Background

Primary brain tumours account for about 2% of all cancers in the US with an incidence of about 23 per 100,000. Gliomas account for 80.6% of all malignant brain tumours.1 The incidence is highest for glioblastoma (3.21 per 100,000 population), followed by diffuse astrocytoma (0.46 per 100,000 population). The age-adjusted mortality rate is 4.4 per 100,000 and the 5-year survival rate is 35%.1 However, the rate varies significantly by age at diagnosis and the histology of the tumour.

Advances in our understanding of the molecular pathogenesis of gliomas has prompted significant changes to the World Health Organization (WHO) classification of central nervous system (CNS) tumours in 2016.2 Previously, the classification criteria was based solely on microscopic features.3 The new criteria reclassifies entities with the incorporation of genetic information in certain tumours. These changes were incorporated because of the impact genetic factors have on tumorigenesis and subsequent therapy.

In today’s era of modern imaging, accurate non-invasive prediction of glioma grade/type, survival, and treatment response remains challenging. Stereotactic brain biopsy, despite being invasive and costly, remains the reference standard for histological and genetic classification; however, pathological diagnosis may still remain inconclusive in 7–15% of patients.4 This necessitates imaging surrogates to characterise tumour heterogeneity. Recently, multiple studies have shown strong association between morphological features from multiparametric magnetic resonance imaging (MRI) and survival.5,6,7,8,9,10 Similarly, functional imaging techniques such as perfusion weighted MRI and magnetic resonance spectroscopy (MRS) have been shown to be beneficial when used along with morphological features, but with limited success and reproducibility.11,12,13,14,15,16 The limitations in current imaging techniques provide an opportunity for more sophisticated sub-visual feature analysis to augment the morphological features and current functional imaging capabilities.

Radiomics refers to the computerised extraction of quantifiable data from radiological images in the form of radiographical cues that are usually sub-visual.17,18 The extraction of these data creates mineable databases from radiological images which can be used for diagnosis, prognosis characterisation, and to assess or predict response to certain therapies.19,20,21 Genetic mutations often determine the aggressiveness of the tumour and have been shown to be associated with a lesion’s growth pattern and response to therapy. Radiomic features have been shown to identify genomic alterations within tumour DNA and RNA. The integrated study of data from radiographical and the genomic scales is termed radiogenomics.

In this review, we describe the applications of radiomics and radiogenomics from the perspective of neuroradiologists, neurosurgeons, and neuro-oncologists. Specifically, we review work that highlights the importance of the evolving field for diagnosing and predicting prognosis of individuals with different brain tumour types. Additionally, we discuss the potential and importance of integrating these applications into radiological workflows to improve patient care and outcome.

Overview of radiomics and radiogenomics workflow

Radiomics is an emergent field that involves converting radiological images into high-dimensional mineable data in a high-throughput fashion. This multi-step process involves (a) image acquisition and reconstruction, (b) image pre-processing, (c) identification of regions of interest, (d) feature extraction and quantification, (e) feature selection, and (f) building predictive and prognostic models using machine learning (Fig. 1).22 To account for MRI intensity non-uniformity, inter- and intra-site scanner variability, image processing routines, such as intensity normalisation, voxel intensity calibration and bias field correction, are used as a precursor to radiomic feature extraction.23,24,25 The segmentation of the regions of interest (ROI) can be achieved by manual, semi-automated or fully-automated methods.19,26,27,28,29 Radiomic features are then extracted from the identified ROIs. Common features can be divided into the following groups: morphological radiomics, textural radiomics, and functional radiomics.

Fig. 1: Radiomics and Radiogenomics workflow.
figure 1

Use of Radiomics and Radiogenomics pipelines in personalized medicine.

