Tiago Fernando Ferreira da Silva
One of the foremost and challenging subfields of MRI is cardiac magnetic resonance imaging (CMR). CMR is becoming an indispensable tool in cardiovascular medicine by acquiring data about anatomy and function simultaneously. For instance, it allows the non-invasive characterization of myocardial tissues via parametric mapping techniques. Parametric mapping techniques provide a spatial visualization of quantitative changes in the myocardial parameters. Several T1 and T2 mapping sequences have been developed via combinations of pulse acquisitions with signal modeling in order to quantify the relaxation parameters accurately and obtain quantitative relaxation maps that provide pixel-wise information regarding the relaxation parameters. The quantification of the relaxation parameters is a fascinating application of CMR since it is a non-invasive way to evaluate disease severity and response to therapy. All in all, parametric mapping provides unique and essential information about the myocardium and is a significant player towards precision medicine.
Groundbreaking developments such as 3D techniques, parallel imaging, compressed sensing, and AI have continuously advanced the CMR field. These developments provide new tools to upgrade and improve parametric mapping sequences, en route to decrease acquisition time, improve spatial resolution, and increase the coverage of the heart. Inspired by the need to develop novel high-quality parametric sequences for 3T, this thesis's primary goal is to introduce an accurate and efficient 3D single breath-hold MR methodology for measuring parametric cardiac mapping on a 3T clinical scanner.
The thesis is divided into two main parts: i) research and development of a new 3D T1 saturation recovery mapping technique, named 3D saturation-recovery compressed SENSE rapid acquisition (3D SACORA), together with a feasibility study regarding the possibility of adding a T2 mapping feature to 3D SACORA concepts and ii) research and implementation of a deep learning-based post-processing method to improve the T1 maps obtained with 3D SACORA.
In the first part of the thesis, 3D SACORA was developed as a new 3D T1 mapping sequence to speed up T1 mapping acquisition of the whole heart. The proposed sequence was validated in phantoms against the gold standard technique (IR-SE) and in-vivo against the previously proposed sequence 3D SASHA. The 3D SACORA pulse sequence design was focused on acquiring the entire left ventricle in a single breath-hold while achieving good quality T1 mapping and stability over a wide range of heart rates (HRs). 3D SACORA consists of three distinct blocks. The sequence begins with proton density (PD) images to avoid T1 effects from prior saturation pulses (first block). Full signal recovery is ensured by waiting a minimum of 6 s between the PD readouts. The second block consists of images acquired at saturation time (TS) 1 and at TS1+RR interval (TS3). Finally, the third block of images is acquired at TS2 and at TS2+nRR intervals (TS4). T1 estimation is optimized not only by the PD images, but also by carefully selecting the acquired saturation time images. TS1 is set to 250 ms to acquire the shortest possible saturation image while maintaining an adequate SNR. For most heart rates, TS2 is set to 500 ms to allow a TS4 similar to the native cardiac T1 (1500 ms at 3T). For high heart rates that make a 500-ms TS2 unfeasible, the sequence automatically computes the longest possible TS2 according to the RR interval derived from the heart rate defined by the user in the acquisition protocol. Full image acquisition is accelerated using a k-space shutter and a spatial domain compressed SENSE factor of 4.5 in order to acquire a whole 3D volume in 2 independent TFE shots of echo train length of 76. For estimating the T1 values, all sampling points are corrected according to the linear readout, and sampling points TS3 and TS4 are further corrected for the magnetization distortion caused by the readout effects of TS1 and TS2, respectively.
The precision and accuracy of 3D SACORA were assessed in phantom experiments. The phantom consisted of eight tubes filled with distilled water and different concentrations of an MR contrast agent selected to obtain T1s spanning the range from 355 ms to 1871 ms. Reference T1 values were obtained using IR-SE. In order to validate further 3D SACORA T1 estimation accuracy and precision, T1 values were also estimated using an in-house version of 3D SASHA. This 3D SASHA variant used a linear spoiled T1-TFE readout without any respiratory trigger for phantom and in-vivo experiments. T1 values (mean ± standard deviation) were measured from manually drawn regions of interest (ROIs) in the generated T1 maps. Passing-Bablok regression and Bland-Altman plots were used to assess correlation and agreement between the three sequences. To explore the proposed methodology's sensitivity to heart rate, we acquired images with 3D SACORA and 3D SASHA at simulated heart rates ranging from 50 to 120 bpm. The estimated T1s were then compared against the T1 reference values obtained with IR-SE.
