Adaptive kernelized evidential clustering for automatic 3D tumor segmentation in FDG–PET images

Automatically and reliably delineating tumor contours in noisy and blurring PET images is a challenging work in clinical oncology. In this paper, we introduce a specific unsupervised learning method to this end. More specifically, a robust clustering algorithm with spatial knowledge enhancement is developed in the framework of belief functions, a formal and powerful tool for modeling and reasoning with uncertain and/or imprecise information. Diverse patch-based image features are extracted to comprehensively describe PET image voxels. Then, informative input features are iteratively selected to learn an adaptive kernel-induced metric in an unsupervised way, so as to precisely grouping voxels into different clusters. The effectiveness of the proposed method has been evaluated on FDG–PET images for lung tumor patients.