Date of Award
Doctor of Philosophy (PhD)
AbstractEach year, 2.2 million pediatric head computed tomography (CT) scans are performed in the United States. Head trauma and craniosynostosis are two of the most common pediatric conditions requiring head CT scans. Head trauma is common in children and one-third of the patients that present to the emergency room undergoes head CT imaging. Craniosynostosis is a congenital disability defined by a prematurely fused cranial suture. Standard clinical care for pediatric patients with head trauma or craniosynostosis uses high-resolution head CT to identify cranial fractures or cranial sutures. Unfortunately, the ionizing radiation of CT imaging imposes a risk to patients, particularly pediatric patients who are vulnerable to radiation. Moreover, multiple CT scans are often performed during follow-up, exacerbating their cumulative risk. The National Cancer Institute reported that radiation exposure from multiple head CT scans will triple the risk of leukemia and brain cancer. Many medical centers have recently removed CT from the postoperative care of craniosynostosis, limiting postoperative evaluation and highlighting the urgent need for radiation-free imaging. Several “Black bone” magnetic resonance imaging (MRI) methods have been introduced as radiation-free alternatives. Despite the initially encouraging results, these methods have not translated into clinical practice due to several challenges, including 1) subjective manual image processing; 2) long acquisition time. Due to poor signal contrast between bone and its surrounding tissues in MR images, existing post-processing methods rely on extensive manual MR segmentation which is subjective, prone to noise and artifacts, hard to reproduce, and time-consuming. As a result, they do not meet the need for clinical diagnosis and have not been employed clinically. A CT scan takes tens of seconds; however, a high-resolution MR scan takes minutes, which may be challenging for pediatric subject compliance and limit clinical adoption. The overall objective of this study is to develop rapid and radiation-free 3D high-resolution MRI methods to provide CT-equivalent information in diagnosing cranial fractures and cranial suture patency for pediatric patients. Two specific aims are proposed to achieve the overall objective. Aim 1: Develop a fully automated deep learning method to synthesize high-resolution pseudo-CT (pCT) of pediatric cranial bone from MR images. Aim 2: Develop a deep learning image reconstruction method to reduce MR acquisition time. Aim 1 is to address the issues of subjective manual image processing. In this aim, we developed a robust and fully automated deep learning method to create pCT images from MRI, which facilitates translating MR cranial bone imaging into clinical practice for pediatric patients. Two 3D patch-based ResUNets were trained using paired MR and CT patches randomly selected from the whole head (NetWH) or in the vicinity of bone, fractures/sutures, or air (NetBA) to synthesize pCT. A third ResUNet was trained to generate a binary brain mask using only MRI. The pCT images from NetWH (pCTNetWH) in the brain area and NetBA (pCTNetBA) in the non-brain area were combined to generate pCTCom. A manual processing method using inverted MR images (iMR) was also employed for comparison. pCTCom had significantly smaller mean absolute errors (MAE) than pCTNetWH and pCTNetBA in the whole head. Dice Similarity Coefficient (DSC) of the segmented bone was significantly higher in pCTCom than in pCTNetWH, pCTNetBA, and iMR. DSC from pCTCom demonstrated significantly reduced age dependence than iMR. Furthermore, pCTCom provided excellent suture and fracture visibility comparable to CT. A fast MR acquisition is highly desirable to translate novel MR cranial to clinical practice in place of CT. However, fast MR acquisition usually results in under-sampled data below the Nyquist rate, leading to artifacts and high noise. Recently, numerous deep learning MR reconstruction methods have been employed to mitigate artifacts and minimize noise. Despite many successes, existing deep learning methods have not accounted for MR k-space sampling density variations. In aim 2, we developed a self-supervised and physics-guided deep learning method by weighting k-space sampling Density in network training Loss (wkDeLo). The proposed method uses an unrolled network with a data consistency (DC) and a regularization (R). A forward Fourier model was used to transform the reconstructed image into k-space. The data consistency between the transformed k-space and the acquired k-space data is enforced in the DC layer. This unrolled network is regularized by k-space deep-learning prior using a convolution neural network. In total, 400 radial spokes were acquired with an acquisition time of 5 minutes. Two disjoint k-space data sets, including the first 1 minute (80 radial spokes) and the remaining 4 minutes (320 radial spokes), were used as the network training input and target. A unique feature of our proposed method is to use a L1 loss weighted by k-space sampling density in an end-to-end training of the unrolled network. Moreover, we also reconstructed images using the same unrolled network structure but without accounting for the k-space sampling density variations in the loss for comparison. In other words, a uniform weighted k-space is used in the training loss (un-wkDeLo). Furthermore, we implemented a well-accepted deep learning reconstruction method, Self-Supervision via Data Undersampling (SSDU) as a baseline method reference. Using the images reconstructed from a 5-min scan as the gold standard, we computed the structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) for reconstructed images from 1-min k-space data using SSDU, un-wkDeLo, and wkDeLo. The SSIM and PSNR of the wkDeLo images are significantly higher than both SSDU and un-wkDeLo. Moreover, the wkDeLo reconstructed images have the highest sharpness and the least artifacts and noise. In aim 2, we have demonstrated that high quality MR images at a spatial resolution of 0.6x0.6x0.8 mm3 could be achieved using only 1 min acquisition time. Finally, we evaluated the clinical utility of the proposed MR cranial bone imaging in identifying cranial fractures and cranial suture patency. Clinicians by consensus evaluated the MR-derived pCT images. Acceptable image quality was achieved in greater than 90% of all MR scans; diagnoses were 100% accurate in the subset of patients with acceptable image quality. We have demonstrated that the proposed 3D high-resolution MR cranial bone method provided CT-equivalent images for pediatric patients with head trauma or craniosynostosis. This work will have a profound impact on pediatric health by providing clinicians with a rapid diagnostic tool without radiation safety concerns.
Dennis Barbour, Kamlesh Patel, Ulugbek Kamilov, Monica Shokeen,