Abstract
Objectives
To develop and share a deep learning method that can accurately identify optimal inversion time (TI) from multi-vendor, multi-institutional and multi-field strength inversion scout (TI scout) sequences for late gadolinium enhancement cardiac MRI.
Materials and methods
Retrospective multicentre study conducted on 1136 1.5-T and 3-T cardiac MRI examinations from four centres and three scanner vendors. Deep learning models, comprising a convolutional neural network (CNN) that provides input to a long short-term memory (LSTM) network, were trained on TI scout pixel data from centres 1 to 3 to identify optimal TI, using ground truth annotations by two readers. Accuracy within 50 ms, mean absolute error (MAE), Lin’s concordance coefficient (LCCC) and reduced major axis regression (RMAR) were used to select the best model from validation results, and applied to holdout test data. Robustness of the best-performing model was also tested on imaging data from centre 4.
Results
The best model (SE-ResNet18-LSTM) produced accuracy of 96.1%, MAE 22.9 ms and LCCC 0.47 compared to ground truth on the holdout test set and accuracy of 97.3%, MAE 15.2 ms and LCCC 0.64 when tested on unseen external (centre 4) data. Differences in vendor performance were observed, with greatest accuracy for the most commonly represented vendor in the training data.
Conclusion
A deep learning model was developed that can identify optimal inversion time from TI scout images on multi-vendor data with high accuracy, including on previously unseen external data. We make this model available to the scientific community for further assessment or development.
Clinical relevance statement
A robust automated inversion time selection tool for late gadolinium–enhanced imaging allows for reproducible and efficient cross-vendor inversion time selection.
Key Points
• A model comprising convolutional and recurrent neural networks was developed to extract optimal TI from TI scout images.
• Model accuracy within 50 ms of ground truth on multi-vendor holdout and external data of 96.1% and 97.3% respectively was achieved.
• This model could improve workflow efficiency and standardise optimal TI selection for consistent LGE imaging.
Similar content being viewed by others
Explore related subjects
Discover the latest articles, news and stories from top researchers in related subjects.Avoid common mistakes on your manuscript.
Introduction
Late gadolinium–enhanced (LGE) imaging is integral to clinical cardiac MRI, with applications including evaluation of tissue viability in coronary artery disease and diagnosis of non-ischaemic cardiomyopathies [1, 2]. LGE relies upon appropriate selection of an optimal inversion time (TI) to suppress background myocardial signal [2,3,4], particularly where only magnitude reconstructed images are available. A Look–Locker TI scout sequence is typically used [5, 6]. The optimal TI varies individually and depends upon contrast agent, dose, acquisition timing and patient factors. Myocardial pathologies impact contrast distribution within the myocardium, for example with increased extracellular volume as is seen in cardiac amyloid, or breach of the intracellular space with myocardial infarction [7]. Whilst phase-sensitive inversion recovery imaging can be performed to minimise dependence upon optimal TI selection [8], TI scout imaging remains widely performed in clinical practice [2].
The optimal TI is typically manually selected by the examining technologist or supervising clinician, subject to factors including level of experience. An automated approach may offer a reproducible, objective and time-efficient alternative. Deep learning methodologies, particularly convolutional neural networks (CNN [9]), have shown promising results in cardiac MRI imaging applications including quantification of LV (left ventricular) function [10], characterisation of tissue scarring [11] and prognostication [12]. Given the spatiotemporal nature of TI scout imaging, an additional recurrent neural network (RNN) may aid in capturing temporal elements [13].
Bahrami et al previously created a deep learning (DL) model for optimal TI selection with promising results in a single-vendor, single-institution setting [14]. However, whether this can translate to a model that performs well on imaging from different institutions is unknown, with TI scout imaging acquired with different pulse sequences and at different field strengths. We hypothesise an integrated CNN and LSTM [15] model which takes advantage of spatiotemporal relations and uses techniques to minimise overfitting can overcome challenges related to generalisability across vendors, institutions and field strengths.
