ESTRO 38 Abstract book

S371 ESTRO 38

prior to treatment planning (Auto+Corr). Two weeks later the same OARs were manually delineated by the same two radiation oncologists (Manual). To quantify efficiency of automated delineation, the time taken for correction of the automated delineation was compared to manual delineation time. Dice similarity coefficients (DSCs) were calculated to quantify overlap between different contours (Auto, Auto+Corr, Manual) where a value of ‘’1’’ represents perfect and ‘’0’’ no overlap (Figure 1).

16 fractions over four weeks. Of 146 patients, 12 patients were irradiated with 65.6 Gy (RBE) and 134 patients were irradiated with 68.8 Gy (RBE). Each dose level is the dose equivalent to 57.6 and 60.8 Gy (RBE) of NIRS. Results The patients consisted of 68 males and 78 females aged from 19 to 86 years with an average age of 54 years. The most frequent primary site was the nasal and paranasal sinus. Median follow-up time was 24.4 months (range, 4.2- 60.9 months). Fourteen patients died because of local recurrence and/or distant metastasis, and 97 patients were alive without local recurrence at the time of this analysis. Median local control time was 21.4 months (range, 4.2-60.7 months). The 2-year local control and overall survival rates were 82% and 88%, respectively. The most frequent acute toxicity of 146 patients, grade 2 (G2) and grade 3 (G3) mucosal reactions were observed in 53 (36%) patients and 34 (23%) patients, respectively. No G3 mucosal reaction was observed in the 65.6 Gy (RBE) group. G2 acute skin reaction of all patients was observed in 30 (21%) patents. In the analysis of late toxicities, 8 (5%) patients showed G3 osteonecrosis and 5 (3%) patient showed G2 brain reaction. Only one patient needed the surgical intervention due to G3 encephalitis. All of these late toxicities were observed in the 68.8 Gy (RBE) group. Conclusion The review article by Mizoe et al.* reported the 2-year local control rate as about 95% and G2 and G3 mucosal reactions as 36% and 6%. Although our results were slightly inferior to this article, they showed acceptable toxicities and therapeutic effectiveness for adenoid cystic carcinomas. It is necessary to evaluate a large number of patients and a long follow up period. *Mizoe J et al. Radiother Oncol 2012; 103(1): 32–7. **Fossati P et al. Phys Med Biol 2012; 57: 7543–7554. Acknowledgment: This study was partially supported by Associazione Italiana per la Ricerca sul Cancro (AIRC, project IG-14300). PO-0723 Benefits of deep learning for delineation of organs at risk in head and neck cancer J. Van der Veen 1,2 , S. Willems 3 , D. Robben 3 , W. Crijns 2 , F. Maes 3 , S. Nuyts 1,2 1 KU Leuven – University of Leuven, Department of Oncology, Leuven, Belgium ; 2 UZ Leuven – University Hospitals Leuven, Department of Radiation Oncology, Leuven, Belgium ; 3 KU Leuven – University of Leuven, Medical Image Computing ESAT/PSI, Leuven, Belgium Purpose or Objective Delineation of organs at risk (OARs) is necessary for correct treatment planning and for comparison of dose- volume histograms between doctors, centres and studies. Although guidelines exist, manual delineation shows significant inter-observer variability. The aim was to develop a 3D convolutional neural network (CNN) that allows automated delineation of OARs resulting in more efficient and consistent contouring of head and neck OARs from 70 HNC patients were manually delineated using the international consensus guidelines by Brouwer et al. (2015). The planning CT scans together with the delineations were used to train a single 3D CNN to delineate 16 different OARs. The 3D CNN is based on the work of Kamnitsas et al. (2017) and has four pathways. Each pathway operates on a different resolution and region of interest (ROI), allowing both a broader ROI and fine details to be processed simultaneously. These pathways are concatenated, followed by extra convolutional layers to predict the final delineation. Next, the automated delineation (Auto) of OARs was performed on planning CT images of 15 new patients. Then, the automated delineations were subsequently corrected by a senior resident and supervised by the treating physician cancer (HNC) patients. Material and Methods

Results Average correction time was significantly shorter than the average time needed for complete manual delineation (21 versus 34 minutes respectively, p<6.5E-4). The CNN performed best for spinal cord, oral cavity, mandible, brainstem and parotid glands with median DSCAuto vs.Manual of 0.92, 0.91, 0.91, 0.87 and 0.86 respectively. Median DSC Auto vs.Auto+Corr for these same OARs was 0.99, 0.96, 0.99, 0.94 and 0.97 respectively and this was in general significantly higher compared to DSC Auto vs.Manual (p<1.1E-6). For all OARs except supraglottic larynx, median DSC Auto vs.Auto+Corr was higher than DSC Auto+Corr vs.Manual , indicating that the corrections applied were smaller than the intra-observer variability (figure 2).

Conclusion We developed a CNN that allows more time efficient delineation of OARs in HNC patients. The CNN is able to produce automated delineations that only need small alterations in most OARs and it has therefore been implemented in clinical practice in our hospital. PO-0724 Texture analysis for predicting laryngeal preservation in advanced laryngo-pharyngeal cancers