A textural feature based tumor therapy response prediction model for longitudinal evaluation with PET imaging

Early therapy response prediction, employing biomarkers such as ¹⁸F-fluorodeoxyglucose (FDG) followed with positron emission tomography (PET), is an actively researched topic. Traditionally, only the first order intensity based feature estimates are used for the response evaluations. In this work, we focus on the predictive power of lesion texture along with traditional features in follow up studies. Both standard and textural features are extracted after delineating the lesions with state-of-the-art methods. We propose subspace learning to reduce the influence of delineation parameters and to represent each patient as a Grassmann manifold spanned by the extracted feature subspace. We also propose parallel analysis (PA) to find out the optimal subspace dimensionality. Weighted projection distance between longitudinal subspaces is checked for concordance with the progression outcome using time dependent receiver operating characteristics (ROC). The preliminary clinical results suggest that higher order lesion textures have an added value in response evaluations.