The search for CT imaging subtypes of colorectal liver metastases and the impacts of slice thickness

This paper explores the potential for computed tomography (CT) imaging subtypes of colorectal liver metastases (CRLMs) through unsupervised analyses of classical radiomic, topological and convolutional neural network (CNN) imaging features while also investigating the influence slice thickness has on the forming of these subtype groupings. A multi-center cohort of 1,199 patients with preoperative, portal-venous phase CT imaging of resectable CRLMs was used for this analysis. PCA and t-SNE were used to visualize potential slice thickness associations with classical radiomic, topological, and CNN features. Clusters in PCA were defined using K-means clustering for each feature type before cluster association with overall survival and hepatic-disease free survival was quantified. We found that classical radiomic first order and second order texture features clustered by slice thickness using both PCA and t-SNE approaches. More distinct groupings were shown in these feature groups when they were extracted from the liver parenchyma. Both topological and CNN imaging features formed no visual patterns associated with slice thickness. Significant differences in hepatic disease-free survival were seen between clusters for liver parenchymal classical radiomic first order and texture features as well as topological persistent statistics features. In overall survival, clusters from classical radiomic shape and topological persistent landscape were significant. When grouping by slice thickness, significant differences in overall survival were seen. The significance seen from the classical radiomic and topological features suggests the existence of imaging subtypes of CRLMs. Our work demonstrates the potential for topological approaches to be used to develop biomarkers robust to protocol influences and the importance of accounting for slice thickness when using classical radiomic features.