Image quality affects deep learning reconstruction of MRI. Jeelani, H., Martin, J., Vasquez, F., Salerno, M., & Weller, D. S. In 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), pages 357–360, Washington, DC, April, 2018. IEEE. ![Image quality affects deep learning reconstruction of MRI [link]](https://bibbase.org/img/filetypes/link.svg "link")
Paper doi abstract bibtex The magnetic resonance imaging (MRI) process is susceptible to a wide range of artifacts caused by various sources. In some cases, artifacts might be confused with pathology. In addition, state-of-the-art dynamic MR reconstruction algorithms are iterative in nature, causing longer reconstruction times. Recently, deep learning has been applied to MRI reconstruction and produces high quality images at high acceleration rates. Since deep learning highly depends on training data, the quality of training images must not be ignored. This article demonstrates how noisy images in the training data affect the quality of MR reconstruction. The proposed method modifies the loss function of the neural network to prefer higher quality target images by using a weighted loss function. In this paper mean squared error loss is used, but the approach can be extended to other types of loss function. Using still frames from cardiac MRI’s, this approach is compared to existing approaches that discard noisy training data or ignore these quality differences. Even a basic weighting strategy improves the deep learning reconstruction quality over such methods.
@inproceedings{jeelani_image_2018,
address = {Washington, DC},
title = {Image quality affects deep learning reconstruction of {MRI}},
isbn = {978-1-5386-3636-7},
url = {https://ieeexplore.ieee.org/document/8363592/},
doi = {10.1109/ISBI.2018.8363592},
abstract = {The magnetic resonance imaging (MRI) process is susceptible to a wide range of artifacts caused by various sources. In some cases, artifacts might be confused with pathology. In addition, state-of-the-art dynamic MR reconstruction algorithms are iterative in nature, causing longer reconstruction times. Recently, deep learning has been applied to MRI reconstruction and produces high quality images at high acceleration rates. Since deep learning highly depends on training data, the quality of training images must not be ignored. This article demonstrates how noisy images in the training data affect the quality of MR reconstruction. The proposed method modifies the loss function of the neural network to prefer higher quality target images by using a weighted loss function. In this paper mean squared error loss is used, but the approach can be extended to other types of loss function. Using still frames from cardiac MRI’s, this approach is compared to existing approaches that discard noisy training data or ignore these quality differences. Even a basic weighting strategy improves the deep learning reconstruction quality over such methods.},
language = {en},
urldate = {2021-09-07},
booktitle = {2018 {IEEE} 15th {International} {Symposium} on {Biomedical} {Imaging} ({ISBI} 2018)},
publisher = {IEEE},
author = {Jeelani, Haris and Martin, Jonathan and Vasquez, Francis and Salerno, Michael and Weller, Daniel S.},
month = apr,
year = {2018},
pages = {357--360},
}
Downloads: 0
{"_id":"s9E5s6cmicWbkKSaL","bibbaseid":"jeelani-martin-vasquez-salerno-weller-imagequalityaffectsdeeplearningreconstructionofmri-2018","author_short":["Jeelani, H.","Martin, J.","Vasquez, F.","Salerno, M.","Weller, D. S."],"bibdata":{"bibtype":"inproceedings","type":"inproceedings","address":"Washington, DC","title":"Image quality affects deep learning reconstruction of MRI","isbn":"978-1-5386-3636-7","url":"https://ieeexplore.ieee.org/document/8363592/","doi":"10.1109/ISBI.2018.8363592","abstract":"The magnetic resonance imaging (MRI) process is susceptible to a wide range of artifacts caused by various sources. In some cases, artifacts might be confused with pathology. In addition, state-of-the-art dynamic MR reconstruction algorithms are iterative in nature, causing longer reconstruction times. Recently, deep learning has been applied to MRI reconstruction and produces high quality images at high acceleration rates. Since deep learning highly depends on training data, the quality of training images must not be ignored. This article demonstrates how noisy images in the training data affect the quality of MR reconstruction. The proposed method modifies the loss function of the neural network to prefer higher quality target images by using a weighted loss function. In this paper mean squared error loss is used, but the approach can be extended to other types of loss function. Using still frames from cardiac MRI’s, this approach is compared to existing approaches that discard noisy training data or ignore these quality differences. Even a basic weighting strategy improves the deep learning reconstruction quality over such methods.","language":"en","urldate":"2021-09-07","booktitle":"2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)","publisher":"IEEE","author":[{"propositions":[],"lastnames":["Jeelani"],"firstnames":["Haris"],"suffixes":[]},{"propositions":[],"lastnames":["Martin"],"firstnames":["Jonathan"],"suffixes":[]},{"propositions":[],"lastnames":["Vasquez"],"firstnames":["Francis"],"suffixes":[]},{"propositions":[],"lastnames":["Salerno"],"firstnames":["Michael"],"suffixes":[]},{"propositions":[],"lastnames":["Weller"],"firstnames":["Daniel","S."],"suffixes":[]}],"month":"April","year":"2018","pages":"357–360","bibtex":"@inproceedings{jeelani_image_2018,\n\taddress = {Washington, DC},\n\ttitle = {Image quality affects deep learning reconstruction of {MRI}},\n\tisbn = {978-1-5386-3636-7},\n\turl = {https://ieeexplore.ieee.org/document/8363592/},\n\tdoi = {10.1109/ISBI.2018.8363592},\n\tabstract = {The magnetic resonance imaging (MRI) process is susceptible to a wide range of artifacts caused by various sources. In some cases, artifacts might be confused with pathology. In addition, state-of-the-art dynamic MR reconstruction algorithms are iterative in nature, causing longer reconstruction times. Recently, deep learning has been applied to MRI reconstruction and produces high quality images at high acceleration rates. Since deep learning highly depends on training data, the quality of training images must not be ignored. This article demonstrates how noisy images in the training data affect the quality of MR reconstruction. The proposed method modifies the loss function of the neural network to prefer higher quality target images by using a weighted loss function. In this paper mean squared error loss is used, but the approach can be extended to other types of loss function. Using still frames from cardiac MRI’s, this approach is compared to existing approaches that discard noisy training data or ignore these quality differences. Even a basic weighting strategy improves the deep learning reconstruction quality over such methods.},\n\tlanguage = {en},\n\turldate = {2021-09-07},\n\tbooktitle = {2018 {IEEE} 15th {International} {Symposium} on {Biomedical} {Imaging} ({ISBI} 2018)},\n\tpublisher = {IEEE},\n\tauthor = {Jeelani, Haris and Martin, Jonathan and Vasquez, Francis and Salerno, Michael and Weller, Daniel S.},\n\tmonth = apr,\n\tyear = {2018},\n\tpages = {357--360},\n}\n\n","author_short":["Jeelani, H.","Martin, J.","Vasquez, F.","Salerno, M.","Weller, D. S."],"key":"jeelani_image_2018","id":"jeelani_image_2018","bibbaseid":"jeelani-martin-vasquez-salerno-weller-imagequalityaffectsdeeplearningreconstructionofmri-2018","role":"author","urls":{"Paper":"https://ieeexplore.ieee.org/document/8363592/"},"metadata":{"authorlinks":{}},"html":""},"bibtype":"inproceedings","biburl":"https://bibbase.org/zotero/sede","dataSources":["pQYee5oovEhJF75n6"],"keywords":[],"search_terms":["image","quality","affects","deep","learning","reconstruction","mri","jeelani","martin","vasquez","salerno","weller"],"title":"Image quality affects deep learning reconstruction of MRI","year":2018}