ESTRO 2020 Abstract book

S240 ESTRO 2020

Conclusion The deep-learning cGAN model is a feasible and generalisable method for anal and rectal sCT generation using T2-SPACE sequences. SCTs with clinically acceptable dosimetric accuracy were produced from data acquired from multiple centres, for male and female anatomy and for anal and rectal cancer sites. The AutoConfidence AI approach enabled individual patient sCT robustness, through highlighting sCT slices and global sCT datasets with lower HU accuracy. PH-0411 Neural network based MR-only treatment planning for cervix cancer using low field MR images M. Buschmann 1 , L. Fetty 1 , G. Heilemann 1 , M. Heilmann 1 , P. Kuess 1 , D. Georg 1 , N. Nesvacil 1 1 Medical University of Vienna, Department of Radiation Oncology, Vienna, Austria Purpose or Objective With the introduction of MR-only workflows, the generation of synthetic CT (sCT) images from MR scans has gained interest. Neural networks are increasingly applied for the task of MR-to-sCT-conversion. However, most sCT studies were based on high-field strength MR scanners. Low field MR scanners were so far not investigated in detail. Images from these scanners usually suffer from poorer image contrast and lower signal to noise ratio. The current study investigates the use of low field MR images and a neural network for MR-only external beam treatment planning of cervix cancer patients. Material and Methods A pretrained generative adversarial network was utilized for sCT generation. It was trained and tested on an independent patient dataset consisting of 41 patients with pelvic malignancies (prostate and cervix; 25 for training + 16 for evaluation) for MR to sCT conversion. Five patients with locally advanced cervical cancer were retrospectively selected. All patients received pre-treatment MR scans in a 0.35T MAGNETOM C! open-MR scanner (Siemens, Erlangen) in treatment position. Target volumes and organs at risk (bladder, rectum, bowel, femoral heads) were contoured on the MR images. For each patient a VMAT plan with two arcs was generated on the sCT image with the aim of delivering 45 Gy in 25 fractions to the PTV. In one patient two simultaneous integrated boosts with 55 Gy were planned for pathological lymph nodes. All treatments were planned for a 10 MV photon beam provided by an Elekta Synergy LINAC. sCT images were rigidly registered (without rotation) to the planning CT (pCT) scan and the VMAT plan was recalculated on the pCT images. The two dose distributions were compared using a gamma analysis (3 mm/3%) and dose volume metrics of target and organ structures (D mean , D 2% , D 98% ; for bowel: D 2cc , V 40Gy , V 30Gy ). Results The gamma agreement index between pCT and sCT images was 98%±1%, when a threshold of 10 Gy was used. Deviations of PTV dose parameters were below 2% for all patients. For bladder and rectum the agreement was in the range of 3%, whereas the femoral heads showed higher deviations up to 4.5% with no systematic difference. The volumes V 40Gy and V 30Gy of the bowel were slightly elevated on sCT images with deviations up to 10% in one patient compared to pCT dose distributions.

Conclusion This study demonstrated the feasibility of generating sCTs from low field MR images for MR-only VMAT planning of cervix cancer. The deviations in dose distributions from the ground truth pCT images were considered small and are rather caused by anatomical variations between MR and pCT scans than from inaccuracies in sCT grey values. The performance of the network was good although having been trained on a high proportion of prostate scans. Future studies will examine larger patient numbers and proton dose distributions. A limitation of this approach is the limited field of view in longitudinal direction of the low field MRs which may be a problem for patients with paraaortic irradiation.