ESTRO 37 Abstract book

S1188

ESTRO 37

Conclusion Although there are spatial differences, total distortion is comparable between the MR-Sim and MR-Linac. Total distortion was found to be <1.3mm for the Philips Ingenia and <1mm for the Elekta Unity MR-Linac in the two anatomies investigated. Future work will implement this methodology into the online MR-Linac workflow to directly assess geometric accuracy. In a second stage, this method can be used to automatically correct for the combined effect of B0 and gradient induced distortions. EP-2148 Approaching Intra-Physician Contouring Variability: Head and Neck Auto-Contouring with Deep Learning J. Chan 1 , V. Kearney 1 , S. Wu 1 , M. Bogdanov 1 , S. Haaf 2 , N. Dixit 1 , A. Sudhyadhom 1 , S. Yom 1 , T. Solberg 1 1 University of California UCSF, Radiation Oncology, San Francisco CA, USA 2 Nimble Therapy- LLC, Department of Artificial Intelligence, San Francisco CA, USA Purpose or Objective To investigate the feasibility of approaching intra- physician contouring accuracy using a deep learning- based automatic contouring system. Material and Methods The accuracy of 4 different contouring approaches were investigated. A ground truth was generated by manually contouring the spinal cord, brainstem, left parotid, right parotid, and mandible for 5 head and neck cancer patients 10 times. The intra-physician variability was calculated by taking the average of all dice coefficients from 10 combinations of each patient and structure separately. Two atlas-based systems, MIM Maestro (MIM), and RayStation (RS), and one deep learning-based system, Nimble ContourTM (NC) (Nimble Therapy, San Francisco, CA), were evaluated. In total 102 patients were considered in this study. 82 patients were used to train the MIM and NC algorithms, while due to limitations of the system, 10 patients were used to train RS. In the latter, 5 patients used as the ground truths were selected from the remaining 20 validation patients. The accuracy of each algorithm was evaluated by calculating the average dice coefficients from all 10 contour variations relative to the ground truth (Figure 1). Results For all contours, the average dice coefficients were the highest for the intra-physician variability, ranging from 0.848 for right parotid to 0.956 for mandible (Table 1). However, NC was able to achieve very similar results and outperformed both of the other algorithms. The average difference between the intra-physician variability and NC for the spinal cord, brainstem, left parotid, right parotid, and mandible was 0.002, 0.009, 0.012, 0.006, and 0.027 respectively, while the average difference between NC and the next best algorithm was 0.032, 0.032, 0.074, 0.093, and 0.034 respectively.

Conclusion By moving towards deep learning-based algorithms, auto- contouring accuracy can approach intra-physician variability. EP-2149 A priori scatter correction of clinical cone- beam CTs to enable on-line proton dose calculations A.G. Andersen 1 , Y. Park 2 , B. Winey 3 , G. Sharp 3 , U. Elstrøm 1 , J. Petersen 1 , L. Bentzen 4 , L. Muren 1 1 Aarhus University Hospital, Department of Medical Physics, Aarhus V, Denmark 2 UT Southwestern, Department of Radiation Oncology, Dallas- Texas, USA 3 Massachusetts General Hospital, Department of Medical Physics, Boston- Massachusetts, USA 4 Aarhus University Hospital, Department of Radiation Oncology, Aarhus V, Denmark Purpose or Objective The favorable dose characteristics of proton therapy can best be exploited if dose degradations caused by density variations and anatomy changes in the patient are accounted for and minimised. Imaging devices such as on- board cone-beam CT (CBCT) open possibilities for strategies such as online monitoring, re-calculation, adaptation and even re-optimisation of dose distributions. However, CBCT reconstructions are influenced by scatter and other artifacts that distort the density values away from their real values. A scatter correction method based on the a priori information of the planning CT has previously shown potential to improve dose calculation accuracy in CBCTs of phantoms. In this first clinical study we apply the a priori scatter correction algorithm on pelvic CBCTs and evaluate the basis for all on-line applications, the correctness of the Hounsfield Unit (HU) values. Material and Methods Image data sets of four prostate cancer patients were used, including planning CTs (pCTs) with organ delineations and CBCT projections from a Varian on-board imager (OBI) along with the clinical reconstruction (VarianCBCT). The CBCT projections were first reconstructed using a regular Feldkamp-Davis-Kress (FDK) algorithm (rawCBCT). We then registered the pCT to the rawCBCT rigidly (rigidCT) and deformably (deformCT). The deformCT was forward projected to the same angles as the CBCT projections followed by a linear intensity correction. The difference between the two projection sets were smoothed to create a scatter map, which was then subtracted from the CBCT projections before a final FDK (corrCBCT). To evaluate the correctness of the HU values in the CBCT reconstructions we propagated all organ delineations from the pCT using rigid registration. The values inside the structures were then plotted against those at the same geometrical position in the deformCT and histograms were created of all values in the CBCTs and CTs. The correctness was scored as the similarity to the histogram of the deformCT using the Euclidean Earth Movers Distance (EMD). As we expected the HU distribution of the bladder and the femur heads to be the least influenced by the physiological changes between CT and CBCT, these structures were used for comparison. Results The EMD of the bladder were similar for the corrCBCT and the VarianCBCT (<5% difference) while the rawCBCT had an EMD between 18 and 62 percent points higher than corrCBCT and VarianCBCT. For the head of the