Fully Incomplete Information for Multiview Clustering in Postoperative Liver Tumor Diagnoses

Multiview clustering (MVC) is a proven, effective approach to boosting the various downstream tasks given by unlabeled data. In contemporary society, domain-specific multiview data, such as multiphase postoperative liver tumor contrast-enhanced computed tomography (CECT) images, may be vulnerable to exploitation by illicit organizations or may not be comprehensively collected due to patient privacy concerns. Thus, these can be modeled as incomplete multiview clustering (IMVC) problems. Most existing IMVC methods have three issues: (1) most methods rely on paired views, which are often unavailable in clinical practice; (2) directly predicting the features of missing views may omit key features; and (3) recovered views still have subtle differences from the originals. To overcome these challenges, we proposed a novel framework named fuzzy clustering combined with information theory arithmetic based on feature reconstruction (FCITAFR). Specifically, we propose a method for reconstructing the characteristics of prevailing perspectives for each sample. Based on this, we utilized the reconstructed features to predict the missing views. Then, based on the predicted features, we used variational fuzzy c-means clustering (FCM) combined with information theory to learn the mutual information among views. The experimental results indicate the advantages of FCITAFR in comparison to state-of-the-art methods, on both in-house and external datasets, in terms of accuracy (ACC) (77.5%), normalized mutual information (NMI) (37.9%), and adjusted rand index (ARI) (29.5%).