• Chao Li, Wang Shuo S The Centre for Mathematical Imaging in Healthcare, Department of Pure Mathematics and Mathematical Statistics, University of Cambridge, Cambridge, UK. , Angela Serra, Turid Torheim, Jiun-Lin Yan, Natalie R Boonzaier, Yuan Huang, Tomasz Matys, Mary A McLean, Florian Markowetz, and Stephen J Price.
• Cambridge Brain Tumour Imaging Laboratory, Division of Neurosurgery, Department of Clinical Neurosciences, University of Cambridge, Box 167 Cambridge Biomedical Campus, Cambridge, CB2 0QQ, UK. cl109@outlook.com.
• Eur Radiol. 2019 Sep 1; 29 (9): 4718-4729.

ObjectivesIntegrating multiple imaging modalities is crucial for MRI data interpretation. The purpose of this study is to determine whether a previously proposed multi-view approach can effectively integrate the histogram features from multi-parametric MRI and whether the selected features can offer incremental prognostic values over clinical variables.MethodsEighty newly-diagnosed glioblastoma patients underwent surgery and chemoradiotherapy. Histogram features of diffusion and perfusion imaging were extracted from contrast-enhancing (CE) and non-enhancing (NE) regions independently. An unsupervised patient clustering was performed by the multi-view approach. Kaplan-Meier and Cox proportional hazards regression analyses were performed to evaluate the relevance of patient clustering to survival. The metabolic signatures of patient clusters were compared using multi-voxel spectroscopy analysis. The prognostic values of histogram features were evaluated by survival and ROC curve analyses.ResultsTwo patient clusters were generated, consisting of 53 and 27 patients respectively. Cluster 2 demonstrated better overall survival (OS) (p = 0.007) and progression-free survival (PFS) (p < 0.001) than Cluster 1. Cluster 2 displayed lower N-acetylaspartate/creatine ratio in NE region (p = 0.040). A higher mean value of anisotropic diffusion in NE region was associated with worse OS (hazard ratio [HR] = 1.40, p = 0.020) and PFS (HR = 1.36, p = 0.031). The seven features selected by this approach showed significantly incremental value in predicting 12-month OS (p = 0.020) and PFS (p = 0.022).ConclusionsThe multi-view clustering method can provide an effective integration of multi-parametric MRI. The histogram features selected may be used as potential prognostic markers.Key Points• Multi-parametric magnetic resonance imaging captures multi-faceted tumor physiology. • Contrast-enhancing and non-enhancing tumor regions represent different tumor components with distinct clinical relevance. • Multi-view data analysis offers a method which can effectively select and integrate multi-parametric and multi-regional imaging features.

29 keywords...

hide…