Purpose
To study the influence of gradient echo–based contrasts as input channels to a 3D patch‐based neural network trained for synthetic CT (sCT) generation in canine and human populations.
Methods
...Magnetic resonance images and CT scans of human and canine pelvic regions were acquired and paired using nonrigid registration. Magnitude MR images and Dixon reconstructed water, fat, in‐phase and opposed‐phase images were obtained from a single T1‐weighted multi‐echo gradient‐echo acquisition. From this set, 6 input configurations were defined, each containing 1 to 4 MR images regarded as input channels. For each configuration, a UNet‐derived deep learning model was trained for synthetic CT generation. Reconstructed Hounsfield unit maps were evaluated with peak SNR, mean absolute error, and mean error. Dice similarity coefficient and surface distance maps assessed the geometric fidelity of bones. Repeatability was estimated by replicating the training up to 10 times.
Results
Seventeen canines and 23 human subjects were included in the study. Performance and repeatability of single‐channel models were dependent on the TE‐related water–fat interference with variations of up to 17% in mean absolute error, and variations of up to 28% specifically in bones. Repeatability, Dice similarity coefficient, and mean absolute error were statistically significantly better in multichannel models with mean absolute error ranging from 33 to 40 Hounsfield units in humans and from 35 to 47 Hounsfield units in canines.
Conclusion
Significant differences in performance and robustness of deep learning models for synthetic CT generation were observed depending on the input. In‐phase images outperformed opposed‐phase images, and Dixon reconstructed multichannel inputs outperformed single‐channel inputs.
This study evaluated the accuracy of synthetic computed tomography (sCT), as compared to CT, for the 3D assessment of the hip morphology. Thirty male patients with asymptomatic hips, referred for ...magnetic resonance (MR) imaging and CT, were included in this retrospective study. sCT images were generated from three‐dimensional radiofrequency‐spoiled T1‐weighted multi‐echo gradient‐echo MR images using a commercially available deep learning‐enabled software and were compared with CT images through mean error and surface distance computation and by means of eight clinical morphometric parameters relevant for hip care. Parameters included center‐edge angle (CEA), sharp angle, acetabular index, extrusion index, femoral head center‐to‐midline distance, acetabular version (AV), and anterior and posterior acetabular sector angles. They were measured by two senior orthopedic surgeons and a radiologist in‐training on CT and sCT images. The reliability and agreement of CT‐ and sCT‐based measurements were assessed using intraclass correlation coefficients (ICCs) for absolute agreement, Bland–Altman plots, and two one‐sided tests for equivalence. The surface distance between CT‐ and sCT‐based bone models were on average submillimeter. CT‐ and sCT‐based measurements showed moderate to excellent interobserver and intraobserver correlation (0.56 < ICC < 0.99). In particular, the inter/intraobserver agreements were good for AV (ICC > 0.75). For CEA, the intraobserver agreement was good (ICC > 0.75) and the interobserver agreement was moderate (ICC > 0.69). Limits of agreements were similar between intraobserver CT and intermodal measurements. All measurements were found statistically equivalent, with average intermodal differences within the intraobserver limits of agreement. In conclusion, sCT and CT were equivalent for the assessment of the hip joint bone morphology.
This study compared magnetic resonance imaging‐based synthetic computed tomography (sCT) to conventional CT for the morphological assessment of the hip joint. The sCT generation technique reconstructed radiodensity contrast accurately, providing bone models within a submillimeter accuracy to CT‐based bone models. CT and sCT images were found statistically equivalent for measuring eight morphometric parameters clinically relevant in hip care. For both modalities, moderate to excellent interobserver and intraobserver reliability were obtained. This provides new possibilities in the radiation‐free diagnosis and surgical planning.
Background
Three-dimensional (3D)-printed saw guides are frequently used to optimize osteotomy results and are usually designed based on computed tomography (CT), despite the radiation burden, as ...radiation-less alternatives like magnetic resonance imaging (MRI) have inferior bone visualization capabilities. This study investigated the usability of MR-based synthetic-CT (sCT), a novel radiation-less bone visualization technique for 3D planning and design of patient-specific saw guides.
