluating machine learning models trained to reconstruct simulated (Track 1), homogeneous in vivo (Track 2), and heterogeneous in vivo (Track 3) GABA-edited MRS data. Four quantitative metrics were used to evaluate the results: mean squared error (MSE), signal-to-noise ratio (SNR), linewidth, and a shape score metric that we proposed. Challenge participants were given three months to create, train and submit their models. Challenge organizers provided open access to a baseline U-NET model for initial comparison, as well as simulated data, in vivo data, and tutorials and guides for adding synthetic noise to the simulations. Results The most successful approach for Track 1 simulated data was a covariance matrix convolutional neural network model, while for Track 2 and Track 3 in vivo data, a vision transformer model operating on a spectrogram representation of the data achieved the most success. Deep learning (DL) based reconstructions with reduced transients achieved equivalent or better SNR, linewidth and fit error as conventional reconstructions with the full amount of transients. However, some DL models also showed the ability to optimize the linewidth and SNR values without actually improving overall spectral quality, pointing to the need for more robust metrics. Conclusion The edited-MRS reconstruction challenge showed that the top performing DL based edited-MRS reconstruction pipelines can obtain with a reduced number of transients equivalent metrics to conventional reconstruction pipelines using the full amount of transients. The proposed metric shape score was positively correlated with challenge track outcome indicating that it is well-suited to evaluate spectral quality.

show abstract

“…The j th sub-band of an image I is approximated by blurring I using a zero-meaned Gaussian kernel with σ 2 ∝ 2 J−j and downsampling the image by factor 2 j . σ 2 N is a parameter of the vision model and was chosen as 0.4 for MRI images 190 . The VIF is bounded by 0, but can reach values greater than one if the reconstructed image shows less noise or improved contrast compared to the target image.…”

Section: Discussionmentioning

confidence: 99%

“…The Calgary-Campinas multi-channel MR dataset 190 is an open, 2D brain MRI dataset providing 12-channel k-space of 167 volunteers, acquired using a T 1 -weighted gradient-recalled echo sequence (TE=2.6-3.1 ms, TR=6.3-7.4 ms, TI=400-650 ms). The MRI was acquired using a 3T system at 1 mm 3 isotropic resolution.…”

Section: Datasetsmentioning

confidence: 99%

Real-time MRI-guided radiotherapy with deep learning

Terpstra

View full text Add to dashboard Cite

Radiotherapy is a common treatment for cancer, which uses radiation to destroy cancer cells. The goal is to deliver a high dose to the tumor without irradiating healthy tissue, but the organ motion due to, for example, respiration makes this difficult. You can see what you treat by irradiating patients with hybrid MRI/radiation devices (MRI-Linac). With real-time adaptive MRI-guided radiotherapy (MRIgRT), the radiation beam is continuously adjusted to compensate for respiratory motion. A major challenge is the speed of MRI acquisitions, and it can take several minutes to acquire high-quality, high-resolution three-dimensional MRI. To compensate for respiratory motion, MRI must be acquired, reconstructed into an image, and the motion computation must be completed within 200 milliseconds. Deep learning is a promising solution due to its speed and quality when minimal MRI data is available. In this thesis we investigated the applicability of deep learning for MRI-guided radiotherapy. The research shows that neural networks can determine three-dimensional motion within 200 milliseconds with an error of less than 1 millimeter while using 25x fewer data than conventional MRI scans. Moreover, we discovered that traditional training of neural networks causes systematic errors in motion models. In particular, minimizing the L2 norm in the complex domain results in an underestimation of the recovered magnitude. This is problematic for radiotherapeutic applications as this will result in systematic errors. To resolve this, we have proposed a new loss function to train neural networks that is fully symmetric with respect to the recovered magnitude and phase. Finally, we have investigated the acceleration of the acquisition and reconstruction of the so-called 4D-MRI, which is important for MRI-guided radiotherapy to quantify the motion magnitude but takes a long time to acquire. Deep learning can potentially reconstruct accelerated 4D-MRI acquisitions with high quality, but training these models is challenging due to the high-dimensional nature of 4D-MRI. This makes it difficult to capture all spatio-temporal information in a computationally feasible way. We have proposed a new deep learning model that uses separate spatial and temporal sub-networks, which enables high-quality 4D-MRI reconstruction with significantly fewer trainable parameters. Applying this model to 4D-MRI reconstruction resulted in a significant acceleration of 4D-MRI acquisition and reconstruction without losing quality. This thesis has shown that deep learning is a feasible technology for radiotherapy and could enable more effective treatments with fewer side effects and reduce the number of treatment fractions. Although challenges remain for clinical implementation, neural networks are a promising technology for real-time adaptive MRIgRT.

show abstract

Multi-Coil MRI Reconstruction Challenge—Assessing Brain MRI Reconstruction Models and Their Generalizability to Varying Coil Configurations

Cited by 19 publications

References 55 publications

End‐to‐end memory‐efficient reconstruction for cone beam CT

End‐to‐end memory‐efficient reconstruction for cone beam CT

Advancing GABA-edited MRS Research through a Reconstruction Challenge

Real-time MRI-guided radiotherapy with deep learning

Contact Info

Product

Resources

About