The Influence of Local Variation and Local Similarity on Tumour Subregional Analysis

The Influence of Local Variation and Local Similarities on Tumour Subregional Analysis
Jola Mirecka1, Benjamin J Irving1, P Danny Allen2, Paul Kinchesh2, Stuart Gilchrist2, Ana L Gomes2, Veerle Kersemans2, Sean Smart2, Michael Chappell1, Julia A Schnabel1,3, Mark Jenkinson4

1 Institute of Biomedical Engineering (Department of Engineering Science), University of Oxford, UK;
2 Department of Oncology, University of Oxford, UK;
3 School of Biomedical Engineering and Imaging Sciences, King’s College London, UK;
4 Oxford Centre for Functional MRI of the Brain (FMRIB), Nuffield Department of Clinical Neurosciences, University of Oxford, UK.


In tumours, the process of angiogenesis often results in increasingly more disorganised and leaky vasculature as tumours progress. This may lead to a highly heterogeneous perfusion, which is known to produce locally varied subregions in tumours. Tracking such subregional differences might be the key to understanding and predicting tumour progression. Dynamic contrast-enhanced MRI (DCE-MRI) enables monitoring such changes in perfusion and vascularity, which can be in turn quantified with pharmacokinetic (PK) modelling resulting in a set of parameter maps. Subsequently, segmentation methods are often applied onto the maps to define tumour subregions that can be later tracked over time.

While there is a multitude of methods available for segmentation, different approaches carry certain assumptions about the composition and distribution of data. The choice of subregional analysis might highly influence the segmentation outcome. In this work, we compare two approaches to subregional analysis – supervoxel-based (from local similarity point of view) [1]2 and patch-based (local variation) [3]. Finally we combine the advantages of both methods in a single variation-guided supervoxels method.


We evaluate both methods on a dataset of 10 mice scanned over 10 days. The scans were acquired with a 4.7T DCE-MRI (TR 1.4 ms, TE 0.64 ms, FA 5°). Extravascular-extracellular space enhancement was modelled using the Toft’s model and population-derived arterial input function. Subsequently, both methods extract different features from 3D PK maps and follow with K-means clustering to learn Bag of Visual Words (BOV) model that describes the most dominant visual words. Using KNN the words are then mapped back onto images resulting in segmentations. In patch-based method [3], raw values from densely-extracted 26-point rigid neighbourhoods were used as features to capture the variability in local perfusion patterns. In supervoxel-based method [1]2, mSLIC was used on the maps to group similarly perfused areas (supervoxels) and mean values from supervoxels were used as features. Finally. in variation-guided mSLIC (VG-mSLIC), patches were used to guide the supervoxels.


We use heatmaps representing the confidence of each voxel belonging to each subregion as a measure of contiguity of segmentations. The confidence is based on the percentile distance from the cluster centre. For each tumour we then calculate the mean percentile value, of which the mean for the whole cohort was 0.002 and 0.0015 for supervoxel and patch methods respectively, suggesting that most voxels were placed close to the cluster centres, resulting in high confidence of segmentations (Fig. 1). While both methods were able to produce contiguous subregions (both visually and based on our metric), there is a noticeable difference in spatial distribution of different subregional labels (Fig. 2).

Fig. 1: Top row: segmentations, bottom row: heatmaps of confidence of segmentation for each voxel. The scale relates to the value of the heatmaps. The heatmaps are based on the percentile distance from the cluster centre representing the confidence of the voxel belonging to the cluster. Most values lay well within the small percentile of the cluster centre (bottom row averages).

Fig. 2: Comparison of different subregional segmentations on a progressing tumour (columns relate to days of progression). First row represents Ktrans maps. Second row represents visual words learned with prior supervoxelisation (local similarities). Third row relates to visual words learned from patches (local variation). Label colours are consistent across the rows and columns.


The choice of segmentation method should be of special importance, as different tumour models can exhibit different homogeneous or heterogeneous properties throughout growth. While both methods in comparison seem to produce meaningful segmentations, the spatial distribution of labels differs, indicating that different types of tumours/tumour models may benefit from different methods.


[1] Irving, B.J., Mirecka, J., Gomes, A.L., Allen, P. D., Kinchesh, P., Keremsans, V., Gilchrist, S., Smart, S., Schnabel, J.A., Brady, J.M., Chappell, M.: “Perfusion- supervoxels for DCE-MRI based tumor subregion assessment.” In proceedings of the 25th Annual Meeting of ISMRM (2017).

[2] Irving, B., Popescu, I. A., Bates, R., Allen, P .D., Gomes, A. L., Kannan, P., Kinchesh, P., Gilchrist, S., Kersemans, V., Smart, S., Schnabel, J. A.: “maskSLIC: Regional Superpixel Generation with Application to Local Pathology Characterisation in Medical Images.”

[3] Mirecka, J., Irving, B., Kannan, P., Kersemans, V., Allen, P. D., Smart, S. C., Kinchesh, P., Muschel, R. J., Jenkinson, M. and Schnabel, J. A.: “Statistical Texture Modeling of Tumour Perfusion Heterogeneity in Dynamic Contrast-Enhanced MRI”. Computational Methods for Molecular Imaging, MICCAI (2015).