The purpose of our study is to develop an artificial intelligence model capable of selecting optimal TI from scout sequences, applicable to imaging obtained from different institutions and different scanner systems. In addition, we aim to share our developed model with the research community.
Materials and methods
Data source
This was an international retrospective cross-sectional study from four centres with institutional ethics approval. The dataset comprised 1136 de-identified studies where TI scout sequences were performed in 1100 patients (mean age 54.5 years, 691 male, 409 female). Studies were performed at 1.5 T or 3 T, with three vendor systems. Thirty-six repeat studies were present in the centre 4 group, not used in model training. The most common indication for centre 1–3 examinations was assessment of scar and myocardial viability (n = 259). For centre 4, aortic stenosis was the predominant study indication (54/67, 80.6%). Magnet systems, administered contrast and patient demographics are summarised in Table 1. Scanner-specific acquisition parameters are provided in Supplementary Materials.
For each study, a median (range) of 34 (9–100) frames was obtained in the TI scout series with median (range) inversion time 477.5 (231.0–1935.0) ms and median (range) temporal resolution 23.0 (8.8–45.3) ms. Images with the highest temporal resolution were acquired at centre 3 with median (range) of 16.7 (8.8–27.0) ms, with lower temporal resolution for one centre 2 system with median (range) 38.8 (23.3–45.3) ms. Images were retrieved in Digital Imaging and Communications in Medicine (DICOM) [16] format and anonymised using Python code.
A portion of the dataset (427/1100 subjects, detailed in Supplementary Materials) overlaps with previously published work, where the aim was to automate sorting of various cardiac MRI sequences, unrelated to the purpose of the current study [17].
Ground truth
The image with optimal myocardial nulling (the null point) was selected from the TI scout series by a radiologist with 17 years’ experience with cardiac MRI (R.P.L., SCMR level 3 equivalent), and a radiology trainee (E.C.) with 2 years’ experience with cardiac MRI interpretation, following training sessions prior to independent labelling. A subset of centre 1–3 data (327/1033, 31.7%) was interpreted independently by both readers, with agreement assessed prior to further data labelling, and differences subsequently resolved in consensus for ground truth labels. As there was excellent agreement between independent readers (Spearman correlation 0.94), the remaining 706 centre 1–3 cases were labelled by the radiology trainee alone. Trainee-labelled data was used solely for model training, and the holdout test set from centres 1 to 3 used only consensus data. All centre 4 data was labelled by the expert radiologist for external validation of the final model.
Data elements
Model input was left ventricular mid short-axis TI scout images. The outcome variable for model development was the index of the optimal TI image. Each frame of the series was given a binary class—before (labelled 0) or after the null point (labelled 1). A soft label of 0.5 was given to the null point frame. Through a sequence-to-sequence approach, the model receives frames individually in a recurrent fashion, and then outputs the probability for the respective frames (Fig. 1). The transition point of these intermediary values is used to derive the outcome variable, i.e. for prediction, the image frame with the label closest to 0.5 is considered the optimal image (detailed under “Model design”).
Data pre-processing
Several oversampling techniques were utilised during the training phase to help the model learn from underrepresented data as well as critical frames corresponding to the null point (further explained in Supplementary Materials—Data Preprocessing).
For every epoch during training, a splice of nine consecutive frames was sampled from each series. Images were standardised with global mean and variance, and then resized to 256 × 256 resolution through interpolation. Two frames were randomly removed from each nine-frame sequence to emulate varying temporal resolutions used in scout acquisition protocols across centres and magnet systems. The remaining seven frames were then grouped with their neighbouring frames (referred to below as a window), such that there were five sets of three adjacent frames across five timesteps, with oversampling around the null point when myocardial signal is suppressed (Fig. 2; additional details in Supplementary Materials).
During the testing and validation and testing phases, all images from each TI scout series were passed as is in groups of three consecutive images to the model, without oversampling or splicing (Fig. 2).
Model design
The model design consisted of two parts: a feature extraction CNN, and a bidirectional LSTM network (Fig. 3).