Full size image

Following feature extraction, different statistical methods are used to select a subset of top features that correlate with the expected outcome.30 Commonly used feature selection algorithms include minimum redundancy maximum relevance (mRMR) algorithm31,32 and sequential feature selection methods.33 Feature selection is performed in order to reduce potential model overfitting associated with the high dimensionality of the radiomic feature set. Once top features are identified, machine learning classifiers and other statistical methods such as the Cox-proportional Hazards modelling techniques34 are used to build predictive and prognostic models. Sanduleanu et al. proposed a “radiomics quality score” tool to assess the quality of the radiomics research study linking tumour biology; however, interpretability of the outcomes of those scores is still questionable.35

The recent advent of radiogenomics has also accelerated the integration of multi-omic data for accurate diagnosis and improved personalised cancer treatments. The first step of the radiogenomic pipeline in neuro-oncology (Fig. 1) is to acquire genomic material via a fresh frozen paraffin embedded (FFPE) sample or a tissue microarray (TMA) sample obtained from a stereotactic brain biopsy from within the brain tumour. Next, bioinformatics techniques, such as sequencing, can detect single-gene mutations. For instance, epidermal growth factor receptor (EGFR) amplification, O6-methylguanine-methyltransferase (MGMT) methylation can be detected by analysing the proteins through immunohistochemical (IHC) analysis and next-generation sequencing (NGS) techniques such as mRNA sequencing. mRNA sequencing, whole-exome sequencing, and whole-genome sequencing can help detect multi-gene expression anomalies. The decisive goal of radiogenomic analysis involves associating gene mutations and pathways directly with distinct imaging phenotypes.

Radiomic feature groups

Morphological radiomics

Morphological radiomic features are used to quantify lesion topology induced by the proliferating boundaries. These can be further divided into global and local morphological features. Global features characterise the contour of the lesion by extracting measurements such as roundness, perimeter, diameters of major and minor axes, and elongation factor. Local morphological features characterise the surface curvature attributes derived from isosurfaces.26,36 These comprise quantitative measurements such as degree of curvature (curvedness) and degree of sharpness.

Textural radiomics

Structural texture analysis

Structural methods describe texture by identifying structural primitives and their placement rules. Multi-scale, multi-resolution steerable bandpass filters like Gabor filter banks37,38 are among the most widely used orientation-based structural descriptors. Gabor descriptors are modelled to mimic the way the human visual system deciphers object appearances, by decomposing the original image into filter responses of a sinusoidal wave of multiple frequencies and orientations. Gabor filters have been shown to distinguish pathologies on histology samples as demonstrated by Doyle et al.39

Statistical texture analysis

Statistical methods analyse the spatial distribution of grey values by computing local features at each image point and deriving a set of statistics from the distribution of local features. One commonly used statistical technique for identifying shape-based object classes is histogram of oriented gradients (HOG).40 Traditionally, the applicability of HOG has been demonstrated for detection of human forms in cluttered images. Multi-coordinate HOG can distinguish different categories of lung tissues in high-resolution tomography images. It characterises local object appearance and shape by computing distribution of local intensity gradients. Grey level co-occurrence matrix (GLCM) features, popularly known as Haralick features41 and originally designed for aerial photography, utilise the values of distance and angle for a combination of grey levels.

Texture analysis using a combination of statistical and structural techniques

Local binary patterns (LBP) is a textural operator that combines statistical and structural methods in appearance classification. LBP is robust with regards to illumination changes and has been shown to be useful in medical datasets which are corrupted by patient motion artefacts. This feature presents texture information as a joint distribution of the intensity of a central pixel and that of its neighbors.42 Li et al.43 demonstrated the use of LBP along with neural networks to classify endoscopic images. Another feature that combines statistical and structural techniques is the co-occurrence of local anisotropic gradient orientations (CoLlAGe) descriptor which seeks to capture and exploit local anisotropic differences in voxel-level gradient orientations to distinguish similar appearing phenotypes.44

