Head and neck organ-at-risk multi-modal semantic segmentation using 3D-UNet

Head and neck (HaN) cancer is a common type of cancer. Radiotherapy is used to treat this cancer by targeting cancerous cells while avoiding healthy organs. A precise description of targeted areas and adjacent organs at risk (OARs) is required, which is done using Computed Tomography (CT) images. Ho...

ver descrição completa

Detalhes bibliográficos
Autores: Ben Aïcha, Takwa|||0000-0002-3786-3649, Kacem Echi, Afef|||0000-0001-9219-5228, Ben Amor, Aziz, Baccar, Hamdi, Nadhir Najjar, Mohamed
Formato: artículo
Fecha de publicación:2025
País:España
Recursos:Universitat Autònoma de Barcelona
Repositorio:Dipòsit Digital de Documents de la UAB
Idioma:inglés
OAI Identifier:oai:dnet:uabarcelona_::3e0c74880fe3bef37eeba3fec3c9659e
Acesso em linha:https://ddd.uab.cat/record/326497
https://dx.doi.org/urn:doi:10.5565/rev/elcvia.2147
Access Level:acceso abierto
Palavra-chave:Multi-modal semantic segmentation
Deep learning
3d-unet
Computed tomography
Magnetic resonance radiation therapy
Descrição
Resumo:Head and neck (HaN) cancer is a common type of cancer. Radiotherapy is used to treat this cancer by targeting cancerous cells while avoiding healthy organs. A precise description of targeted areas and adjacent organs at risk (OARs) is required, which is done using Computed Tomography (CT) images. However, some OARs in the head and neck region are better observed in magnetic resonance (MR) images. Therefore, we propose a fully automated system for OAR segmentation using CT images and other imaging modalities. More specifically, we want to use the patient's CT and MR images to identify 30 organs that may be at risk. We proposed 3D-UNet, a model for volumetric segmentation that accurately captures spatial relationships. The model has to skip connections for feature propagation, improving segmentation. In addition, it can handle multimodal inputs to integrate complementary imaging information for more precise segmentation. Our proposed model achieved a training accuracy of 94.2% and a test accuracy of 94.1% which competes with the related works.