ESTRO 2024 - Abstract Book

S3869

Physics - Image acquisition and processing

ESTRO 2024

resection cavity. In clinical practice, pre-treatment imaging for PT involves the acquisitions of a Computed Tomography (CT) scan to acquire the tissues Stopping Power Ratio (SPR) for proton range prediction, and of a Magnetic Resonance Imaging (MRI) for CTV delineation. Dual-energy CT (DECT) is the gold standard for SPR prediction, achieving the lowest uncertainties in particle range. While MRI offers a better soft tissue contrast, a direct conversion from MRI information to SPR is currently unfeasible. The implementation of an MRI-only workflow, directly predicting SPR information from MRI, would omit patient dose from CT imaging, streamline the clinical workflow and thus facilitate fast adaptive PT.

In this study, we apply state-of-the-art deep learning approaches to compute synthetic SPR (sSPR) maps from MRI for primary brain tumour patients, utilizing highly precise, DECT-derived SPR information for training.

Material/Methods:

For 159 patients, combined datasets of DECT-derived SPR and MRI were used. DECT datasets were acquired with a Dual-Spiral DECT scanner Somatom Definition AS (Siemens Healthineers, Forchheim, Germany); SPR maps were generated in clinical routine using the DirectSPR approach (Siemens Healthineers) (1) for 146 patients. Regarding the MRI, they were acquired with the Aera, Avanto, MAGNETOM Lumina, MAGNETOM Vida, Prisma, Skyra, Verio, Verio.Dot (all Siemens Healthineers, Erlangen, Germany), SIGNA HDxt (GE Healthcare, Chicago, United-States), Ingenuity TF PET/MR (Philips Healthcare, Best, Netherlands) for 1, 1, 1, 32, 2, 13, 8, 15, 4 and 77 patients respectively. No scanner information was available for the remaining 5 patients. This resulted in 78 pairs of SPR/T1-weighted MRI and 81 pairs of SPR/contrast-enhanced T1-weighted MRI. Image pre-processing involved an MRI bias field correction with the N4 filter, images B-spline spatial resampling to 1x1x1mm3, a rigid registration of the MRI onto the SPR, a zero mean-unit variance MRI intensities standardization as well as a MRI and SPR intensities rescaling to [0; 1] – for SPR the same global scaling factor was applied for all datasets. 112 and 25 patients were used to train and validate three modified 2D U-Nets, respectively, either based on axial, coronal or sagittal slices. The remaining 22 patients were used as an independent testing set. A 2.5D approach was additionally explored, considering the median of the predictions from the three different views. To evaluate the sSPR maps, the Mean Absolute Error (MAE) was computed within the whole head, further referred as body, as well as for 3 different SPR intervals, namely air, bone and soft tissues volumes. Since SPR are described in HU, with 0 HU corresponding to an SPR of 1 with an increment of 0.001 per 1 HU, the latter three volumes were obtained considering SPR intensities of: £ -150 HU for air, ³ 200 HU for bone and between [20 HU-60 HU] for soft tissues, respectively. Intervals were chosen to assure a clear separation of the different volumes. For the statistical analysis, the MAE mean +/- standard deviation were calculated for the testing cohort. In addition, Wilcoxon tests were performed, and a significance threshold of 5% was set.

Results:

Figure 1 shows the qualitative sSPR performance for the patients with the lowest and highest MAE when using the 2.5D approach. For the worst-scenario case, inaccurate reconstructions were mostly localized in the maxillary sinuses.