For in-vivo validation, seven large healthy pigs were scanned with 3D SACORA and 3D SASHA. In all pigs, images were acquired before and after administration of MR contrast agent. Bulls-eye plots of T1 and coefficient of variation were generated from 3D SACORA and 3D SASHA data according to the American Heart Association (AHA) standard 17-segment model of the LV. Septal T1 values (mean ± standard deviation) were measured from ROIs manually drawn on the septal myocardium. These measurements were used to compare the septal T1 values and coefficient of variation obtained by both sequences.
The Passing-Bablok regression plots show a good correlation and no significant bias between methods. The Bland-Altman plots show good agreement between both methods and IR-SE. 3D SACORA showed less dependence on heart rate, particularly for pre-contrast cardiac T1 values at 3T; estimated T1 values with 3D SASHA tended to increase with higher heart rate, whereas 3D SACORA T1s seemed to be relatively independent of heart rate. Regarding accuracy, both sequences performed equally well. Although precision was good with both sequences, 3D SASHA performed slightly better, particularly on short T1s. The coefficient of variation appeared to be independent of the heart rate in both sequences.
In the in-vivo experiments, all T1-weighted images showed good contrast and quality, and the T1 maps correctly represented the information contained in the T1-weighted images. Bulls-eye plots of mean T1 values showed homogeneity across the LV myocardium for both sequences. The bulls-eye plots for the coefficient of variation showed good precision in measuring T1 values in myocardium and blood for both sequences. Septal T1s and coefficients of variation did not considerably differ between the two sequences, confirming good accuracy and precision. Mean septal native and post-contrast T1s measured with 3D SACORA were 1453 ± 44 ms and 824 ± 66 ms, respectively. For 3D SASHA, the mean septal native T1 was 1460 ± 60 ms and the mean septal post-contrast T1 was 824 ± 60 ms. The mean coefficient of variation for native septal T1 was 0.041 ± 0.010 for 3D SACORA and 0.039 ± 0.010 for 3D SASHA. The post-contrast values were 0.050 ± 0.008 and 0.041 ± 0.008, respectively. 3D SACORA images showed good contrast, homogeneity, and were comparable to corresponding 3D SASHA images, despite the shorter acquisition time (15s vs. 188s, for a heart rate of 60 bpm).
In conclusion, the proposed 3D SACORA successfully acquired a whole-heart 3D T1 map in a single breath-hold at 3T, estimating T1 values in agreement with those obtained with the IR-SE and 3D SASHA sequences. 3D SACORA sequence acquired pre-contrast and post-contrast T1 maps of the whole heart with good accuracy, precision, and image quality for LV analysis at 3T. The sequence was optimized for speed and can acquire a 3D T1 map in 15 heartbeats for a heart rate of 60 bpm.
Following the successful validation of 3D SACORA, a feasibility study was performed to assess the potential of modifying the acquisition scheme of 3D SACORA in order to obtain T1 and T2 maps in a single breath-hold. Although this adds in a whole new level of complexity, the possibility of using 3D SACORA concepts to acquire simultaneous, co-registered 3D T1 and T2 maps is exciting and could be groundbreaking. In this way, a new 3D T1/T2 sequence, named 3D dual saturation-recovery compressed SENSE rapid acquisition (3D dual-SACORA), was built on the main concepts of 3D SACORA.
