Real time on-the-fly dynamic MRI for radiotherapy
External beam radiotherapy (EBRT) utilizes high energy radiation (primarily photons or electrons) generated from a linear accelerator (linac) for the treatment of cancer. Current linac technologies lack the ability to continuously monitor the tumour and surround area during treatment. Real-time imaging in combination with EBRT is ideal, as it can allow for adaptive radiotherapy, or advanced gating to reduce the radiation damage to healthy surrounding tissues and organs at risk. To achieve this,
... several groups around the world have developed systems that combine a linac with a magnetic resonance imaging (MRI) system (LMR). In fact, many LMRs have already been installed around the world and have already started clinical trials. With the increasing popularity of LMR systems, real-time MR imaging is becoming an important aspect for achieving adaptive radiotherapy or advanced gating techniques. Real-time on-the-fly MRI presents a challenge, given the length of time it takes to acquire fully sampled MR data. Speeding up the acquisition can be achieved via the undersampling of k-space; however, this violates the Nyquist criterion resulting in aliasing artefacts occurring through the reconstructed image. Therefore, there exists a trade-off between image quality and image acquisition speed. The work presented within this thesis investigates MR image reconstruction techniques that are able to speed up the data acquisition (via k-space undersampling), while preserving the image fidelity; allowing for real-time MRI. In particular, this work focuses on image reconstruction for patients with non-small cell cancer (lung tumours), as this provides a challenging scenario of tracking a tumour during involuntary breathing motion. Two novel techniques were developed for real-time MR imaging. Compressed sensing principal component analysis (CS-PCA), which uses an incoherent phase encoding ii scheme to undersample k-space. Principal component analysis is conducted on previously acquired data, to generate principal components that are used to fill the missing k-space in order to rapidly reconstruct undersampled data in under 10 ms per image. Secondly, 3D convolutional neural networks (CNNs) were used to reconstruct undersampled data. Two phase encoding schemes were investigated, incoherent undersampling and coherent low-resolution undersampling. A CNN was trained for each patient individually, making it-what we have termed-a patient-specific CNN. The CNN utilized a cascading network of convolutional layers and data consistency layers for reconstruction. Training the CNN took approximately 6 hours to using our hardware, which consisted of an Intel Xeon E5-2650 CPU and NVIDIA GTX 1080Ti GPU. The image reconstruction time using the trained CNN parameters was less than 65 ms per image. iii Throughout the duration of my PhD I have accomplished many milestones in my life.