ESTRO 2020 Abstract book

S1001 ESTRO 2020

5x5x5). The input of the network was a 3D volume of size 192x192x64 replicated 16 times to start with 16 channels. For training and validation the input sub-volume was randomly selected from the entire CT scanning volume. The output of the network was another volume of size 192x192x64x9 (9 channels for the 8 classes and 1 for the background). A weighted dice score was used as loss function (weighted equally across all 9 classes). Training was done on a GPU card (Tesla P100, NVIDIA) using an Adam optimizer using a decaying learning rate until no improvement was seen in weighted dice score in the validation set. Results After 90 epochs of training the network produced delineations with an overall weighted dice score of 0.79. The dice scores for all regions of interest are given in table 1. Errors were primarily visible in the cranio-caudal direction (i.e. errors in the upper and lower boundaries of the regions of interest) especially for the femoral heads and all clinical target volumes. Deviations from the manual delineations in the left-right and anterior-posterior direction were generally small, except for the pre-sacral lymph nodes. This is likely caused by the absence of clear anatomical boundaries. An example of the output of the network is given in figure 1. Conclusion Deep learning based auto-contouring for both the organs at risk and clinical target volumes in rectal cancer patients resulted in reasonable delineations. These delineations might be used as a starting point to speed up clinical workflow. Table 1. Dice scores for the network after training evaluated on the test set (n=13).

body outline using an autosegmentation tool (MIRS, Varian Medical Systems). The labelled mask is warped into the other 19 motion phases by applying the MVFs determined by acMoCo. The 4D mask is used to anatomically constrain a second acMoCo run: Zero motion is enforced on the spine and in voxels surrounding the patient, and a sliding lung motion filter is applied to allow for a tangential discontinuity at the pleural cavity.

The results of the anatomically constrained acMoCo (acacMoCo) algorithm are compared with the unconstrained acMoCo approach and with the advanced McKinnonBates (aMKB) algorithm [MedPhys 45(8):3783ff, 2018] (iTools Reconstructions, Varian Medical Systems). Results The overall visual quality is significantly improved with acacMoCo compared with acMoCo. The movement of the spine is suppressed which results in a highly realistic 4D display. The sharpness of the diaphragm is increased. Compared with aMKB far less artifacts are remaining in acacMoCo and small details such as lung vessels are clearly visible.

Conclusion Anatomical constraints help to improve the image quality of MoCo algorithms, increasing sharpness of the diaphragm and reducing non-physiological artifacts such as bones “breathing” in synchrony with respiration. PO-1718 Deep learning based auto-contouring of planning CT scans for rectal cancer L. Bokhorst 1 , M.H.F. Savenije 1 , M.P.W. Intven 1 , C.A.T. Van den Berg 1 1 UMC Utrecht, Radiotherapy, Utrecht, The Netherlands Purpose or Objective Determine the accuracy of a deep learning based auto- contouring approach for CT based organ at risk and clinical target volume (CTV) delineation in rectal cancer. Material and Methods A total of 116 non-contrast enhanced pelvic planning CT scans were available for this study, all with clinically used delineations of the body contour, bladder, femoral heads (left and right), and CTVs (mesorectal, internal iliac lymph nodes (left and right) and pre-sacral lymph nodes). Data was split in a training (n=90), validation (n=13), and test (n=13) set. A 3D encoder-decoder style network was used consisting of 4 down/up convolution levels (kernel size = 2x2x2, stride = 2) with each level having 1,2,3, and 3 regular convolution layers (kernel size = 5x5x5) respectively. Additionally, at the bottom of the network there are 3 regular convolution layers (kernel size =

Figure 1. Example of the results of the network (Top rows show the prediction of the network. Bottom rows shows the difference with the manual segmentation. Errors were primarily visible in the cranio-caudal direction and with the pre-sacral lymph nodes.)