ESTRO 2022 - Abstract Book

S474

Abstract book

ESTRO 2022

Conclusion The choice of dose comparison method has the largest impact on error identification for pre-treatment QA using DL, compared to image preprocessing. Model performance can improve by applying mean/stdev normalization and high image resolution, but the latter needs more computational resources and longer training times. While this is not a major issue for 2D images, it may be for 2D images per treatment segment or for 3D reconstructed dose volumes.

1. Nyflot et al. 2019 Med Phys 46: 456-464 2. Potter et al. 2020 Med Phys 47: 4711-4720 3. Kimura et al. 2021 Med Phys 48: 4769-4783

MO-0549 No need for manual adjustments of deep learning segmentation in oropharyngeal cancer?

H. van de Glind 1 , I.G. van Bruggen 2 , J.A. Langendijk 2 , S. Both 2 , C.L. Brouwer 2

1 Universitair Medisch Centrum Groningen , Department of Radiation Oncology, Groningen, The Netherlands; 2 Universitair Medisch Centrum Groningen, Department of Radiation Oncology, Groningen, The Netherlands Purpose or Objective Delineation of organs at risk (OAR’s) plays a critical role in radiotherapy treatment planning. However, the segmentation of OAR’s can be very time-consuming, especially in the head and neck region. The accuracy of our deep learning based automated segmentation is currently within interobserver variability, however the influence of its use in treatment planning – without performing manual adjustments – is still unclear. We compared dose and normal tissue complication probability (NTCP) of fully automated vs. adjusted deep learning contouring (DLC) for volumetric modulated arc therapy (VMAT) and intensity-modulated proton therapy (IMPT). Materials and Methods A test set of 10 patients, who were treated for oropharyngeal cancer between February 2021 and July 2021 in the UMCG, was selected. The patients were treated with a prescribed dose of 70 Gy in 35 fractions to the primary tumour and 54.25 Gy to the elective lymph node areas. DLC (Mirada Medical, Oxford, United Kingdom) was used for automated segmentation of OAR’s. Treatment plans were created in RayStation Development 10B using internally validated machine learning based automated planning models (RaySearch Laboratories, AB, Stockholm, Sweden). For every patient four plans were created: DLC VMAT, DLC adjusted VMAT, DLC IMPT and DLC adjusted IMPT. The DLC plan is optimized using automatically segmented DLC. The DLC adjusted plan is optimized using manual adjusted DLC by a radiation therapy technician and approved by a radiation oncologist. NTCP values for the development of late xerostomia and dysphagia were calculated and compared between the DLC and DLC adjusted plans. The input variables to the NTCP model for xerostomia were the mean planned dose to the parotid and submandibular glands, and for dysphagia the mean dose to the oral cavity and pharyngeal constrictor muscles based on the DLC and DLC adjusted plans. All dose and NTCP variables were derived using the manually adjusted contours. Results The mean dose differences between the DLC and DLC adjusted plans for the OAR’s relevant for the NTCP’s were within 0.07 Gy (Table 1a). The mean difference in NTCP (NTCP DLC – NTCP DLC adjusted) was small: -0.08 percentage point (pp) for VMAT and 0.11 pp for IMPT for xerostomia grade II, and 0.16 pp for both VMAT and IMPT for dysphagia grade II (Table 1b). Statistically significant difference was found between the NTCP of the DLC and DLC adjusted IMPT plans for dysphagia (p = 0.01), as shown in Figure 1.