Joint variational segmentation of CT/PET data using non-local active contours and belief functions
Foued Derraz, Antonio Pinti, Laurent Peyrodie, Miloud Bousahla, Hechmi Toumi · Pattern Recognition and Image Analysis · 2015
In this paper, we have proposed a new framework to use both PET and CT images simultaneously for tumor segmentation. Our method combines the strength of each imaging modality: the superior contrast of PET and the superior spatial resolution of CT. We formulate this problem as a Non-Local Active Contours (NL-AC) based-variational segmentation framework incorporating Belief Functions (BFs). The proposed method used all features issued from both modalities (CT and PET) as a descriptor to drive the NL-AC curve evolution. The new segmentation framework allows us to incorporate in the same framework heterogeneous knowledge in order to reduce the imprecision due to noise poor contrast, weak or missing boundaries of objects, inhomogeneities, etc. The proposed method was evaluated on relevant tumor segmentation problems. The results showed that our method can effectively make use of both PET and CT image information, yielding segmentation accuracy of 81.52% in Dice Similarity Coefficient (DSC) and the Average Symmetric Surface Distance (ASSD) of 1.2 ± 0.8 mm, which is 10% (resp., 16%) improvement compared to two state of art segmentation methods using the PET (resp., CT) images.