Doctor of Philosophy (Ph.D.)
Degree Granting Department
Wilfrido Moreno, Ph.D.
Natarajan Raghunand, Ph.D.
Ashwin Parthasarathy, Ph.D.
Andres Tejada, Ph.D.
Christos Ferekides, Ph.D.
Eduardo Moros, Ph.D.
Image Reconstruction, Intratumoral Habitats, Joint Total Variation, Radiation Dose Maps, Tissue Phenotyping
MR guided Radiotherapy (MRgRT) marks an important paradigm shift in the field of radiotherapy. Superior tissue contrast of MRI offers better visualization of the abnormal lesions, as a result precise radiation dose delivery is possible. In case of online treatment planning, MRgRT offers better control of intratumoral motion and quick adaptation to changes in the gross tumor volume. Nonetheless, the MRgRT process flow does suffer from some challenges that limit its clinical usability. The primary aspects of MRgRT workflow are MRI acquisition, tumor delineation, dose map prediction and administering treatment. It is estimated that the acquisition of MRI takes around 50% of the entire process. Further, delineating the tumor volumes and generating the dose map plans are labor-intensive and time-consuming yet necessary to prevent radio necrosis and associated toxicity. To this end, this dissertation focuses on the two important aspects of MRgRT. First, acceleration of reconstruction of multiparametric MRI (mpMRI). Second, prediction of precise dose maps from the pre-radiation therapy mpMRI sequences without the need of manual contouring.
A joint reconstruction algorithm to accelerate the reconstruction of a series of complex T1w images, T1 and proton density maps simultaneously from the undersampled k-space data is presented. The ambiguity introduced by undersampling is resolved using model-based constraints, and structural information from a reference fully sampled image as the joint total variation prior. The algorithm is extended with minor modifications to accelerate the reconstruction of complex T2w, T2*w images and their parameter maps. Validation of the reconstructed images and parameter maps was carried out by computing tissue-type maps, as well as the maps of the Proton Density Fat Fraction (PDFF), Proton Density Water Fraction (PDwF), fat relaxation rate R_2f^* and water relaxation rate R_2w^* from the reconstructed data by comparing them with Ground Truth (GT) equivalents. It is demonstrated that using only 18% k-space data, it is possible to identify the tissue type maps like fluid, muscle, tumor and adipose with the same fidelity as that obtained using GT data. The mean T1 and T2 values in each tissue type were computed using only 18% k-space data, which were within 8%-10% of the GT values from fully sampled data. The PDFF and PDwF maps computed using 27% k-space data were within 3%-15% of GT values and showed good agreement with the expected values for the four tissue types. The next task focuses on directly predicting the optimum Radiation Therapy (RT) dose maps from the pre-RT mpMRI. It is now well established that the tumor volume comprises several different microenvironments. Hence, predicting a voxel-wise dose map from the pre-RT and prescribed/desirable post-RT mpMRI will yield better control of radionecrosis-related toxicity. Furthermore, it is also important for the radiation oncologist to simulate voxel-wise radiologic outcomes of specific RT dose map prescriptions on post-RT mpMRI. To accomplish these two tasks, end-to-end deep neural networks are trained. The forward model is used to predict post-RT changes on mpMRI using pre-RT mpMRI when administered with the radiation dose map. A variant of the pix2pix GAN network is trained to predict post-RT ADC maps, T1wCE, T2w, T1w, FLAIR MRI from pre-RT mpMRI and the radiation dose maps. The results of the forward model are validated by identifying the tissue type maps like blood volume, gray matter, white matter, edema, non-enhancing tumor, contrast enhancing tumor, hemorrhage, fluid and comparing them with the GT maps. Further, the quantitative validation is carried out by comparing the percentage of volumes of these tissue type maps from pre-RT, post-RT and predicted post-RT mpMRI. The results of the forward model are also tested with the simulated dose maps and comparing the changes on the predicted post-RT ADC maps that are mechanistically relatable to voxel-level tumor response to therapies. Next, a variant of pix2pix GAN is trained to predict the radiation dose maps from the pre-RT ADC maps and the prescribed post-RT ADC maps. This is called as the inverse model. It is determined from the simulated results that to achieve higher ADC values, higher RT dose maps are required.
In summary, the results of the feasibility study showed that it is possible to identify various tissue type habitats from the reconstructed mpMRI scans using only 18% k-space data. This dissertation also highlights that it is possible to alleviate the manual aspects of Radiation Therapy planning by using pre-RT and post-RT mpMRIs to predict the Radiation dose maps.
Scholar Commons Citation
Pandey, Shraddha, "Accelerating Multiparametric MRI for Adaptive Radiotherapy" (2022). USF Tampa Graduate Theses and Dissertations.