Contribution à l’amélioration du contourage de l’image médicale en radiothérapie

Abstract

Radiation therapy is one of the main cancer treatments. Contouring of the tumor and the surrounding organs in computed tomography (CT) images constitutes an essential step in the radiotherapy workflow. This operation is generally performed manually by the oncologist. Therefore, manual segmentation takes a long time and suffers from inter and intra-observer variations. Recently, the contribution of artificial intelligence (AI) and particularly Deep Learning (DL) techniques became very important in this field. Successful results have been obtained with automatic segmentation for different imaging modalities (CT, MRI, PET scan). In this thesis, we presented a contribution to the automatic of clinical target volumes (CTVs) and organs at risk (OARs) in CT images. First, different applications of artificial intelligence in radiotherapy were presented. Subsequently, we presented a study of the segmentation of CTV target volume, heart and lungs in CT images taken from a public database. We used three segmentation models based on Deep Learning. Using evaluation metrics, we compared and evaluated the results obtained with the three models. Besides, we considered another public database (abdominal CT images). We performed kidney segmentation using the U-Net model. Evaluation of this segmentation was carried out using several evaluation metrics. Furthermore, we compared our results to those obtained by other authors. Finally, we used a publicly available dataset to segment the liver and the brain using the U-Net model. The segmentation performance was evaluated through geometric metrics. Additionally, three clinicians qualitatively assessed the segmentations for clinicalrelevance, and correlations between geometric metrics and clinical evaluations were analyzed.