Deep blind arterial input function: signal correction in perfusion cardiac magnetic resonance

Habib Rebbah, Magalie Viallon, Pierre Croisille, Timothé Boutelier
Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning, PMLR 250:1237-1256, 2024.

Abstract

Objectives: The non-linear relationship between gadolinium concentration and the signal in perfusion cardiac magnetic resonance (CMR) poses a significant challenge for accurate quantification of pharmacokinetic parameters. This phenomenon primarily impacts the arterial input function (AIF), causing it to appear saturated in comparison to the temporal concentration profile. This study aims to leverage a blind deconvolution strategy through a deep-learning approach to address the saturation in the AIF.Methods: We propose the utilization of a convolutional neural network (CNN) architecture with the saturated AIF and a set of myocardial tissue signals as inputs, generating the corrected AIF as the output. To train the network, a dataset comprising over 3×10^6 simulated AIFs with associated signals from five simulated tissues response for each instance was employed. To assess the effectiveness of the approach, the trained network was evaluated using a dual-saturation sequence to compare the corrected AIF with the unsaturated version. The clinical dataset encompassed scans from 43 patients.Results: The mean square error (MSE) for the testing subset of the simulated database was 0.69% of the peak. In the in vivo dataset, the coefficient of determination R2 was 0.26 and 0.86 for the saturated and corrected AIF, respectively, in comparison to the unsaturated AIF.Conclusion: The proposed network successfully corrects the acquisition-induced effects on the AIF. Moreover, the extensive simulated database, featuring diverse acquisition parameters, facilitates the robust generalization of the networkś application.

Cite this Paper

BibTeX
@InProceedings{pmlr-v250-rebbah24a, title = {Deep blind arterial input function: signal correction in perfusion cardiac magnetic resonance}, author = {Rebbah, Habib and Viallon, Magalie and Croisille, Pierre and Boutelier, Timoth\'e}, booktitle = {Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning}, pages = {1237--1256}, year = {2024}, editor = {Burgos, Ninon and Petitjean, Caroline and Vakalopoulou, Maria and Christodoulidis, Stergios and Coupe, Pierrick and Delingette, Hervé and Lartizien, Carole and Mateus, Diana}, volume = {250}, series = {Proceedings of Machine Learning Research}, month = {03--05 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v250/main/assets/rebbah24a/rebbah24a.pdf}, url = {https://proceedings.mlr.press/v250/rebbah24a.html}, abstract = {Objectives: The non-linear relationship between gadolinium concentration and the signal in perfusion cardiac magnetic resonance (CMR) poses a significant challenge for accurate quantification of pharmacokinetic parameters. This phenomenon primarily impacts the arterial input function (AIF), causing it to appear saturated in comparison to the temporal concentration profile. This study aims to leverage a blind deconvolution strategy through a deep-learning approach to address the saturation in the AIF.Methods: We propose the utilization of a convolutional neural network (CNN) architecture with the saturated AIF and a set of myocardial tissue signals as inputs, generating the corrected AIF as the output. To train the network, a dataset comprising over 3×10^6 simulated AIFs with associated signals from five simulated tissues response for each instance was employed. To assess the effectiveness of the approach, the trained network was evaluated using a dual-saturation sequence to compare the corrected AIF with the unsaturated version. The clinical dataset encompassed scans from 43 patients.Results: The mean square error (MSE) for the testing subset of the simulated database was 0.69% of the peak. In the in vivo dataset, the coefficient of determination R2 was 0.26 and 0.86 for the saturated and corrected AIF, respectively, in comparison to the unsaturated AIF.Conclusion: The proposed network successfully corrects the acquisition-induced effects on the AIF. Moreover, the extensive simulated database, featuring diverse acquisition parameters, facilitates the robust generalization of the networkś application.} }
Endnote
%0 Conference Paper %T Deep blind arterial input function: signal correction in perfusion cardiac magnetic resonance %A Habib Rebbah %A Magalie Viallon %A Pierre Croisille %A Timothé Boutelier %B Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2024 %E Ninon Burgos %E Caroline Petitjean %E Maria Vakalopoulou %E Stergios Christodoulidis %E Pierrick Coupe %E Hervé Delingette %E Carole Lartizien %E Diana Mateus %F pmlr-v250-rebbah24a %I PMLR %P 1237--1256 %U https://proceedings.mlr.press/v250/rebbah24a.html %V 250 %X Objectives: The non-linear relationship between gadolinium concentration and the signal in perfusion cardiac magnetic resonance (CMR) poses a significant challenge for accurate quantification of pharmacokinetic parameters. This phenomenon primarily impacts the arterial input function (AIF), causing it to appear saturated in comparison to the temporal concentration profile. This study aims to leverage a blind deconvolution strategy through a deep-learning approach to address the saturation in the AIF.Methods: We propose the utilization of a convolutional neural network (CNN) architecture with the saturated AIF and a set of myocardial tissue signals as inputs, generating the corrected AIF as the output. To train the network, a dataset comprising over 3×10^6 simulated AIFs with associated signals from five simulated tissues response for each instance was employed. To assess the effectiveness of the approach, the trained network was evaluated using a dual-saturation sequence to compare the corrected AIF with the unsaturated version. The clinical dataset encompassed scans from 43 patients.Results: The mean square error (MSE) for the testing subset of the simulated database was 0.69% of the peak. In the in vivo dataset, the coefficient of determination R2 was 0.26 and 0.86 for the saturated and corrected AIF, respectively, in comparison to the unsaturated AIF.Conclusion: The proposed network successfully corrects the acquisition-induced effects on the AIF. Moreover, the extensive simulated database, featuring diverse acquisition parameters, facilitates the robust generalization of the networkś application.
APA
Rebbah, H., Viallon, M., Croisille, P. & Boutelier, T.. (2024). Deep blind arterial input function: signal correction in perfusion cardiac magnetic resonance. Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 250:1237-1256 Available from https://proceedings.mlr.press/v250/rebbah24a.html.

Related Material