A 3D difference-of-Gaussian-based lesion detector for brain PET

Positron emission tomography (PET) plays an important role in neurodegenerative disorder diagnosis and neurooncology applications, especially detecting the early metabolism anomalies in human brains. Current lesion detection algorithms can be roughly classified into voxel-based, region of interest (ROI)-based, and global algorithms. These methods may capture the scale and/or location of the lesions in brain, but other important properties, such as lesion metabolism rate and contrast to non-lesion parts are often ignored. To capture these important features, we propose a novel lesion detector with three lesion-centric feature descriptors for brain PET. We analyze the lesion patterns of 331 PET datasets from the ADNI baseline cohort and further perform t-test between different disorder groups to validate the new lesion-centric features. The preliminary results show that the proposed lesion detector is robust in capturing the brain lesions and has a great potential to be a predictive biomarker for neurological disorders.