Automated thresholded region classification using a robust feature selection method for PET-CT

Abstract

Fluorodeoxyglucose Positron Emission Tomography - Computed Tomography (FDG PET-CT) is the preferred imaging modality for staging the lymphomas. Sites of disease usually appear as foci of increased FDG uptake. Thresholding is the most common method used to identify these regions. The thresholding method, however, is not able to separate sites of FDG excretion and physiological FDG uptake (sFEPU) from sites of disease. sFEPU can make image interpretation problematic and so the ability to identify / label sFEPU will improve image interpretation and the assessment of the total disease burden and will be beneficial for any computer aided diagnosis software. Existing classification methods, however, are sub-optimal as there is a tendency for over-fitting and increased computational burden because they are unable to identify optimal features that can be used for classification. In this study, we propose a new method to delineate sFEPU from thresholded PET images. We propose a feature selection method, which differs from existing approaches, in that it focuses on selecting optimal features from individual structures, rather than from the entire image. Our classification results on 9222 coronal slices derived from 40 clinical lymphoma patient studies produced higher classification accuracy when compared to existing feature selection based methods.