Methods
Eight human cadaveric lower arms (mean age: 78y) received MRI and CT scans as well as high-resolution micro-CT. From the MRI scans, sCT were generated using a conditional generative adversarial network. Digital 3D bone surface models based on the sCT and general CT were compared to the surface model from the micro-CT that was used as ground truth for image resolution. From both the sCT and CT digital bone models saw guides were designed and 3D-printed in nylon for one proximal and one distal bone position for each radius and ulna. Six blinded observers placed these saw guides as accurately as possible on dissected bones. The position of each guide was assessed by optical 3D-scanning of each bone with positioned saw guide and compared to the preplanning. Eight placement errors were evaluated: three translational errors (along each axis), three rotational errors (around each axis), a total translation (∆T) and a total rotation error (∆R).
Results
Surface models derived from micro-CT were on average smaller than sCT and CT-based models with average differences of 0.27 ± 0.30 mm for sCT and 0.24 ± 0.12 mm for CT. No statistically significant positioning differences on the bones were found between sCT- and CT-based saw guides for any axis specific translational or rotational errors nor between the ∆T (
p
= .284) and ∆R (
p
= .216). On Bland-Altman plots, the ∆T and ∆R limits of agreement (LoA) were within the inter-observer variability LoA.
Conclusions
This research showed a similar error for sCT and CT digital surface models when comparing to ground truth micro-CT models. Additionally, the saw guide study showed equivalent CT- and sCT-based saw guide placement errors. Therefore, MRI-based synthetic CT is a promising radiation-less alternative to CT for the creation of patient-specific osteotomy surgical saw guides.
To extend the double echo steady-state (DESS) sequence to enable chemical-shift corrected water-fat separation.
This study proposes multiple-echo steady-state (MESS), a sequence that modifies the ...readouts of the DESS sequence to acquire two echoes each with bipolar readout gradients with higher readout bandwidth. This enables water-fat separation and eliminates the need for water-selective excitation that is often used in combination with DESS, without increasing scan time. An iterative fitting approach was used to perform joint chemical-shift corrected water-fat separation and T
estimation on all four MESS echoes simultaneously. MESS and water-selective DESS images were acquired for five volunteers, and were compared qualitatively as well as quantitatively on cartilage T
and thickness measurements. Signal-to-noise ratio (SNR) and T
quantification were evaluated numerically using pseudo-replications of the acquisition.
The water-fat separation provided by MESS was robust and with quality comparable to water-selective DESS. MESS T
estimation was similar to DESS, albeit with slightly higher variability. Noise analysis showed that SNR in MESS was comparable to DESS on average, but did exhibit local variations caused by uncertainty in the water-fat separation.
In the same acquisition time as DESS, MESS provides water-fat separation with comparable SNR in the reconstructed water and fat images. By providing additional image contrasts in addition to the water-selective DESS images, MESS provides a promising alternative to DESS.
Purpose
To extend the double echo steady‐state (DESS) sequence to enable chemical‐shift corrected water‐fat separation.
Methods
This study proposes multiple‐echo steady‐state (MESS), a sequence that ...modifies the readouts of the DESS sequence to acquire two echoes each with bipolar readout gradients with higher readout bandwidth. This enables water‐fat separation and eliminates the need for water‐selective excitation that is often used in combination with DESS, without increasing scan time. An iterative fitting approach was used to perform joint chemical‐shift corrected water‐fat separation and T2 estimation on all four MESS echoes simultaneously. MESS and water‐selective DESS images were acquired for five volunteers, and were compared qualitatively as well as quantitatively on cartilage T2 and thickness measurements. Signal‐to‐noise ratio (SNR) and T2 quantification were evaluated numerically using pseudo‐replications of the acquisition.
Results
The water‐fat separation provided by MESS was robust and with quality comparable to water‐selective DESS. MESS T2 estimation was similar to DESS, albeit with slightly higher variability. Noise analysis showed that SNR in MESS was comparable to DESS on average, but did exhibit local variations caused by uncertainty in the water‐fat separation.