The CNN first iterates through all windows of a given series, where the channels correspond to the earliest, middle and latest frames within a given window, and thereby help identify both spatial features and temporal relationships between neighbouring frames (e.g. change in intensity). Three CNN architectures were evaluated—VGG16 [18], ResNet18 [19] and Squeeze-and-Excitation [20] SE-ResNet18, a modified ResNet architecture utilising Squeeze-and-Excitation blocks to perform channel-wise feature recalibration and enhance channel interdependencies. The CNN compresses features into a feature array with a dimension of 1024, to produce embedded representations of the windows.
The bidirectional LSTM network was then used to capture global temporal information beyond neighbouring frames. A final dense layer performs logistic regression and outputs a sequence of probabilities, where each value corresponds to a window. The predicted optimal TI was extracted from DICOM metadata (“InversionTime” for GE and “TriggerTime” for Siemens and Philips images) of the centre image of the window whose probability was closest to 0.5.
Data partitioning
A total of 154 of 1033 samples (14.9%) from centres 1 to 3 were held out for testing. The remainder were used for training and validation, using five-fold cross-validation with an 80/20 training/validation split. The best-performing model from five-fold cross-validation was selected as the final model, whose performance was tested on the holdout test set from centres 1 to 3.
Model training
Training was performed on an Nvidia Tesla V100 virtual partition with 16 gigabytes of VRAM, using Python 3.9.6 and PyTorch 1.9.0. Details of other packages are provided in Supplementary Materials.
Data augmentation was performed using an open-source framework (MONAI [21]) to expose the model to more heterogeneous training data and improve generalisability. Grid search was performed to identify the most suitable learning rate, decay rate and batch size. The CNN-LSTM model was trained end-to-end using ADAM optimisation over 30 epochs with a batch size of 32 and weight decay of 1 × 10−4. A differential learning rate was utilised, with an initial learning rate of 1.5 × 10−2 for the feature extractor and 3 × 10−3 for the LSTM. A weighted Binary Cross-Entropy loss function was used. For the loss of each sample, larger weights were applied around the null point:
where Ti is the inversion time at window i.
Detailed description of augmentations employed, model details including hyperparameters and the weighted loss function are provided in Supplementary Materials.
Model evaluation
Metrics
Model performance was evaluated with mean absolute error (MAE), accuracy (defined as model selection within 50 ms of ground truth, based on degree of variability of TI scout series of differing temporal resolutions and typical temporal resolution of cardiac MR sequences) [3], Lin’s concordance correlation coefficient (LCCC) and Bland–Altman analysis. Reduced major axis regression (RMAR) was also used to assess fixed and proportional bias of the predictions, with the latter not assessed with Bland–Altman analysis [22]. RMAR accounts for potential errors in both prediction and ground truth, as opposed to linear regression, which only incorporates prediction error. In this study, MAE was prioritised for model selection, as it most directly reflects the discrepancy between predicted inversion time and ground truth. Each metric was selected to reflect an aspect of model performance but may not completely correlate, as optimising for one metric may negatively impact others.
Inter-reader agreement on the holdout test set was also assessed with the same metrics for direct comparison. Subset analysis of model performance was carried out for individual centres, vendors and field strengths.
Visualisations
2D histograms of prediction and ground truth distributions were created to visualise concordance between predicted inversion time obtained from the model and ground truth in 25-ms intervals. This bin size was selected to adequately display the density information of these distributions. Saliency maps were also employed to assist with model interpretability, further detailed in Supplementary Materials [23].
External data
To assess generalisability, the final model was tested on all centre 4 data from two vendors (n = 103), previously unseen by the model during training and validation.
