ICHNO-ECHNO 2022 - Abstract Book

S33

ICHNO-ECHNO 2022

Conclusion Sarcopenia can be easily assessed on pre-treatment CT scans and is a promising adverse prognostic factor for OS an DDFS in HNSCC patients

PO-0065 Automatic segmentation of organs at risk (OARs) in the head and neck (H&N) region using 3D Unets

A. Traverso 1 , S. Datta 1 , S. Pai 1 , I. Hadzic 1 , D. Bontempi 1 , S. Mehrkanoon 2 , A. Dekker 3 , A. Briassouli 2

1 Maastro Clinic, Radiotherapy, Maastricht, The Netherlands; 2 Maastricht University, Department of Knowledge Engineering , Maastricht, The Netherlands; 3 Maastro Clinic, Department of Radiotherapy, Maastricht, The Netherlands Purpose or Objective Traditional deep learning (DL) methods struggle to segment organs with a poor surrounding contrast when using the standard loss function (Sorensen-Dice coefficient) for “training”. This is a relevant issue in the H&N region, where tiny OAR structures are present. We explored several combinations of current state of the art approaches to improve OAR auto contouring for H&N. Materials and Methods A total of 50 CT scans from nasopharynx patients with 22 annotated OARs from the MICCAI 2019 challenge were used for training; while 30 CT scans from a custom curated TCGA-HNSC dataset were used for validation. A histogram of all the OARs was computed and the region where most of the information was concentrated was cropped and scaled to [0,1] which is referred to as Naive Linear Function (NLF). The segmental linear function (SLF) was used to preprocess the images. The SLF mimic the fact that radiologists would use different window widths/levels to better differentiate OARS with diverse HU values. As loss functions, both the traditional Sorensen-Dice coefficient and the generalized surface Dice loss, based on the distance map, were investigated. The distance map is produced by calculating the volumetric distance from each closed surface for the labelled organs. Two pre-processing techniques were investigated: