ESTRO 2020 Abstract book

S1021 ESTRO 2020

implementing CT-based attenuation maps into the PET reconstruction path. References: [1] Carney et al. DOI: 10.1118/1.2175132 [2] Paulus et al. DOI: 10.1088/0031-9155/58/22/8021 Table 1: Comparison of the average percentage difference of PET signal intensities without device attenuation correction (uncorrected scan) and with device attenuation correction (corrected scan) and the reference scan, given as absolute values.

this methodology preserves data privacy by training models without the data ever leaving the firewall of its original healthcare provider. References [1] P. Bilic et al. , “The Liver Tumor Segmentation Benchmark (LiTS),” arXiv:1901.04056 [cs] , Jan. 2019. [2] M. Abadi et al. , “Deep Learning with Differential Privacy,” Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security - CCS’16 , pp. 308–318, 2016. PO-1745 PET/MRI for RT planning: CT-based attenuation correction of radiation treatment positioning devices L. Taeubert 1 , Y. Berker 2 , B. Beuthien-Baumann 3 , A.L. Hoffmann 4 , E.G.C. Troost 4 , A. Pfaffenberger 1 , M. Kachelrieß 2 , C. Gillmann 1 1 German Cancer Research Center, Medical Physics in Radiation Oncology, Heidelberg, Germany ; 2 German Cancer Research Center, X-ray imaging and Computed Tomography, Heidelberg, Germany ; 3 German Cancer Research Center, Radiology, Heidelberg, Germany ; 4 Helmholtz-Zentrum Dresden-Rossendorf, Radiooncology, Dresden, Germany Purpose or Objective To correct for the PET signal attenuation of RT patient positioning hardware and MRI coils by implementing user- derived CT-based attenuation maps into the PET The RT positioning hardware consisted of a flat RT table overlay, coil holders for abdominal scans, coil holders for head and neck scans (all Medibord) and an MRI compatible ProStep system (Innovative Technologie Völp) for hip and leg immobilization. Each hardware element was scanned at a CT (Somatom Confidence, Siemens Healthineers) using 120kV/255mA for low attenuating hardware (RT table overlay, coil holders, ProStep) and 140kV/467mA for high attenuating hardware (coil holders with coils). From the CT images, attenuation maps were derived using the bilinear (low attenuating hardware) [1] and the adapted bilinear (high attenuating hardware) [2] approach. Attenuation maps were implemented into the PET reconstruction path. For experimental validation of attenuation correction accuracy, PET images with each positioning device mounted independently (RT scans) and one reference scan (without any RT positioning hardware present) were acquired at a PET/MRI scanner (Biograph mMR, Siemens Healthineers) using an active 68Ge phantom (32 MBq, 10 min scan time). For each positioning device, PET reconstructions of the RT scans were performed twice: once without hardware attenuation correction (uncorrected RT scan) and once with hardware attenuation correction implemented (corrected RT scan). The average percentage difference in PET signal intensities of the uncorrected and corrected reconstructions to the reference reconstruction was calculated. Results Table 1 compares the average percentage difference of PET signal intensities for the corrected and uncorrected RT scans with the reference scan for each positioning device. The PET signal attenuation ranges from 7.9±0.8% (head and neck coil holder with coils) to 1.6±0.4% (abdominal coil holder). With attenuation correction, these values can be reduced to 2.5±2.0% and 0.3±0.7%, respectively. Conclusion Integrating hybrid PET/MRI into radiation treatment planning has great potential to improve tumor delineation and dose prescription. Since these scans must be acquired under treatment conditions, additional RT positioning hardware is necessary. The additional patient positioning hardware necessary causes a PET signal attenuation in the order of up to 8% which can be corrected to 2.5% or less by reconstruction path. Material and Methods

PO-1746 Deep learning artificial intelligence for auto- segmentation of the bowel bag. J. Miskell 1 , C. Thomas 1 , M. Pearson 1 1 Guy's and St. Thomas' NHS Foundation Trust, Radiotherapy, London, United Kingdom Purpose or Objective During the radiotherapy (RT) treatment planning process, it is essential that organs at risk (OAR) are delineated accurately and consistently to ensure that healthy tissue is spared. Manual OAR delineation is time-consuming and prone to large inter and intra-operator variability. Methods for automatic segmentation have the potential to introduce significant time savings and reduce variability. Neural network auto-segmentation models have thus far concentrated on segmentation of individual OARs. This work formulates a network for auto-segmentation of the bowel bag (BB) – a structure encompassing all bowel loops – which is manually delineated at our centre for all patients receiving RT to pelvic nodes. Material and Methods A 2D CNN, based on the U-net architecture, was developed and trained using clinically approved BB delineations from 20 pelvic node patients. Input data consisted of 2D CT slices labelled with binary masks of the BB. The network was trained using a binary cross entropy loss function for 500 epochs. The model was then tested on a further 5 independent patient CT datasets to auto-segment BB. Similarity metrics were used to quantitatively assess the performance of the model. Results Despite low patient numbers, initial performance of the trained model is promising. The segmented BB structures were directly compared to gold-standard segmentations, resulting in dice similarity coefficients for each of the 5 auto-segmented BB structures of 0.87, 0.89, 0.90, 0.90 and 0.88, and the corresponding Jaccard indexes of 0.78, 0.80, 0.83, 0.82 and 0.79. Dice coefficients were measured for each slice and ranged between 0.39 and 0.96, with the network performing better on the superior slices of the BB and worse on inferior slices. The inferior slices across all patients show consistently poorer performance in areas where the bowel loops descend around the bladder.