Table 2.
Habitat imaging in glioblastoma: the current evidence.
| Paper | Patients | Imaging habitats | Survival/progression | Histopathology | Molecular | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Stage of imaging | MRI sequences/biomarkers | Intensity normalisation | Biomarker clustering | Habitat generation | |||||||
| Lee 2015 Texas, USA (22) |
Glioblastoma N = 74 from The Cancer Genome Atlas |
Preoperative | T1+C FLAIR |
Yes | Gaussian mixture modelling - 2 clusters | Overlap of each cluster to produce 4 habitats | Top 5 spatial features had good area under curve = 0.76 for predicting survival | Top 5 spatial features had high accuracy for predicting subtypes - pro-neural 0.93, classical 0.88, neural 0.85, mesenchymal 0.70 | |||
| Lee 2015 Texas, USA (23) |
Glioblastoma N = 65 from The Cancer Genome Atlas |
Preoperative | T1+C FLAIR |
Yes | As above | Top 5 spatial features had high accuracy for predicting EGFR status (AUC 0.845) | |||||
| McGarry 2016 Single-centre Wisconsin, USA (24) |
Glioblastoma N = 81 |
Preoperative | T1 T1+C FLAIR ADC |
Yes | Automated tissue segmentation – 3 clusters | Overlap of each cluster to produce 83 habitats | Identified 5 habitats associated with shorter overall survival High intensity of both FLAIR and contrast enhancement present in 4 out of 5 habitats associated with overall survival |
One habitat could be hyper-cellular | |||
| Zhou 2016 Florida (25) |
Glioblastoma N = 32 from The Cancer Genome Atlas N = 22 internal cohort |
Preoperative | T1+C T2 FLAIR |
Yes | Otsu thresholding – 2 clusters | Interested in two habitats formed by overlap of high/high or low/low clusters | Spatial mapping between habitats were better at predicting survival than the presence of habitats themselves In particular, spatial mapping between high T2/high FLAIR region achieved >80% accuracy |
||||
| Khalifa 2016 Toulouse, France (16) |
Glioblastoma N = 15 Primary glioblastoma enrolled in a trial, ≤5cm diameter |
Postoperative pre-radiotherapy | FLAIR rCBV ADC |
NR | Gaussian mixture modelling to produce 2 (ADC) and 3 (rCBV) clusters | Interested in 4 habitats formed by various overlap combinations | Positive predictive value for recurrence was highest at 31.9% in overlap region of high FLAIR, hyper-perfusion and restricted diffusion | ||||
| Dextraze 2017 Texas, USA (26) |
Glioblastoma N = 85 from The Cancer Genome Atlas |
Preoperative | T1 T1+C T2 FLAIR |
Yes | K-means – 2 clusters | Overlap of each cluster to produce 16 habitats | All three habitats associated with survival present in high enhancing segment One habitat predictive of survival occurred in high FLAIR/high T1+C intensity One habitat in very periphery, two in enhancing core |
One habitat predictive of survival correlated with necrosis quantification | Habitats associated with various pathways including NFkB, DNA damage response/transduction and STAT1/NK activation | ||
| You 2018 Michigan, USA (27) |
Glioblastoma N = 21 |
Post-resection pre-radiotherapy | T1 T1+C T2 FLAIR DWI ADC FA CBV kTrans MET PET |
Yes | Hierarchical clustering to derive habitats | 5/10 patients had recurrence related to MET only area (metabolically active rim) | 1. High T2/FLAIR habitat mostly in necrotic core 2. High DWI habitat surrounded FLAIR region – core cellular component 3. High methionine habitat in periphery – metabolically active rim |
||||
| Stringfield 2019 Multi-centre (28) |
Glioblastoma N = 74 (37 in each cohort of long and short term survivors) |
Preoperative | T1 T1+C FLAIR |
Yes | Otsu thresholding – 2 (FLAIR) or 3 (T1+C) clusters | Overlap to produce 6 habitats | High FLAIR/high T1c habitat present in significantly higher volume in long term survivors | ||||
| Li 2019 Cambridge, UK (29) |
Glioblastoma Maximal resection Performance status 0-1 N = 112 |
Preoperative | ADC rCBV |
Yes | Quartiles – 4 clusters each | Interested in two habitats – lowest quartile rCBV/lowest quartile ADC, lowest quartile rCBV/highest quartile ADC | Higher volume of these habitats associated with better PFS Lactate/Creatine ratio in these regions associated with shorter PFS and OS |
Minimally invasive phenotype defined on DTI had lower proportion of the low rCBV/low ADC habitat | |||
| Alvarez-Torres 2019 Multi-centre (30) |
Glioblastoma (N = 184) | Preoperative | rCBV rCBF |
NR | Gaussian mixture modelling to yield 4 habitats per patient | Several habitats predictive of survival | |||||
| Park & Kim 2020-2021 Seoul, South Korea (31, 32) |
Glioblastoma (IDH wildtype; various N) | Post chemo-radiotherapy | T1+C ADC rCBV +/- EPT |
NR | K-means clustering to define various habitats using combinations of ADC, rCBV and EPT | Hypovascular cellular habitat (low rCBV and ADC) and hypovascular low conductivity (low rCBV and EPT) habitats strongly correlated with site of future progression. | |||||
| Xu 2021 New-York USA (33) |
Glioblastoma N = 263 from BraTS 2020 training dataset |
Preoperative | T1+C T2 FLAIR |
NR | Used simple linear interactive clustering (SLIC) – method that depends on intensity of pixels and their location | Graph features of habitats improved overall survival cox regression model | |||||
| Bailo 2022 Milan, Italy (34) |
High grade gliomas (including 12 glioblastomas) | Preoperative | Vp MD FAZA PET |
NR | Otsu thresholding to identify high/low regions of each biomarker in enhancing tumour or oedema | Multiple overlaps between clusters – 8 habitats possible | Habitats with high Vp/high FAZA uptake (regardless of MD) correlated with hyperplastic vessels and cellularity with low rate of necrosis. Largest volumetric representation was by ‘less aggressive’ habitats comprising low Vp/low FAZA uptake. These correlated with low cellularity and no signs of necrosis/angiogenesis. |
||||
| Yang 2022 Xi’an, China (9) |
Glioblastoma (test cohort of 122, validation cohort of 65 patients) | Preoperative | T1 T1+C T2 FLAIR |
Yes | Investigated oedema region only. Determined optimal number of K-means clusters using elbow plot method. Performed K-means clustering and produced 4 habitats. | Defined high risk habitat in oedema region based on radiomic features. This habitat improved performance of cox regression model of overall survival. | High risk habitat not correlated with MGMT methylation status | ||||
This table summarises data from 15 studies that have performed habitat imaging in glioblastoma. T1 + C, T1 with contrast; FLAIR, Fluid Attenuated Inversion Recovery; ADC, Apparent Diffusion Coefficient; DWI, Diffusion Weighted Imaging; FA, Fractional Anisotropy; CBV, cerebral blood volume; MET PET, methionine positron emission tomography; FAZA PET, 8F-labeled fluoroazomycinarabinoside PET, localises to hypoxic regions; EPT, Electrical Properties Tomography imaging.