Conclusion
In the same acquisition time as DESS, MESS provides water‐fat separation with comparable SNR in the reconstructed water and fat images. By providing additional image contrasts in addition to the water‐selective DESS images, MESS provides a promising alternative to DESS.
Purpose
The development of advanced estimators for intravoxel incoherent motion (IVIM) modeling is often motivated by a desire to produce smoother parameter maps than least squares (LSQ). Deep neural ...networks show promise to this end, yet performance may be conditional on a myriad of choices regarding the learning strategy. In this work, we have explored potential impacts of key training features in unsupervised and supervised learning for IVIM model fitting.
Methods
Two synthetic data sets and one in‐vivo data set from glioma patients were used in training of unsupervised and supervised networks for assessing generalizability. Network stability for different learning rates and network sizes was assessed in terms of loss convergence. Accuracy, precision, and bias were assessed by comparing estimations against ground truth after using different training data (synthetic and in vivo).
Results
A high learning rate, small network size, and early stopping resulted in sub‐optimal solutions and correlations in fitted IVIM parameters. Extending training beyond early stopping resolved these correlations and reduced parameter error. However, extensive training resulted in increased noise sensitivity, where unsupervised estimates displayed variability similar to LSQ. In contrast, supervised estimates demonstrated improved precision but were strongly biased toward the mean of the training distribution, resulting in relatively smooth, yet possibly deceptive parameter maps. Extensive training also reduced the impact of individual hyperparameters.
Conclusion
Voxel‐wise deep learning for IVIM fitting demands sufficiently extensive training to minimize parameter correlation and bias for unsupervised learning, or demands a close correspondence between the training and test sets for supervised learning.
Purpose
To develop and evaluate a patch‐based convolutional neural network (CNN) to generate synthetic computed tomography (sCT) images for magnetic resonance (MR)‐only workflow for radiotherapy of ...head and neck tumors. A patch‐based deep learning method was chosen to improve robustness to abnormal anatomies caused by large tumors, surgical excisions, or dental artifacts. In this study, we evaluate whether the generated sCT images enable accurate MR‐based dose calculations in the head and neck region.
Methods
We conducted a retrospective study on 34 patients with head and neck cancer who underwent both CT and MR imaging for radiotherapy treatment planning. To generate the sCTs, a large field‐of‐view T2‐weighted Turbo Spin Echo MR sequence was used from the clinical protocol for multiple types of head and neck tumors. To align images as well as possible on a voxel‐wise level, CT scans were nonrigidly registered to the MR (CTreg). The CNN was based on a U‐net architecture and consisted of 14 layers with 3 × 3 × 3 filters. Patches of 48 × 48 × 48 were randomly extracted and fed into the training. sCTs were created for all patients using threefold cross validation. For each patient, the clinical CT‐based treatment plan was recalculated on sCT using Monaco TPS (Elekta). We evaluated mean absolute error (MAE) and mean error (ME) within the body contours and dice scores in air and bone mask. Also, dose differences and gamma pass rates between CT‐ and sCT‐based plans inside the body contours were calculated.
Results
sCT generation took 4 min per patient. The MAE over the patient population of the sCT within the intersection of body contours was 75 ± 9 Hounsfield Units (HU) (±1 SD), and the ME was 9 ± 11 HU. Dice scores of the air and bone masks (CTreg vs sCT) were 0.79 ± 0.08 and 0.70 ± 0.07, respectively. Dosimetric analysis showed mean deviations of −0.03% ± 0.05% for dose within the body contours and −0.07% ± 0.22% inside the >90% dose volume. Dental artifacts obscuring the CT could be circumvented in the sCT by the CNN‐based approach in combination with Turbo Spin Echo (TSE) magnetic resonance imaging (MRI) sequence that typically is less prone to susceptibility artifacts.
Conclusions
The presented CNN generated sCTs from conventional MR images without adding scan time to the acquisition. Dosimetric evaluation suggests that dose calculations performed on the sCTs are accurate, and can therefore be used for MR‐only radiotherapy treatment planning of the head and neck.
The aim of this study was to investigate the influence of variable density and data-driven k-space undersampling patterns on reconstruction quality for compressed sensing (CS) magnetic resonance ...imaging to provide recommendations on how to avoid suboptimal CS reconstructions.
