Integrating Deep Learning into Magnetic Resonance-based Thermometry: Enhancing Precision in Clinical Hyperthermia

This thesis presents a novel deep learning-based framework for accelerating Magnetic Resonance (MR) imaging and thermometry, with a specific focus on applications in MR-guided hyperthermia. Hyperthermia, a treatment modality for cancer that involves heating tumour tissues, requires precise and real-time temperature monitoring. Magnetic Resonance Thermometry (MRT) provides non-invasive, real-time temperature maps essential for controlling heat delivery during treatment. However, the long acquisition times associated with fully-sampled MR data can introduce motion artefacts and limit the temporal resolution needed for effective therapy monitoring. This research addresses these challenges by proposing and evaluating complex-valued Convolutional Neural Networks (CV-CNNs) for reconstructing MR images from undersampled data. The work demonstrates that the proposed deep learning models, specifically Fourier-PDNet and Fourier-PDUNet, are capable of accurately reconstructing both magnitude and phase images from significantly undersampled MR data, achieving up to 10-fold acceleration. The results highlight that while both models effectively reconstruct magnitude images, they face challenges with phase image reconstruction, leading to minor temperature inaccuracies. In addition to image reconstruction, this thesis also cover a deep learning-based segmentation approach using U-Net models to distinguish between homogeneous and heterogeneous temperature distributions during hyperthermia. This automatic segmentation is crucial for effective monitoring and control of heat delivery in solid cancer treatments. Notably, previous studies have shown that MR thermometry data correlate with pathological response in soft tissue sarcoma. Our findings suggest that incorporating information on temperature distribution patterns—whether homogeneous or heterogeneous—could further refine temperature parameters linked to therapeutic success. From a biological standpoint, a heterogeneous temperature distribution might even be advantageous: lower temperature areas may promote immune cell infiltration, while higher temperature regions could induce immunogenic tumour cell death. This aligns with concepts observed in spatially fractionated radiotherapy, where varying conditions within the tumour microenvironment can enhance treatment efficacy. The study also highlights that CV-CNNs outperform traditional CNNs in complex imaging scenarios, indicating their potential for more accurate phenotype extraction in biomedical research. This capability is particularly relevant for image-driven genetic studies, where precise imaging data is essential for identifying genetic associations and understanding disease mechanisms. The findings underscore the clinical relevance of these techniques, as accelerated MR imaging and improved segmentation can significantly reduce scan times, mitigate motion artefacts, and enhance temporal resolution, thereby improving the precision and effectiveness of hyperthermia treatments. This work lays the foundation for integrating deep learning into clinical MR-guided hyperthermia, offering a pathway towards more efficient and reliable temperature monitoring in cancer therapy.