The link between cardiac and renal pathology is a well-recognized especially in chronic kidney disease (CKD) patients, MRI playing an important role to investigate the connection between both organs in medical research as well as clinical practice. However, MRI suffers from limitations considering the long acquisition time of cardiac MRI sequences and the tedious and largely manual analysis of renal images. The focus of this thesis is, therefore, to explore novel methods to facilitate the use of cardiac and renal MRI. The first part of this thesis aims to accelerate the acquisition and reconstruction of 2D cine cardiac MRI, and the second part, to streamline and automate the analysis of renal MRI data for CKD patients.
To speed-up cardiac MRI, two acceleration methods were investigated. First, the sc-GROG k-t ESPIRiT with Total Variation (TV) constraint method for reconstructing unaliased cardiac real-time, multi-slice, single breath-hold 2D cine radial MR images. The results were compared to the NUFFT k-t ESPIRiT approach. The second work proposed an image-based modified deep learning radial cardiac U-Net (RC-Net) using scGROG and NUFFT for the efficient reconstruction of real-time, multi-slice, single breath-hold, highly accelerated radial 2D cine MR data that could potentially speed up the long post-processing time considered as a drawback of ESPIRiT reconstructions. Furthermore, a transfer learning approach (TLRC-Net) was presented to address the limited training data issue, to reduce training time and to overcome the generalization problem of MR data across different acceleration factors (AFs). The experimental results in this study were rigorously evaluated using AP, RMSE, SSIM, PSNR.
The second part of the thesis aimed to accelerate renal MRI analysis. The manual renal contouring to measure cortical and medullary signal intensity on the MRI is a difficult and time-consuming task as observed during an active participation to clinical studies. In a first step, this work proposed a deep learning-based RCM U-Net to automatically segment the renal cortex and medulla from T1 maps. A correlation analysis between the automatically measured cortico-medullary difference (ΔT1) values measured by the RCM U-Net, clinical manual regions of interest (ROI) values, eGFR, and percentage fibrosis was performed. In second step, a multi-class supervised deep learning based MdU-Net was proposed to automatically segment the renal cortex and medulla from ADC maps of allograft CKD kidneys. The correlations between MdU-Net measured ΔADC values were significantly correlated with 3D Slicer manual ground truth values and clinical ROI.
In conclusion, this thesis introduced significant development in cardiac MRI and renal MRI segmentation that could promote the use of MRI in patients with cardio-renal syndrome.