First, we investigated the influence of randomness and sampling density on the reconstruction quality when using random variable density and variable density Poisson disk undersampling. Compressed sensing reconstructions on 1 knee and 2 brain data sets were compared with fully sampled data sets and reconstruction errors were measured. Sampling coherence was evaluated on the undersampling patterns to investigate whether there was a relation between this coherence measure and reconstruction error.Second, we investigated whether data-driven undersampling methods could improve reconstruction quality when 1 or more fully sampled scans are available as a training set. We implemented 3 different data-driven undersampling methods: (1) Monte Carlo optimization of variable density and variable density Poisson disk undersampling, (2) calculating sampling probabilities directly from the k-space power spectra of the training data, and (3) iterative design of undersampling patterns based on CS reconstruction errors in k-space.Two cross-validation experiments were set up using retrospective undersampling to evaluate the 3 data-driven methods and the influence of the size of the training set. Furthermore, in an experiment that included prospective under sampling, we show the practical applicability of 2 of the data-driven methods. Compressed sensing reconstruction quality was measured with both the normalized root-mean-square error metric and the mean structural similarity index measure.
Different optimal variable sampling densities were found for each of the data sets, showing that the optimal sampling density is data dependent. Choosing a sampling density other than the optimal density decreased reconstruction quality. These results suggest that choosing a sampling density without having any reference scans is likely suboptimal. Furthermore, no meaningful correlation was found between sampling coherence and reconstruction error.For the data-driven methods, the iterative method yielded statistically significantly higher reconstruction quality in both retrospective and prospective experiments. In retrospective experiments, the power spectrum method yielded a reconstruction quality that was comparable with the data-driven variable density method. The size of the training set had only a minor influence on the reconstruction quality.
Data-driven undersampling methods can be used to avoid suboptimal reconstruction quality in CS magnetic resonance imaging, provided that at least 1 fully sampled scan is available to train the data-driven method. The iterative design method resulted in the highest reconstruction quality.
Display omitted
•Satisfactory synthetic CT images were derived from planning T1w and T2w MR images.•Deep learning-based MRI-only radiotherapy is feasible in pediatric abdominal tumors.•CT-sCT dose ...differences were clinically acceptable (<2%) for photon & proton plans.•Larger differences were caused by existing interscan differences (eg bowel filling)
To assess the feasibility of magnetic resonance imaging (MRI)-only treatment planning for photon and proton radiotherapy in children with abdominal tumours.
The study was conducted on 66 paediatric patients with Wilms’ tumour or neuroblastoma (age 4 ± 2 years) who underwent MR and computed tomography (CT) acquisition on the same day as part of the clinical protocol. MRI intensities were converted to CT Hounsfield units (HU) by means of a UNet-like neural network trained to generate synthetic CT (sCT) from T1- and T2-weighted MR images. The CT-to-sCT image similarity was evaluated by computing the mean error (ME), mean absolute error (MAE), peak signal-to-noise ratio (PSNR) and Dice similarity coefficient (DSC). Synthetic CT dosimetric accuracy was verified against CT-based dose distributions for volumetric-modulated arc therapy (VMAT) and intensity-modulated pencil-beam scanning (PBS). Relative dose differences (Ddiff) in the internal target volume and organs-at-risk were computed and a three-dimensional gamma analysis (2 mm, 2%) was performed.
The average ± standard deviation ME was −5 ± 12 HU, MAE was 57 ± 12 HU, PSNR was 30.3 ± 1.6 dB and DSC was 76 ± 8% for bones and 92 ± 9% for lungs. Average Ddiff were <0.5% for both VMAT (range −2.5; 2.4%) and PBS (range −2.7; 3.7%) dose distributions. The average gamma pass-rates were >99% (range 85; 100%) for VMAT and >96% (range 87; 100%) for PBS.
The deep learning-based model generated accurate sCT from planning T1w- and T2w-MR images. Most dosimetric differences were within clinically acceptable criteria for photon and proton radiotherapy, demonstrating the feasibility of an MRI-only workflow for paediatric patients with abdominal tumours.
PDF
