Residual learning for 3D motion corrected quantitative MRI: Robust clinical T1, T2 and proton density mapping

Carolin Pirkl, Matteo Cencini, Jan W. Kurzawski, Diana Waldmannstetter, Hongwei Li, Anjany Sekuboyina, Sebastian Endt, Luca Peretti, Graziella Donatelli, Rosa Pasquariello, Michela Tosetti, Mauro Costagli, Guido Buonincontri, Marion I. Menzel, Bjoern H. Menze
Proceedings of the Fourth Conference on Medical Imaging with Deep Learning, PMLR 143:618-632, 2021.

Abstract

Subject motion is one of the major challenges in clinical routine MR imaging. Despite ongoing research, motion correction has remained a complex problem without a universal solution. In advanced quantitative MR techniques, such as MR Fingerprinting, motion does not only affect a single image, like in single-contrast MRI, but disrupts the entire temporal evolution of the magnetization and causes parameter quantification errors due to a mismatch between the acquired and simulated signals. In this work, we present a deep learning-empowered retrospective motion correction for rapid 3D whole-brain multiparametric MRI based on Quantitative Transient-state Imaging (QTI). We propose a patch-based 3D multiscale convolutional neural network (CNN) that learns the residual error, i.e. after initial navigator-based correction, between motion-affected quantitative T1, T2 and proton density maps and their motion-free counterparts. For efficient model training despite limited data availability, we propose a physics-informed simulation to apply continuous motion-patterns to motion-free data. We evaluate the performance of the residual CNN on 1.5T and 3T MRI data of ten healthy volunteers. We analyze the generalizability of the model when applied to real clinical cases, including pediatric and adult patients with large brain lesions. Our study demonstrates that image quality can be significantly improved after correcting for subject motion. This has important implications in clinical setups where large amounts of motion affected data must be discarded.

Cite this Paper

BibTeX
@InProceedings{pmlr-v143-pirkl21a, title = {Residual learning for 3D motion corrected quantitative {MRI}: Robust clinical T1, T2 and proton density mapping}, author = {Pirkl, Carolin and Cencini, Matteo and Kurzawski, Jan W. and Waldmannstetter, Diana and Li, Hongwei and Sekuboyina, Anjany and Endt, Sebastian and Peretti, Luca and Donatelli, Graziella and Pasquariello, Rosa and Tosetti, Michela and Costagli, Mauro and Buonincontri, Guido and Menzel, Marion I. and Menze, Bjoern H.}, booktitle = {Proceedings of the Fourth Conference on Medical Imaging with Deep Learning}, pages = {618--632}, year = {2021}, editor = {Heinrich, Mattias and Dou, Qi and de Bruijne, Marleen and Lellmann, Jan and Schläfer, Alexander and Ernst, Floris}, volume = {143}, series = {Proceedings of Machine Learning Research}, month = {07--09 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v143/pirkl21a/pirkl21a.pdf}, url = {https://proceedings.mlr.press/v143/pirkl21a.html}, abstract = {Subject motion is one of the major challenges in clinical routine MR imaging. Despite ongoing research, motion correction has remained a complex problem without a universal solution. In advanced quantitative MR techniques, such as MR Fingerprinting, motion does not only affect a single image, like in single-contrast MRI, but disrupts the entire temporal evolution of the magnetization and causes parameter quantification errors due to a mismatch between the acquired and simulated signals. In this work, we present a deep learning-empowered retrospective motion correction for rapid 3D whole-brain multiparametric MRI based on Quantitative Transient-state Imaging (QTI). We propose a patch-based 3D multiscale convolutional neural network (CNN) that learns the residual error, i.e. after initial navigator-based correction, between motion-affected quantitative T1, T2 and proton density maps and their motion-free counterparts. For efficient model training despite limited data availability, we propose a physics-informed simulation to apply continuous motion-patterns to motion-free data. We evaluate the performance of the residual CNN on 1.5T and 3T MRI data of ten healthy volunteers. We analyze the generalizability of the model when applied to real clinical cases, including pediatric and adult patients with large brain lesions. Our study demonstrates that image quality can be significantly improved after correcting for subject motion. This has important implications in clinical setups where large amounts of motion affected data must be discarded.} }
Endnote
%0 Conference Paper %T Residual learning for 3D motion corrected quantitative MRI: Robust clinical T1, T2 and proton density mapping %A Carolin Pirkl %A Matteo Cencini %A Jan W. Kurzawski %A Diana Waldmannstetter %A Hongwei Li %A Anjany Sekuboyina %A Sebastian Endt %A Luca Peretti %A Graziella Donatelli %A Rosa Pasquariello %A Michela Tosetti %A Mauro Costagli %A Guido Buonincontri %A Marion I. Menzel %A Bjoern H. Menze %B Proceedings of the Fourth Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2021 %E Mattias Heinrich %E Qi Dou %E Marleen de Bruijne %E Jan Lellmann %E Alexander Schläfer %E Floris Ernst %F pmlr-v143-pirkl21a %I PMLR %P 618--632 %U https://proceedings.mlr.press/v143/pirkl21a.html %V 143 %X Subject motion is one of the major challenges in clinical routine MR imaging. Despite ongoing research, motion correction has remained a complex problem without a universal solution. In advanced quantitative MR techniques, such as MR Fingerprinting, motion does not only affect a single image, like in single-contrast MRI, but disrupts the entire temporal evolution of the magnetization and causes parameter quantification errors due to a mismatch between the acquired and simulated signals. In this work, we present a deep learning-empowered retrospective motion correction for rapid 3D whole-brain multiparametric MRI based on Quantitative Transient-state Imaging (QTI). We propose a patch-based 3D multiscale convolutional neural network (CNN) that learns the residual error, i.e. after initial navigator-based correction, between motion-affected quantitative T1, T2 and proton density maps and their motion-free counterparts. For efficient model training despite limited data availability, we propose a physics-informed simulation to apply continuous motion-patterns to motion-free data. We evaluate the performance of the residual CNN on 1.5T and 3T MRI data of ten healthy volunteers. We analyze the generalizability of the model when applied to real clinical cases, including pediatric and adult patients with large brain lesions. Our study demonstrates that image quality can be significantly improved after correcting for subject motion. This has important implications in clinical setups where large amounts of motion affected data must be discarded.
APA
Pirkl, C., Cencini, M., Kurzawski, J.W., Waldmannstetter, D., Li, H., Sekuboyina, A., Endt, S., Peretti, L., Donatelli, G., Pasquariello, R., Tosetti, M., Costagli, M., Buonincontri, G., Menzel, M.I. & Menze, B.H.. (2021). Residual learning for 3D motion corrected quantitative MRI: Robust clinical T1, T2 and proton density mapping. Proceedings of the Fourth Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 143:618-632 Available from https://proceedings.mlr.press/v143/pirkl21a.html.

Related Material