Analysis of PET-CT Scans through Machine Learning for the Diagnosis of Skeletal Metastases

We examine the predictive value of computed tomography (CT) data for high uptake in positron emission tomography (PET) for patients suspected of skeletal metastases. Our database consists of PET-CT scans of 457 patients, where training and testing was performed on a relevant subset of 142 patients. 1524 unique pairs of registered CT and PET slices are used to examine whether PET uptake can be predicted. We show spectral total- variation (STV) features are highly useful for the task. STV features are based on the nonlinear eigenfunctions of the one-Laplacian, and are well suited for medical imaging. The low-correlation of spectral bands facilitates the use of STV within ensembles. We compare our approach to deep-learning methods and to radiomic features. This turns out to be a complex classification problem. Our findings are that all methods are able to predict, to some extent, the PET uptake: the best neural net configuration (transfer learning based on RadImageNet) with AUC=0.75, Radiomics with AUC=0.79 and STV with AUC=0.87. For the available data, classical methods outperform neural nets and expose more transparently features relevant for classification. We observe that fine STV scales in CT images are especially indicative for the presence of high uptake in PET.

M.Sc. student under the supervision of Prof. Guy Gilboa.