ESTRO 2024 - Abstract Book

S3133

Physics - Autosegmentation

ESTRO 2024

2766

Mini-Oral

Automatic segmentation method for OARs and GTVs delineation for nasopharyngeal carcinoma treatments

Mehdi Astaraki 1,2 , Iuliana Toma-Dasu 1,2

1 Stockholm University, Medical Radiation Physics, Stockholm, Sweden. 2 Karolinska Institutet, Oncology-Pathology, Stockholm, Sweden

Purpose/Objective:

The quality of the delineated contours of Gross Target Volumes (GTVs) and Organs-at-Risks (OARs) plays an important role in the treatment of Nasopharyngeal Carcinoma (NPC). However, the delineation of the OARs poses multiple challenges due to the presence of organs with borders difficult to distinguish on morphological images [1]. In addition, the definition of GTV borders which consist of the primary tumor and the affected lymph nodes in the absence of MRI or PET modalities is challenging [2].

We therefore aimed to develop and externally validate a model for the automatic segmentation of tumors (NPC) and healthy organs relevant for the treatment of NPC.

We participated with this work in the Segmentation of OARs and GTVs of NPC for Radiotherapy Planning (MICCAI SegRap 2023) challenge [3]. Our model won the first place for the GTVs segmentation in the test phase and the second place for both the GTVs and OARs tasks in the validation phase.

Material/Methods:

The SegRap dataset consists of co-registered CT and contrast-enhanced CT volumes of 200 NPC patients for segmenting 54 healthy structures to form 45 OARs as well as 2 GTVs including primary tumors and affected lymph nodes. From this set, 120 subjects with corresponding annotations were released as the training subset. The developed models were later submitted for an objective evaluation over 20 subjects in the validation and 60 subjects in the test phase. Our proposed pipeline consists of three steps: (i) enhancing the contrast between the surrounding soft tissues by applying intensity clamping in each of the modalities, (ii) automatically cropping the volumes around the H&N regions, and (iii) 3D U-Net model with convolutional blocks for segmentation. Two models were trained independently for OARs and GTVs tasks. The OARs model was trained with a patch size of 64 ✖ 192 ✖ 160 for 2500 epochs and the GTVs model was trained with a patch size of 80 ✖ 192 ✖ 128 for 700 epochs. A combination of Dice and binary cross entropy was used as the objective function [4].