Results
Model evaluation
The best-performing model was SE-Resnet-18, which produced accuracy of 96.1%, MAE 22.9 ms and LCCC 0.472 on the holdout test set from centres 1 to 3 (Table 2), with further details of model selection provided in Supplementary Materials. RMAR demonstrated no proportional bias and fixed bias of 17.3 ms, indicating a slight tendency to overestimate the optimal TI (Fig. 4a, b). 2D histogram also shows strong performance of the final model to predict the optimal TI (Fig. 4c), with most errors within 50 ms of ground truth optimal TI, erring more on the side of overestimation of optimal TI. For comparison, inter-reader agreement yielded LCCC of 0.91, with no major proportional bias and similar fixed bias of 16.2 ms. The model performed better on 1.5 T data (accuracy 97.8%) than limited 3 T data (accuracy 72.7%), when data from centres 1 and 2 (where imaging at both field strengths was available) was analysed. Details of inter-reader agreement and subanalysis of model performance at different field strengths are provided in Supplementary Materials. Distribution of errors by centre and vendor is summarised in Table 3.
Spatial saliency heatmaps of the model demonstrated a patchy grid-like pattern with attention primarily in the heart region (Fig. 5a). Attention was also present in surrounding tissues including the lung and intra-abdominal organs. The model paid little attention to aliasing artefacts in the image peripheries. The temporal saliency plot displayed strong attention in frames near the predicted null point (Fig. 5b).
External data performance
The final model achieved accuracy of 97.3% on external data from centre 4, with MAE 15.2 ms and LCCC 0.641 (Table 2). RMAR demonstrated a predominantly proportional bias, with a tendency of the model to overestimate at low TIs and underestimate optimal TI at higher TIs (Fig. 4d, e). Results by centre and vendor are summarised in Table 3.
The model achieved 100% accuracy within 50 ms for images from one vendor (Siemens) with a tendency to underestimate the optimal TI. Underestimations were within 50 ms of ground truth (35/90 optimal TIs underestimated, median ± IQR 17.5 ± 2.5 ms), with little discernible difference between ground truth and prediction on image review (Fig. 6a). There was 84.6% accuracy for images from the other vendor in the centre 4 dataset (Phillips), with a tendency to overestimate TI (10/13 optimal TIs overestimated, 2/13 by greater than 50 ms of the optimal TI, median ± IQR 25.0 ± 25.0 ms) (Fig. 6b). Subanalysis of field strength performance again yielded high accuracy for 1.5 T data for the external test set (91.1%), with 84.6% accuracy observed for 3 T data (further detail in Supplementary Materials).
The final model is made available at this site: https://github.com/AH-AI-in-Radiology/Cardiac_MRI_TI.
Discussion
We have developed a serial CNN-RNN approach that can select optimal TI from multicentre multi-vendor MRI TI scout imaging with high accuracy. The potential benefit of an automated approach to TI selection is a method that is 100% reproducible, i.e. given the same input, the model will predict the same TI every time. Through reliable optimal TI selection, it could potentially assist with more accurate diagnosis, obviating poor or inconsistent nulling of normal myocardium that may occur for inexperienced or time-pressured operators. Our model was able to discern the transition from early to late frames, incorporating both spatial and temporal features. Although the RNN component of our approach was fixed as a LSTM, three CNN feature extractors were compared, VGG16 [18], Resnet18 [19] and SE-Resnet-18. The model containing a SE-Resnet-18 CNN achieved the best results, likely due to Squeeze and Excite Blocks. These blocks enable CNNs to perform dynamic channel-wise feature recalibration [20]. We postulate this allowed our model to extract a more discriminative set of features from a temporal series of images at the CNN level. Saliency mapping demonstrated model attention for both spatial (morphologic) and temporal features in predicting optimal TI.
The small (17.3 ms) overestimation of predicted optimal TI in centre 1–3 test data was uniform across TI (i.e. fixed bias). This may relate to the weighted loss function, which consistently applies larger weights to frames after the null point. In practice, minor overestimation is an acceptable error, as selection of later TIs than the optimal TI will still provide late gadolinium–enhanced imaging images where abnormal myocardium should still appear brighter than normal myocardium. Underestimation is potentially more problematic, as the relationship between normal and abnormal myocardium may be inverted if magnitude images alone are relied upon, with a more negative impact on diagnostic accuracy [24]. However, the underestimation observed in our external test data remained within 50 ms of ground truth, with similar visual appearance of predicted and ground truth images.