The 3D T1/T2 sequence proposed for this feasibility study is designed to acquire T2-weighted images with excellent signal by not acquiring these images in the same RR interval of the saturation pulse, but by acquiring these images with an n amount of recovery beats depending on the heart rate. The acquisition scheme can be divided into three blocks. In the first block, the proton density (PD) is acquired to avoid effects from previous saturation pulses. The second block consists of the acquisition of TS1 and TS2. TS1 is set to 300 ms. TS2 is the maximum saturation time allowed by the heart rate. The third block is focused on the T2 mapping part of the acquisition. Here, TS3 and TS4 are calculated to be always acquired as close as possible to the native cardiac T1 (1500 ms at 3T), and therefore the amount of recovering beats changes according to the heart rate. These two T2 weighted images are acquired with T2- prep echo times of 25 and 45 ms, respectively. The relaxation parameters T1 and T2 are estimated jointly with a three-parameter combined signal model.
A phantom of eight tubes was built to validate the 3D dual-SACORA sequence. The phantom consisted of five tubes for T1 values between 338 and 1819 ms and three tubes for T2 values between 19 and 104 ms. The phantom was scanned with the 3D dual-SACORA sequence with a simulated heart rate of 60 bpm. Under the same conditions, reference T1 and T2 values were estimated using IR-SE and gradient and spin-echo (GraSE) sequence, respectively. Mean and standard deviation were measured from the parametric maps with manually drawn regions of interest. The 3D dual-SACORA sequence was compared against the reference sequences with linear regression analysis. The coefficient of variation (CV) and relative error were calculated for assessing precision and accuracy. An in-vivo study was performed with a healthy volunteer to evaluate the parametric maps' image quality obtained with 3D dual-SACORA. T1 and T2 values (mean ± standard deviation) were measured from ROIs manually drawn on the septal myocardium and blood pool.
T1 and T2 maps of the phantom were successfully obtained with the 3D dual-SACORA sequence. The linear regression plots showed good agreement between the T1 and T2 values obtained by the proposed sequence and IR-SE and GraSE, respectively. However, the proposed sequence seemed to underestimate the very long T2s. The coefficient of variation and relative error plots show that the 3D dual-SACORA sequence achieved good precision and accuracy for most values. As already seen in the linear regression analysis, the relative error also suggests that the 3D dual-SACORA sequence underestimated the very long T2s (104 ms).
A volunteer was successfully scanned with the 3D dual-SACORA sequence (acquisition duration of approximately 20s) in a single breath-hold. T1-weighted images (TS1 and TS2) and T2-weighted images (TS3 and TS4) showed good contrast and quality. Although TS4 has a T2 preparation of 45 ms, the image's SNR and quality were excellent as it was acquired with a recovery beat (TS4 = 1603 ms). The parametric maps accurately represented the information contained in the saturation time images. The maps from the in-vivo experiments showed good contrast and homogeneity. The mean septal T1 of the three slices was 1443.7 ± 51.3 ms, while the mean septal T2 of the three slices was 51.7 ± 2.8 ms. The septal T1 and T2 values are in good agreement with reference sequences and published work. The T2-weighted images presented excellent quality, showing that the use of recovering beats for acquiring T2-weighted images in saturation recovery sequences is auspicious and a valuable idea.
In conclusion, this feasibility study's findings open the door to the possibility of using 3D SACORA concepts to develop a successful 3D T1/T2 sequence. Although the 3D dual-SACORA sequence has room for improvement, the results show that a large study based on the proposed sequence is welcome and can be groundbreaking. To the best of our knowledge, the proposed sequence is the first saturation recovery sequence using recovering beats for acquiring T2-weighted images.
In the second part of the thesis, a deep learning-based super-resolution model was implemented to improve the image quality of the T1 maps of 3D SACORA, and a comprehensive study of the performance of the model in different MR image datasets and sequences was performed.
After careful consideration, the selected convolutional neural network to improve the image quality of the T1 maps was the Residual Dense Network (RDN). This network can make full use of all the image hierarchical features (local features and global features) and has shown outstanding performance against state-of-the-art methods on benchmark datasets; however, it has not been validated on medical datasets. In this way, the model was initially validated on two benchmark medical datasets, one containing brain MR images and the other containing cardiac MR images, achieving a reasonable assessment of the model's robustness. After this validation, a self-acquired cardiac dataset was carefully put together, and the RDN model was validated on improving the T1 maps of 3D SACORA and 3D SASHA.
