ESTRO 2020 Abstract book

S1019 ESTRO 2020

The spatial analysis provided information about the location where the Atlas and DLC based delineations performed differently, and which areas need further editing. Furthermore, differences in contouring style of the two contouring algorithms became obvious. This enables a more structured way of improving delineation [1] Lustberg T, et al; Clinical evaluation of atlas and deep learning based automatic contouring for lung cancer; Radiother Oncol. 2018 Feb; 126(2):312-317 PO-1742 Post-processing 4DCT to improve delineation of heart substructures M. Van Herk 1 , A. McWilliam 1 , K. Banfill 1 , C. Faivre-Finn 1 1 University of Manchester, Radiotherapy Related Research, Manchester, United Kingdom Purpose or Objective Dose constraints for cardiac substructures such as left and right coronary arteries (LCA, RCA) will likely be implemented clinically in the near future. However, many centers delineate organs at risk on the average of a 4DCT. Because the heart moves with respiration as well as heartbeat, cardiac sub-structures are poorly visible. It is our aim to compare substructure definition between the average scan, a scan processed with respiratory motion compensation (mid-position scan), and with an experimental post-processing method that sorts for cardiac motion. Material and Methods 4DCTs were collected of 4 patients in an ethics-approved prospective study on cardiac damage after radiotherapy. First, the average (AVG) was calculated. Second, the 10 frames were deformable registered, deformed to mid- position and averaged (MIDP). After motion compensation for respiration, the dominant motion that remains is cardiac, causing motion blurring that differs from frame to frame and slice by slice (the heart beats once every few CT rotations). The least motion blurring occurs at end diastole with the least heart motion. Cardiac self-sorting (CSS) is a method that aims to select the slice closest to end-diastole from the 10 respiratory corrected frames. Selected slices are stacked to create a new 3D volume. To evaluate image quality, 13 observers were asked for qualitative feedback and then to circle the origin of LCA and RCA. Scans were analysed in random order and observers were blind to the applied image processing. We report qualitative results as well as observer variation. Cardiac CT taken on a different day was available as ground truth (GT) but not provided to the observers. Results The quality of CSS was judged from sagittal reconstructions through the heart. Self- sorting failed on average once every 10 slices and this was noted by some observers. They reported AVG to be blurry, MIDP less so and CSS as sharpest (Fig. 1). In 3 out of 4 cases the observers commented that the quality of the CSS scan was similar to the scan quality of a cardiac CT scan. In the fourth case RCA remained hard to visualize, even on the gated cardiac CT GT. Outliers over 1 cm from median occurred for 20% of arteries for AVG, 18% for MIDP and 13% for CSS. Observer variation vector length was 5.4 mm, 4.8 mm and 4.7 mm, respectively. The greatest improvements occurred for the RCA, generally the LCA is easier to visualize. algorithms. References

Conclusion It is possible to reduce motion blurring due to respiration and heartbeat in 4DCT scans. Both respiratory mid-position as well as slice by slice self-sorting techniques are useful. It is likely that results will improve with more advanced sorting methods, or by recording the heartbeat during scanning. This method opens up the way for better definition of heart substructures during treatment planning. We acknowledge the help of all the observers. PO-1743 automatic segmentation of nasopharyngeal carcinoma: a solution for single institution Y. Guo 1 , Y. Qing 1 1 Fudan University Shanghai Cancer Center, Department of Radiation Oncology, Shanghai, China Purpose or Objective To build a solution for nasopharyngeal carcinoma automatic segmentation on single institution with MR images. Material and Methods This study employed a Unet with VGG16 backbone to build end-to-end segmentation models. 371 nasopharyngeal carcinoma patients’ GTV-nx and GTV-nd were delineated by a junior oncologist on MR images. (GTV-nx was defined as the primary nasopharyngeal tumor mass and GTV-nd was defined as the metastatic lymph node, GTV-nd-L and GTV- nd-R respectively represented the left and right cervical lymph nodes). These contours were used to train the initial model (model ini ). We predicted contours on training dataset by this model. Then auto-segmented contours of these patients were generated and fine-tuned by the junior oncologist. The fine-tuned contours were used to train two auto-segmentation models (model co and model re ). model co inherited weights in model ini , while model re start with random weights. All of them had been trained for 200 epochs and evaluated by an independent validation dataset. The validation dataset was delineated by a senior oncologist, including 38 patients. Results For GTV-nx, the average values of DSCs of three models were 0.83±0.06, 0.83±0.06 and 0.83±0.07 for model ini , model re and model co respectively. And no statistical difference was observed among them. For GTV-nd-R, the average values of DSCs of three models were 0.72±0.15, 0.74±0.16 and 0.73±0.16 for model ini , model re and model co , respectively. The DSC of model re was statistically significantly higher than model ini (P=0.019<0.05). For GTV- nd-L, the average values of DSCs of three models were 0.74±0.14(model ini ), 0.76±0.13 (model re ) 0.76±0.14 (model co ). Statistically significant difference was observed for the DSC between model ini and model re (P=0.004<0.05). Some examples of GTV-nx and GTV-nd were respectively shown in Figure 1 and Figure 2