Poorer performance was noted for Philips data from centres 2 and 4. We postulate acquisition differences as a potential cause. This includes differences in the pulse sequence used to obtain TI scout imaging (balanced steady-state free precession for Siemens systems versus spoiled gradient echo or ultrafast gradient echo images from Philips and GE systems in our cohort), impacting tissue contrast and signal-to-noise. Temporal resolution likely also impacted accuracy, as systems with poorer temporal resolution/fewer frames were likely penalised in terms of accuracy, since true optimal TI is more likely to lie between frames. This is exemplified by the comparatively higher MAE for test set data from centre 2 (median temporal resolution 38.8 ms), and low MAE for centre 3 (median temporal resolution 16.7 ms). However, this was difficult to confirm in external test data which was predominantly obtained from a single-vendor system.
However, high model accuracy for the external data was observed. The heterogeneity of training data from various vendors may have contributed to its performance. The relative homogeneity of the centre 4 dataset (indication of aortic stenosis in the majority) may have contributed to the high observed accuracy. High prevalence of left ventricular hypertrophy in this disease might minimise difficulties with partial voluming of the LV myocardium that can be encountered in patients with low myocardial mass. However, we observed overestimation of optimal TI at lower TIs and underestimation at higher TI on centre 4 data, largely driven by results from the Siemens magnet. However, 101/103 (98.1%) predictions fell within 50 ms of the true optimal TI, and the degree of underestimation (potentially more problematic for late gadolinium–enhanced imaging) was small (all < 50 ms), with little visual difference in predicted versus ground truth imaging in these cases. Underlying pathology (with known prolongation of T1 myocardial values in aortic stenosis [25]), acquisition parameters, contrast type [26], dosage and timing of TI scout acquisition post contrast administration may all have influenced our results, highlighting the real-world variability of the TI scout sequence used between institutions.
Whilst Bahrami et al previously explored automatic detection of optimal TI with CNN and LSTM in a single-institution and single-vendor setting [14], we have developed a model using multicentre data at two field strengths that was able to accurately predict optimal TI for previously unseen data from two vendors from a centre not used for model training. We have shown that prediction of optimal TI is possible from pixel data alone, even for heterogeneous imaging acquired with different pulse sequences, at different field strengths and varying in terms of contrast type, dose and timing protocols. Additionally, we developed a distinct architecture which reduced the parameter counts substantially for improved training efficiency (from over 400 million to 32,258,625). A next step is implementation of such a model into clinical practice, as recently described by Ohta et al in a single-institution, single-vendor study utilising a VGG-19 model to aid TI selection in real time with a smartphone [27].
A limitation of the study was the unbalanced data distribution between centres and magnet systems. Centre 3 data was underrepresented, with only 6 datapoints in the holdout test set, and no representation of the centre 3 vendor in the external dataset. External test data was mainly from a single vendor with a relatively homogeneous patient population, which may bias model performance. Vendor-specific performance should therefore be interpreted with caution. 3-Tesla (3 T) scanner data was also underrepresented. Notably, only 5.6% (49/879) of training and validation data from centres 1 to 3 were from 3 T, probably accounting for poorer model performance on 3 T data. The model’s performance on patients with amyloid was not adequately assessed, with relatively few patients referred for this clinical indication. In clinical practice, optimal TI is difficult to determine in this population, but the TI scout itself can assist with diagnosis due to abnormal contrast kinetics [28]. Variable temporal resolution of the scout sequence in our data likely influenced model training and accuracy, particularly if the ideal TI lies between two discrete samples, impacting series with poorer temporal resolution. However, our dataset reflects heterogeneity of real-world clinical imaging.
