Validation and Generalizability of Self-Supervised Image Reconstruction Methods for Undersampled MRI

Thomas Yu1, Tom Hilbert2, Gian Franco Piredda2, Arun Joseph2, Gabriele Bonanno2, Salim Zenkhri3, Patrick Omoumi3, Meritxell Bach Cuadra4, Erick Canales Rodriguez1, Tobias Kober2, Jean-Philippe Thiran1
1: Electrical Engineering, Ecole Polytechnique Federale de Lausanne, 2: Siemens Healthcare AG, 3: Radiology, Lausanne University Hospital, 4: University of Lausanne
Publication date: 2022/09/13
PDF · Dataset · Raw data · arXiv


Deep learning methods have become the state of the art for undersampled MR reconstruction. Particularly for cases where it is infeasible or impossible for ground truth, fully sampled data to be acquired, self-supervised machine learning methods for reconstruction are becoming increasingly used. However potential issues in the validation of such methods, as well as their generalizability, remain underexplored. In this paper, we investigate important aspects of the validation of self-supervised algorithms for reconstruction of undersampled MR images: quantitative evaluation of prospective reconstructions, potential differences between prospective and retrospective reconstructions, suitability of commonly used quantitative metrics, and generalizability. Two self-supervised algorithms based on self-supervised denoising and the deep image prior were investigated. These methods are compared to a least squares fitting and a compressed sensing reconstruction using in-vivo and phantom data. Their generalizability was tested with prospectively under-sampled data from experimental conditions different to the training. We show that prospective reconstructions can exhibit significant distortion relative to retrospective reconstructions/ground truth. Furthermore, pixel-wise quantitative metrics may not capture differences in perceptual quality accurately, in contrast to a perceptual metric. In addition, all methods showed potential for generalization; however, generalizability is more affected by changes in anatomy/contrast than other changes. We further showed that no-reference image metrics correspond well with human rating of image quality for studying generalizability. Finally, we showed that a well-tuned compressed sensing reconstruction and learned denoising perform similarly on all data. The datasets acquired for this paper will be made available online; see for details.


Deep Learning · Self-Supervised Learning · MR Image Reconstruction · Validation · Generalizability · Inverse Problems

Bibtex @article{melba:2022:022:yu, title = "Validation and Generalizability of Self-Supervised Image Reconstruction Methods for Undersampled MRI", author = "Yu, Thomas and Hilbert, Tom and Piredda, Gian Franco and Joseph, Arun and Bonanno, Gabriele and Zenkhri, Salim and Omoumi, Patrick and Cuadra, Meritxell Bach and Canales Rodriguez, Erick and Kober, Tobias and Thiran, Jean-Philippe", journal = "Machine Learning for Biomedical Imaging", volume = "1", issue = "September 2022 issue", year = "2022", pages = "1--31", issn = "2766-905X", doi = "", url = "" }
RISTY - JOUR AU - Yu, Thomas AU - Hilbert, Tom AU - Piredda, Gian Franco AU - Joseph, Arun AU - Bonanno, Gabriele AU - Zenkhri, Salim AU - Omoumi, Patrick AU - Cuadra, Meritxell Bach AU - Canales Rodriguez, Erick AU - Kober, Tobias AU - Thiran, Jean-Philippe PY - 2022 TI - Validation and Generalizability of Self-Supervised Image Reconstruction Methods for Undersampled MRI T2 - Machine Learning for Biomedical Imaging VL - 1 IS - September 2022 issue SP - 1 EP - 31 SN - 2766-905X DO - UR - ER -

2022:022 cover