The high-resolution subsets of the datasets include different 2D slices obtained from several high-resolution 3D acquisitions. The low-resolution subsets were obtained by downsampling the high-resolution images with the Lanczos algorithm. The datasets prepared to evaluate the model's performance on improving MR images' quality have a subset for testing. This subset's images are used to compare the RDN model results against a state-of-the-art method. In this way, the testing subset's low-resolution images were upsampled with the trained model and Lanczos resampling. Afterward, these upsampled images were compared with the original high-resolution images using subjective and objective metrics of image quality assessment to evaluate the model's performance quantitatively. After showing the performance of the trained models in increasing the quality of MR images, the goal was to evaluate the quality of the T1 maps estimated through T1-weighted images upsampled by the trained model. In this case, no ground-truth image is available because the only images available are the T1-weighted images acquired with the T1 mapping sequences; therefore, the acquired images were upsampled using both the trained model and the state-of-the-art Lanczos resampling in order to compare the T1 maps obtained by the RDN model with a reference method. The acquired images were upsampled with the trained model and the Lanczos resampling, and the resulting upsampled images were fitted to obtain the T1 maps following the respective Bloch equations modeling of 3D SACORA and 3D SASHA. The computed T1 maps were visually assessed to evaluate the performance of both methods. ROIs were drawn in the septal myocardium and blood pool of 20 slices (5 per pig) to analyze the upsampling effect on the T1 values. The mean T1 and the standard deviation were obtained from the ROIs. Furthermore, line ROIs were drawn crossing the myocardium to elaborate a profile of the pixel values on the edges between myocardium and blood and evaluate the resolution. The model improved the images successfully for the two benchmark datasets, achieving better performance with the brain dataset than with the cardiac dataset. This result was expected as the brain dataset's high-resolution images presented a better resolution and quality than those of the cardiac dataset. Additionally, the brain images have more well-defined edges than the cardiac images, making the resolution enhancement more evident. Then, on the self-acquired cardiac dataset, the trained RDN model also obtained results with improved image quality assessment metrics and visual assessment, particularly on well-defined edges. These results validated the use of these acquired high-resolution images to improve the saturation time images.
Regarding the improvement of the T1 maps of 3D SACORA and 3D SASHA, the model improved the image quality of the saturation time images and the T1 maps. The model was able to decrease the noise and eliminate motion artifacts, enhancing the T1 maps analytically and visually. Analytically, the model did not considerably modify the T1 values while improving the standard deviation in both myocardium and blood. Visually, the model improved the T1 maps by removing the noise and motion artifacts without losing resolution on the edges. The enhancement of the saturation time images and T1 maps seemed to be more related to the denoising and artifacts correction than to the upsampling, which may be explained by the already decent resolution of the saturation time images. The capability of the model to learn that the motion artifacts should not be present on the images is an exciting and auspicious result.
In conclusion, a deep learning-based post-processing method (RDN model) was used to improve the image quality of MR images with the end goal of improving the T1 maps obtained with 3D SACORA. A comprehensive study of the model's performance in different MR image datasets and sequences showed its capability to improve MR images and T1 maps by increasing resolution, reducing noise, and eliminating artifacts. The model was able to take 3D SACORA T1 maps' quality to a whole new level by correcting many of the issues of acquiring 3D T1 maps in a single breath-hold, opening the door to the possibility of using deep learning-based post-processing models as a standard MR image enhancement approach.
In summary, a 3D single breath-hold MR methodology was introduced for measuring parametric cardiac mapping on a 3T clinical scanner. This methodology includes a ready-to-go 3D single breath-hold T1 mapping sequence for 3T (3D SACORA); the ideas for an upgrade of 3D SACORA, which led to a very promising 3D T1/T2 mapping sequence (3D dual-SACORA); and a deep learning-based post-processing implementation able to improve the image quality of 3D SACORA T1 maps.
© 2001-2024 Fundación Dialnet · Todos los derechos reservados