We have developed a model capable of automatic detection of optimal TI with high accuracy to within 50 ms of optimal myocardial nulling that can be employed upon a range of sequences obtained from different institutions and vendors. We also share our model with the scientific community, to encourage further development or further testing on real-world data. With refinement, it could assist with efficient and reproducible TI selection for LGE imaging in cardiac MRI and could ultimately be incorporated into a fully automated clinical workflow.
Abbreviations
- ADAM Optimiser:
-
Adaptive Moment Estimation
- CNN:
-
Convolutional neural network
- DICOM:
-
Digital Imaging and Communications in Medicine
- DL:
-
Deep learning
- LCCC:
-
Lin’s concordance correlation coefficient
- LGE:
-
Late gadolinium–enhanced imaging
- LSTM:
-
Long short-term memory
- LV:
-
Left ventricular
- MAE:
-
Mean absolute error
- MONAI:
-
Medical Open Network for Artificial Intelligence
- MRI:
-
Magnetic resonance imaging
- ResNet:
-
Residual neural network
- RMAR:
-
Reduced major axis regression
- SCMR:
-
Society of cardiovascular magnetic resonance
- TI:
-
Time inversion
- VGG:
-
Visual Geometry Group
References
Kim RJ, Fieno DS, Parrish TB et al (1999) Relationship of MRI delayed contrast enhancement to irreversible injury, infarct age, and contractile function. Circulation 100(19):1992–2002
Kramer CM, Barkhausen J, Bucciarelli-Ducci C, Flamm SD, Kim RJ, Nagel E (2020) Standardized cardiovascular magnetic resonance imaging (CMR) protocols: 2020 update. J Cardiovasc Magn Reson 22(1):1–18
Fratz S, Chung T, Greil GF et al (2013) Guidelines and protocols for cardiovascular magnetic resonance in children and adults with congenital heart disease: SCMR expert consensus group on congenital heart disease. J Cardiovasc Magn Reson 15(1):1–26
Kim RJ, Shah DJ, Judd RM (2003) How we perform delayed enhancement imaging: HOW I DO…. J Cardiovasc Magn Reson 5(3):505–514
Look DC, Locker DR (1970) Time saving in measurement of NMR and EPR relaxation times. Rev Sci Instrum 41(2):250–251
Amano Y, Takayama M, Kumita S (2009) Contrast-enhanced myocardial T1-weighted scout (Look–Locker) imaging for the detection of myocardial damages in hypertrophic cardiomyopathy. J Magn Reson Imaging 30(4):778–784
Doltra A, Hoyem Amundsen B, Gebker R, Fleck E, Kelle S (2013) Emerging concepts for myocardial late gadolinium enhancement MRI. Curr Cardiol Rev 9(3):185–190
Kellman P, Arai AE, McVeigh ER, Aletras AH (2002) Phase-sensitive inversion recovery for detecting myocardial infarction using gadolinium-delayed hyperenhancement. Magn Reson Med 47(2):372–383
Krizhevsky A, Sutskever I, Hinton GE (2017) ImageNet classification with deep convolutional neural networks. Commun ACM 60(6):84–90
Zhang D, Icke I, Dogdas B et al (2018) A multi-level convolutional LSTM model for the segmentation of left ventricle myocardium in infarcted porcine cine MR images. IEEE 470–473
Moccia S, Banali R, Martini C et al (2018) Automated scar segmentation from CMR-LGE images using a deep learning approach. 2018 Computing in Cardiology Conference (CinC), IEEE 1-4
Bello GA, Dawes TJ, Duan J et al (2019) Deep-learning cardiac motion analysis for human survival prediction. Nat Mach Intell 1(2):95–104
Hernandez KAL, Rienmüller T, Baumgartner D, Baumgartner C (2021) Deep learning in spatiotemporal cardiac imaging: a review of methodologies and clinical usability. Comput Biol Med 130:104200
Bahrami N, Retson T, Blansit K, Wang K, Hsiao A (2019) Automated selection of myocardial inversion time with a convolutional neural network: spatial temporal ensemble myocardium inversion network (STEMI-NET). Magn Reson Med 81(5):3283–3291
Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735–1780
Mustra M, Delac K, Grgic M (2008) Overview of the DICOM standard. 2008 50th International Symposium ELMAR, IEEE 39-44
