Hierarchical modelling for unsupervised tumour segmentation in PET

This paper presents a fully automated and unsupervised method for the segmentation of tumours in PET images. The segmentation technique incorporates a pre-processing stage and a hierarchical approach based on an improved region-scalable energy fitting model. The advantages of the approach lie in its multi-level processing. It first considers the whole range of grey levels in the image volume, which is able to avoid local maxima. Subsequently, the local grey levels range is utilized to refine the segmentation which effectively avoids false negative segmentations. We validate our method using real PET images of head-and-neck cancer patients as well as custom-designed phantom PET images. Compared with other popular approaches, the experimental results on both data sets show that our method can accurately segment tumours in PET images.