Evaluation of feature-driven clustering of dynamic contrast enhanced and oxygen enhanced MRI data to assess tumour microenvironment heterogeneity
Adam K Featherstone1,2, James P B O'Connor2,3, Ross A Little1, Yvonne Watson1, Sue Cheung1, Kaye J Williams2,4, Julian C Matthews1,2, and Geoff J M Parker1,2,5

1Centre for Imaging Sciences, The University of Manchester, Manchester, United Kingdom, 2CRUK & EPSRC Cancer Imaging Centre in Cambridge and Manchester, Cambridge and Manchester, United Kingdom, 3Institute of Cancer Sciences, The University of Manchester, Manchester, United Kingdom, 4School of Pharmacy, The University of Manchester, Manchester, United Kingdom, 5Bioxydyn Ltd., Manchester, United Kingdom


DCE-MRI and OE-MRI scans were performed on 8 preclinical U87 tumour xenografts. Heuristic features (area-under-curve and rate-of-enhancement) were calculated from tumour voxel enhancement curves for each imaging modality. Clustering algorithms (k-means clustering and Gaussian mixture modelling) were applied to these features and native tissue T1 to investigate their utility in characterising physiological heterogeneity in tumours. Efficacy in identifying large regions where there is agreement between features is shown. Further optimisation is needed to optimise the approach to characterise smaller, and potentially important, regions where there is a lack of concordance between features.


There is well-documented variation in biological structure and physiology within tumours. Heterogeneity of tumour tissue, specifically the occurrence of necrotic and hypoxic regions, is a marker of poor prognosis.[1] Non-invasive determination of the location of distinct physiological characteristics is a clinically relevant challenge whose solution could have application in therapeutic planning and assessment of response to intervention.

DCE-MRI can provide information on tumour tissue perfusion and contrast agent leakage and OE-MRI can provide information on tissue oxygenation.[2, 3] Previous work has characterised tumour tissue type by applying ad hoc thresholds to area under the curve (AUC) values for both modalities and has shown that biomarkers from combined DCE-MRI and OE-MRI relate to tumour hypoxia and necrosis.[4] In the current work we investigate the application of data clustering methods to a larger range of features calculated from combined DCE-MRI and OE-MRI.


Data Acquisition

U87 tumour xenografts were grown subcutaneously in 8 nude mice (labelled A-H). MR imaging was performed on a 7 T Bruker preclinical system running ParaVision 5.1. All mice were imaged at baseline and mice A-D were also imaged 3 days later. Each session consisted of 1) a 3D variable flip angle (VFA) spoiled gradient recalled echo (SPGR) acquisition 2) a 3D SPGR OE-MRI acquisition in which the gas supply was switched from air to 100% oxygen via a breathing mask, 3) a second 3D VFA SPGR acquisition and 4) a 3D SPGR DCE-MRI acquisition during which Gd-DOTA was injected via a tail vein.

Data Processing

Native T1 (T10) maps were calculated via a linear least squares fitting of the SPGR signal equation to VFA data. Dynamic DCE-MRI and OE-MRI signals were converted to ∆R1(t) values (change from native R1 = 1/T1) and corrected for baseline drift. Four extra features were calculated from each tumour voxel’s ∆R1(t) curves: initial AUC for DCE-MRI (iAUCDCE), total AUC for OE-MRI (AUCOE) and mean enhancement rate for DCE-MRI and OE-MRI (rateDCE, rateOE).

Data Clustering

K-means clustering (KMC) and Gaussian mixture modelling (GMM) were applied to the pooled features from all tumours with the number of clusters (k) ranging between 2-6. Average DCE-MRI and OE-MRI ∆R1(t) enhancement curves per cluster were calculated for each result. AIC (for GMM) and silhouette score (for KMC) values were calculated for clustering results. These were considered in combination with visual inspection of the clustering results in feature space and separation of average enhancement curves to select an appropriate clustering method.

Results and Discussion

Example clustering outputs, using GMM with k = 3, are shown in Figure 1. Strong correlations are observed between DCE-MRI features (r = 0.91) and between OE-MRI features (r = 0.71). Figure 2 shows average enhancement curves per cluster.

Figure 3 shows the clustering result mapped into image space, alongside maps of the features that were clustered on. Cluster maps show a high level of region contiguity and distributions of functional activity mostly consistent with a rim-core structure. Large areas of agreement between DCE-MRI and OE-MRI features are identified in the cluster maps, but smaller areas of disagreement (Figure 3, regions highlighted by arrows) are not. These regions hold potentially important physiological information[5] but the relatively low numbers of voxels in these regions is most likely why these regions are neither readily apparent in feature space (Figure 1) nor identified by clustering in the tumours studied.


This work has shown that applying clustering algorithms to combined DCE-MRI and OE-MRI data can locate physiological variation within tumours. However, in the tumours studied, this approach doesn’t identify regions containing relatively small numbers of voxels that may contain important information. Additional information is likely to be required to enable automated clustering methods to identify these regions, followed by histopathological comparison. The approach may then be optimised for clinical data to create a non-invasive tool for characterising tumour tissue perfusion and oxygenation heterogeneity, useful for predicting and monitoring treatment response.


This is a contribution from the CRUK & EPSRC Cancer Imaging Centre in Cambridge and Manchester (C8742/A18097).


[1] A. A. Alizadeh et al. Toward understanding and exploiting tumor heterogeneity. Nat Med 21(8):846-853.

[2] T. E. Yankeelov and J. C. Gore. Dynamic contrast enhanced magnetic resonance imaging in oncology: Theory, data acquisition, analysis, and examples. Curr Med Imaging Rev, 3(2):91–107, 2009.

[3] Y. Ohno and H. Hatabu. Basics concepts and clinical applications of oxygen-enhanced mr imaging. Eur J Radiol, 64(3):320–8, 2007.

[4] J. P. B. O’Connor et al. Oxygen-enhanced MRI accurately identifies, quantifies, and maps tumor hypoxia in preclinical cancer models. Cancer Res 76, In Press (2016).

[5] I. Linnik et al. Noninvasive tumor hypoxia measurement using magnetic resonance imaging in murine U87 glioma xenografts and in patients with glioblastoma. Magn Reson Med, 71(5):1854–62, 2014.


Figure 1: Matrix of feature plots. Colours illustrate the results of clustering the data across the five dimensional feature space. Correlation coefficients (r) are displayed for each pair of features.

Figure 2: Average ∆R1 enhancement curves for DCE-MRI and OE-MRI for each cluster. Colours match those in Figure 1 and in cluster maps in Figure 3.

Figure 3: Cluster maps and feature maps for a central slice of tumours A-H at session 1. Arrows indicate regions of large disagreement between DCE-MRI and OE-MRI features.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)