MRI-only Stopping Power Ratio using Deep Learning for Ion-Beam Therapy Application

In ion beam therapy, stopping power ratio (SPR) is required for estimating the dose distribution and ion range. Currently, it is estimated in clinical routines using single-energy CT. SPR derived from single-energy CT leads to large uncertainties in the dose distribution, as no material-specific information is available, and the different patient geometries are not taken into account. Therefore, additional safety margins are applied at the cost of high doses to the surrounding healthy tissue. With the introduction of dual-energy CT (DECT), the uncertainties in the ion range can be better managed as it has the remarkable ability to capture material- and tissue-specific information required to create an accurate SPR map. In DECT imaging, two CT scans are acquired with different X-ray spectra, allowing the relative electron density and ionisation potential to be determined simultaneously. Exclusively planning radiation treatment using MRI has proven beneficial in the most demanding radiotherapy workflows, as it bypasses MRI-CT registration and avoids unnecessary radiation dose exposure (as CT is omitted), allows more accurate delineation of organs-at-risk, reduces financial costs and facilitates rapid online dose adjustment as with the MRI linear accelerator. However, MRI intensities are not correlated with physical properties such as electron density, so tissue heterogeneity corrections required for dose calculation are impossible. Therefore, MRI images can only be used directly in treatment planning with further processing. The established approach to overcome this limitation is the synthetic generation of CT from MRI images, which serve as a surrogate for dose calculation. Among the established techniques for MRI-CT synthesis, the deep learning-based approach can process any MRI sequence as input and generate an accurate pseudo-CT with continuous Hounsfield unit values in a very short time (about a few seconds), meeting the time requirements for online-only MRI workflows. The main contribution of this work is the investigation of pseudo-SPR maps from Dixon MRI on 25 patients with head and neck cancer. Due to the considerable uncertainties arising from a SECT-derived SPR, a DECT-derived SPR was used to generate the pseudo-SPR map in this study. Four experimental cases were used to investigate the different combinations of Dixon MRI sequences on the resulting pseudo-SPR using a 3D U-Net architecture. Of the defined experimental cases, the single input of the Dixon In-phase MRI resulted in a more reliable pseudo-SPR than the Dixon Water MRI, which can be attributed to the fat signal suppressed in the Dixon Water MRI sequence. From this observation, it can be inferred that the fat signal plays a vital role in accurately assigning HU numbers to account for tissue heterogeneity in the predicted pseudo-SPR. As a result, the predicted pseudo-SPR resulted in a visible increase in performance after the Dixon Water and Fat combination was used as the input MRI. Also, since distinguishing air from bone on MRI images is challenging, a CT-based bone segmentation was integrated into the U-Net architecture as an additional input channel. This investigation yielded an enhanced bone identification on the predicted pseudo-SPR, thereby improving its overall quality. Furthermore, assessment of the prediction accuracy of organs at risk (spinal cord, carotid artery, thyroid and oesophagus) based on the pseudo-SPR and the target SPR resulted in more reliable segmentation on the pseudo-SPR than on the target SPR map, which can be attributed to the influence of MRI to incorporate soft tissue contrast. In addition, the presence of metal implants affected the predicted pseudo-SPR to varying degrees. For this reason, the robustness of the proposed model can be further improved to account for such devices by exposing the network to a more diverse training cohort. Besides, the proposed solution approach significantly outperformed both Cycle-GAN and Label-GAN from a previously published work, although this study was trained on a very limited dataset. Finally, to measure the clinical relevance of this study, a photon-based treatment planning for pseudo-CT was performed using the RayStation treatment planning system. The evaluation showed that the pseudo-CT was comparable to the target CT in terms of quality and dose difference, especially for rigid regions (such as the head), with significant variations observed in the region below the neck due to the inherent registration error between MRI and CT. In conclusion, this study revealed a new optimised training strategy for deep-learning-based style transfer: the inclusion of fat and water signals of Dixon MRI plays a vital role in correctly translating MRI to pseudo-SPR for soft tissue regions. Due to the intrinsic challenge of distinguishing air from bone on most MRI sequences, a CT-based bone segmentation utilised in the U-Net architecture significantly improves bone identification on the resulting pseudo-SPR. This newly developed method outperformed previously published state-of-the-art deep learning approaches. Finally, the proposed solution subjected to various degrees of evaluation strategies has proven to be robust, and a preliminary clinical dosimetric evaluation yielded an acceptable pseudo-CT compared to its corresponding target CT in terms of quality and dose difference for a photon-based treatment planning, particularly for rigid regions such as the head. Regardless, the clinical dosimetric evaluation of pseudo-SPR is limited in this study due to the lack of a standard clinical DECT-converted SPR map for ion-beam treatment planning. Therefore, as a further direction, the predicted pseudo-SPR can be validated with a clinical DECT-converted SPR map, enabling the dosimetric analysis of clinical target volumes and organ-at-risk to be feasible in the treatment planning of ion-beam therapy.