Semiautomated thoracic and abdominal computed tomography segmentation using the belief functions theory: application to 3D internal dosimetry.

Abstract : AIM: Segmentation of computed tomography (CT) images is an important step in three-dimensional (3D) internal dosimetry. To this end, a semiautomated method was developed to delineate organs, using the belief functions theory. MATERIALS AND METHODS: The membership degree of each voxel to each volume of interest is estimated by computing a basic belief assignment (bba). For each voxel V, bbas corresponding to each neighbor are aggregated to obtain a unique bba, using a merging procedure. Before aggregating information, a 3D filter is applied, in order to take into account the fact that the more the voxel V(i) is close to V, the more the information coming from V(i) is reliable. The aim is to weaken the contribution of voxels according to their distance with respect to the voxel to be classified. The algorithm was applied on 10 CT scans (pixel size, 0.98 x 0.98 mm2, slice thickness 3 mm, 120 kV). For each organ (i.e., the lung, liver, kidney, and spleen), the algorithm was applied on a part of the CT volume. First, the lung was segmented using two classes with characteristic values, C, defined by the K-means clustering algorithm. Second, the liver and kidneys were segmented using three classes with C-values defined by local mean Hounsfield Unit (HU) measurements, corresponding to fat, liver, and kidney. Third, the spleen was segmented using three classes corresponding to fat, kidney, and spleen, using local mean HU measurements. The semiautomated segmentation was compared with manual segmentation using the volume difference and the agreement (overlap) index. RESULTS: For organ segmentation, the computation duration was between 5 and 20 minutes (2.5 GHz, RAM of 1 GByte), depending on the number of classes and the volume size to classify. On the 10 patients, manual correction was needed for none on the lung, 1 on the spleen, 2 on the kidneys, and 7 on the liver, mostly owing to intercostal structures. The mean relative volume difference (+/-1 standard deviation [SD]) between manual and automated segmentation was 5.0% +/- 3.7%, 5.7% +/- 3.6%, 6.2% +/- 3.3%, and 7.5% +/- 6.5% for the lung, liver, kidney, and spleen, respectively. The corresponding mean agreement index (+/-1 SD) was 0.94 +/- 0.03, 0.90 +/- 0.01, 0.88 +/- 0.03, and 0.84 +/- 0.04. CONCLUSIONS: The algorithm allows for the delineating of organs on thoracic and abdominal CT images, and will be integrated in a 3D internal dosimetry dedicated software.
Type de document :
Article dans une revue
Cancer Biotherapy & Radiopharmaceuticals, 2007, 22 (2), pp.275-80. 〈10.1089/cbr.2006.318〉
Liste complète des métadonnées

http://www.hal.inserm.fr/inserm-00467287
Contributeur : Céline Breton <>
Soumis le : vendredi 26 mars 2010 - 13:03:11
Dernière modification le : mardi 5 juin 2018 - 10:14:41

Identifiants

Citation

Arnaud Dieudonné, Peng Zhang, Patrick Vannoorenberghe, Isabelle Gardin. Semiautomated thoracic and abdominal computed tomography segmentation using the belief functions theory: application to 3D internal dosimetry.. Cancer Biotherapy & Radiopharmaceuticals, 2007, 22 (2), pp.275-80. 〈10.1089/cbr.2006.318〉. 〈inserm-00467287〉

Partager

Métriques

Consultations de la notice

94