Lim RP, Kachel S, Villa AD et al (2022) CardiSort: a convolutional neural network for cross vendor automated sorting of cardiac MR images. Eur Radiol 32(9):5907–5920
Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:14091556. 1409:1556
He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition 770–778
Hu J, Shen L, Sun G (2018) Squeeze-and-Excitation networks 7132–7141
Cardoso MJ, Li W, Brown R et al (2022) MONAI: an open-source framework for deep learning in healthcare. arXiv preprint arXiv:221102701. 2211:02701
Smith RJ (2009) Use and misuse of the reduced major axis for line-fitting. Am J Phys Anthropol 140(3):476–486
Simonyan K, Vedaldi A, Zisserman A (2013) Deep inside convolutional networks: visualising image classification models and saliency maps. arXiv preprint arXiv:13126034. 1312:6034
Kellman P, Arai AE (2012) Cardiac imaging techniques for physicians: late enhancement. J Magn Reson Imaging 36(3):529–542
Elbaum C, Iacuzio L, Bohbot Y et al (2021) Non-contrast myocardial T1 global and regional reference values at 3 Tesla cardiac magnetic resonance in aortic stenosis. Arch Cardiovasc Dis 114(4):293–304
Doeblin P, Schilling R, Wagner M et al (2014) Intraindividual comparison of T1 relaxation times after gadobutrol and Gd-DTPA administration for cardiac late enhancement imaging. Eur J Radiol 83(4):660–664
Ohta Y, Tateishi E, Morita Y, et al (2023) Optimization of null point in Look-Locker images for myocardial late gadolinium enhancement imaging using deep learning and a smartphone. Eur Radiol 1–10.
Pandey T, Jambhekar K, Shaikh R, Lensing S, Viswamitra S (2013) Utility of the inversion scout sequence (TI scout) in diagnosing myocardial amyloid infiltration. Int J Cardiovasc Imaging 29:103–112
Acknowledgements
The authors wish to acknowledge Alana Malandra BAppSc (Med Rad) for her invaluable assistance with data collection.
Funding
Open Access funding enabled and organized by CAUL and its Member Institutions The authors state that this work has not received any funding.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Guarantor
The scientific guarantor of this publication is Ruth P. Lim.
Conflict of interest
The authors of this manuscript declare no relationships with any companies, whose products or services may be related to the subject matter of the article.
Statistics and biometry
One of the authors has significant statistical expertise (Professor L Churilov).
Informed consent
Written informed consent was waived by the Institutional Review Board.
Ethical approval
Institutional Review Board approval was obtained.
Study subjects or cohorts overlap
Some study subjects or cohorts have been previously reported in “CardiSort: a convolutional neural network for cross vendor automated sorting of cardiac MR images,” Eur Radiol 2022 Sep;32(9):5907-5920, in unrelated work. In the current study, 60/718 centre 1 subjects’, 268/268 centre 2 subjects’, 47/47 centre 3 subjects’ and 52/65 centre 4 subjects’ image data was used in the previous publication. However, note the very different purposes of the two studies, with the prior study’s purpose to purely sort different cardiac MRI sequences, and the current work focusing upon one single sequence and automated prediction of optimal inversion time to facilitate scanning workflow. The majority of subjects in this current study have not previously been reported.
Methodology
• retrospective
• cross-sectional study
• multicentre study
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Cheng Xie and Rory Zhang are co-first authors.
Supplementary Information
Below is the link to the electronic supplementary material.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Xie, C., Zhang, R., Mensink, S. et al. Automated inversion time selection for late gadolinium–enhanced cardiac magnetic resonance imaging. Eur Radiol 34, 5816–5828 (2024). https://doi.org/10.1007/s00330-024-10630-w
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00330-024-10630-w