Visualising spatial heterogeneity in glioblastoma using imaging habitats

Abstract

Glioblastoma is a high-grade aggressive neoplasm characterised by significant intra-tumoral spatial heterogeneity. Personalising therapy for this tumour requires non-invasive tools to visualise its heterogeneity to monitor treatment response on a regional level. To date, efforts to characterise glioblastoma's imaging features and heterogeneity have focussed on individual imaging biomarkers, or high-throughput radiomic approaches that consider a vast number of imaging variables across the tumour as a whole. Habitat imaging is a novel approach to cancer imaging that identifies tumour regions or 'habitats' based on shared imaging characteristics, usually defined using multiple imaging biomarkers. Habitat imaging reflects the evolution of imaging biomarkers and offers spatially preserved assessment of tumour physiological processes such perfusion and cellularity. This allows for regional assessment of treatment response to facilitate personalised therapy. In this review, we explore different methodologies to derive imaging habitats in glioblastoma, strategies to overcome its technical challenges, contrast experiences to other cancers, and describe potential clinical applications.

Keywords: MRI; biomarker; glioblastoma; habitats; heterogeneity; imaging; preoperative.

Figure 1

Clinical utility of habitat imaging in glioblastoma: assessment of changes pre and post-radiotherapy. This figure demonstrates the clinical utility of habitat imaging in glioblastoma pre and post-radiotherapy. Top row – structural imaging (T1 with contrast) demonstrates no significant changes in tumour anatomy. Middle two rows – diffusion and perfusion MRI scans demonstrate changes in tumour physiology with treatment with a decrease in rCBV for example (red to yellow represents low to high values for each biomarker). Bottom row – imaging habitats map where each voxel is labelled according to both rCBV and ADC values. This method produced 16 different habitats for this patient. After radiotherapy, the biggest increase was in a habitat defined by low rCBV and low ADC (10.5% increase). The biggest decrease was in a habitat defined by high rCBV and medium ADC (5.7% decrease). Habitats that are more resistant to treatment can be spatially visualised and offered targeted therapy. RT, radiotherapy; T1C, T1 with contrast; rCBV, relative cerebral blood volume normalised to contralateral white matter; ADC, apparent diffusion coefficient.

Figure 2

Habitat imaging methods in glioblastoma. This figure provides an overview of the two main approaches to deriving imaging habitats utilising local preoperative data from 12 patients with glioblastoma undergoing surgery. (A) one step approach: a multi-dimensional dataset can be produced utilising multiple imaging biomarkers from the same MRI acquisition (to avoid interpolation/registration errors), in this case Dynamic Contrast Enhanced (DCE) MRI. Data from R1_N – defined in Table 1 and three DCE-MRI imaging biomarkers (K ^trans, v_p and v_e ) were input into a machine learning K-means clustering algorithm to produce four distinct imaging habitats, that were distinct on Principal Component Analysis (PCA; right). A disadvantage of this approach is its ‘black-box’ nature, such that it is not straightforward to define each habitat for prospective validation. (B) Two step approach: this step first requires clustering of individual imaging biomarkers, in this case ADC and rCBV (left). Each pixel is then assigned to a habitat based on its ADC/rCBV cluster, with multiple cluster combinations defining each habitat (grey box). The advantage of this approach is that imaging biomarkers from different MRI acquisitions (e.g. diffusion and perfusion MRI) can be utilised. It is also easier to define each habitat as the definition of each is derived from its individual ADC/rCBV cluster composition. This approach therefore allows for prospective validation with pre-defined cluster thresholds.

Figure 3

The importance of considering group level data during clustering. This figure demonstrates the necessity of combining patient data for clustering. The top panel shows preoperative ADC data from 12 glioblastoma patients after clustering, demonstrating a histogram with a smooth gaussian shape. The bottom panel shows the results of clustering when data from only one individual patient is considered, revealing a more irregular histogram and different cut off values for each cluster. Corresponding cluster regions are displayed visually on the left of each panel. This technical consideration is of particular importance as it has implications for prospective habitat generation in validation cohorts, which is dependent on robust predefined cut offs.