ESTRO 38 Abstract book

S1043 ESTRO 38

Each training set was used for constructing a Support Vector Machine (SVM) classifier that was employed to segment the specific aBM subregions on CT images of the five patients. A comparison between PET aBM ROIs and CT ones was carried out with respect to DICE index, precision and recall for elements with at least 64 pixels. Results Fig.1 shows an example of aBM manually drawn on PET and the segmentation of the classifier on CT images. Tab. 1 sums the values of DICE index, precision and recall found for the five patients for LSMB , IBM and LPBM ROIs. The highest indices values were obtained for LSBM and IBM subregions, where the DICE index was ≥ 0.75 and precision was ≥ 0.80 in 4 out of 5 patients. The high values of recall for these two subregions (average values: 0.80 and 0.73 respectively) means that the aBM was correctly recognized by the classifier on CT images. For the LPBM , supoptimal results were achieved (average DICE index was 0.5). The identification of aBM in the LPBM might be influenced by the presence of the coxa.

Physics- Division of Radiation Oncology, Houston, USA ; 9 Copenhagen University, Faculty of Health and Medical Science, Copenhagen, Denmark Purpose or Objective Clinical proton radiotherapy (RT) currently assumes constant spatial relative biological effectiveness (RBE). In vitro data have demonstrated a dependence of RBE on local linear energy transfer (LET), but in vivo data are lacking. In order to model the relationship between dose, LET and RBE in clinical settings, tissue response must be assessed on the same voxel-to-voxel level as the 3D treatment plan information. We investigated a hybrid image registration methodology that utilizes multimodal anatomical imaging to deform functional images (diffusion tensor imaging [DTI] and 18F-FET-PET) between different timepoints, to relate local tissue response to dose and LET on a voxel-to-voxel basis. Material and Methods Multimodal MRI sequences were acquired at baseline and in follow-up as part of a prospective study. Here we present the first results of multimodal hybrid image registration on a single patient with suspicion of progression due to anatomical changes leading to 18F-FET- PET scans on clinical indication in addition to the protocol MRI scans. All anatomical scans at a given timepoint were rigidly co-registered to create a hybrid template. Multimodal deformable image registration was used to relate templates between timepoints. The resulting deformation matrix was used to map functional imaging between baseline and all follow-ups. Functional imaging response was related to dose and LET on a voxel-by-voxel basis by co-registering planning CT to baseline MRI. Dose cubes were corrected for fraction size effects and LET RBE dependence by calculating: EQD(d x ) x = D p · [((α/β) x · RBE max + RBE min 2 · d p ) / ((α/β) x + d x )] (where d x is dose per fraction, d p is the proton dose per fraction and D p is the total proton dose, (α/β) x = 2.1Gy, RBE max = α p /α x and RBE min 2 = β p /β x ). Two models for dose correction were examined, in order to evaluate the impact of LET variation on RBE: (1) Fixed RBE max = RBE min = 1.1; and (2) model-based (McNamara et al 2015) RBE max = 0.99+LET*0.36/(α/β) x , RBE min = 1.10-0.0039*(α/β) x 0.5 *LET. Response was evaluated using change in fractional anisotropy (FA) MRI (loss of signal related to white matter changes) as well as 18F-FET-PET. Results The image registration pathway is illustrated in Figure 1. For the example patient, imaging changes were noted at follow-up nine months after start of treatment (FU2), where multimodal MRI and FET-PET were performed. Average change in FA from baseline (DFA) and 18F-FET- PET signal as function of EQD2 are shown in Figure 2: Fig 2a for constant RBE (1), and Fig 2b for LET-dependent RBE (2). The constant RBE (1) dose shows very limited dose- response relations, while the LET-dependent RBE dose

Conclusion A classifier based on texture analysis features for the automatic segmentation of active bone marrow in pelvic region was created. The obtained results were promising, especially for the lumbosacral and iliac structures. A larger population of patients will be included in future studies, to better test the generalization capability of this approach. Furthermore, to improve the detectability in the LPBM subregion, other classifiers, may be employed and/or combined among them. EP-1919 Voxel-based assessment of proton RBE in paediatric brain cancer radiotherapy from multimodal imaging M. Skaarup 1 , A. Appelt 2 , M. Lundemann 1 , S. Darkner 3 , M. Jørgensen 4,5 , C. Thomsen 6 , I. Law 7 , D. Mirkovic 8 , R. Mohan 8 , D. Grosshans 8 , C. Peeler 8 , I. Vogelius 1,9 1 The Finsen Center - Rigshospitalet, Clinic of Oncology, Copenhagen, Denmark ; 2 University of Leeds and Leeds Cancer Centre- St. James's University Hospital, Institute of Cancer and Pathology, Leeds, United Kingdom ; 3 University of Copenhagen, Department of Computer Science, Copenhagen, Denmark ; 4 Rigshospitalet, Clinic of Nephrology, Copenhagen, Denmark ; 5 Sjællands Universitetshospital, Radiology department, Roskilde, Denmark ; 6 Rigshospitalet, Department of Radiology, Copenhagen, Denmark ; 7 Rigshospitalet, Department of Clinical Phsiology- Nuclear Medicine and PET, Copenhagen, Denmark ; 8 The University of Texas- MD Anderson Cancer Center, Department of Radiation