Functional radiomics

A critical obstacle to the clinical adoption of traditional radiomic features is its low biological interpretability. To qualify as a biomarker, an attribute should not only be measurable and reproducible, but also be reflective of the underlying anatomy or physiology. It is imperative to discover radiomic signatures that are biologically relevant. Functional radiomic markers are a new class of markers which specifically target the issue of ‘interpretability’ by modelling features that directly capture underlying physiological properties such as angiogenesis. Properties of vessels feeding the lesions (such as convolutedness, density) play an important role in the drugs’ ultimate response. Recently, tortuosity-based features capturing local and global disorder in vessel network arrangement have been shown to be effective in diagnosis and treatment response assessment.45 Deformation descriptors are another class of functional radiomics markers which seek to measure tissue deformation in the brain parenchyma due to mass effect.46 These features provide an insight into the microenvironment outside the visible surgical margins.

Vessel architecture imaging (VAI) MRI is a technique that non-invasively measures parameters to describe structural heterogeneity of brain microvasculature.47,48,49 The different gradient echo (GE) and spin echo (SE) images produce an apparent different variation in the MRI readout based on the structural and physiological properties of the vessels. Stadlbauer et al.47 examined gliomas (n = 60) using vascular architectural map** (VAM). They introduced three new VAM biomarkers (i) microvessel type indicator (MTI), (ii) vascular-induced bolus peak time shift (VIPS), and (iii) the curvature (Curv) and adapted known parameters, microvessel radius (RU) and density (NU). MTI and VIPS parameters were helpful in detecting neovascularisation, especially in the tumour core of the HGGs, whereas curvature showed peritumoral vasogenic oedema which correlated with neovascularisation in the tumour core of HGGs. These biomarkers gave insight into complexity and heterogeneity of vascular changes in gliomas to differentiate HGGs versus LGGs.50 Furthermore, combining multiparametric quantitative blood oxygenation level-dependent approach (qBOLD) with VAM parameters helped distinguish LGGs versus HGGs and identify isocitrate dehydrogenase (IDH) mutation status with higher sensitivity.50 Stadlbauer et al.51 also performed analysis of vascular hysteresis loop (VHLs) in combination with the VAM biomarkers to assess response of glioblastoma to anti-angiogenic therapy. MTI was found to be useful to predict responding versus non-responding regions, whereas, Curv was better to assess severity of vasogenic oedema. Price et al.52 used diffuse tensor imaging (DTI) with MR perfusion and MRS imaging to determine changes in the invasive versus non-invasive margins of glioblastomas to better predict treatment efficacy and overall survival.53,54

Semantic features

Semantic features, such as tumour location, shape, and geometric properties on structural MRI,19,55 are qualitative features used by neuroradiologists to describe the tumour environment. Previous studies have found that semantic features are related to the genetic phenotype of brain tumours.56 The Visually AcceSAble Rembrandt Images (VASARI) project by TCIA established a feature set to enable consistent description of gliomas using a set of defined visual features and controlled vocabulary.57 Studies have shown that these features are highly reproducible and provide meaningful guidance in glioblastomas.5 Semantic features are also robust to changes in image acquisition parameters and noise and can be used along with more sophisticated radiomic features in machine learning settings.22

Diagnostic applications

Differentiating tumours based on texture analysis

Many studies have shown the application of textural analysis for differentiating HGGs from LGGs. Skogen et al.58 applied a filtration-histogram technique for characterising tumour heterogeneity. In a cohort of 95 patients (27-grade II, 34-grade III, and 34-grade IV), by using standard deviation (SD) at a fine texture scale, they were able to distinguish LGGs from HGGs with sensitivity and specificity of 93% and 81% (AUC 0.91, P < 0.0001). Tian59 et al. applied textural analysis on multiparametric MRI of 153 patients and reported an accuracy of 96.8% for classifying LGGs from HGGs and 98.1% for classifying grade III from grade IV using an SVM classifier. ** may allow for personalised therapeutic decisions in LGG.

Multiple groups have evaluated radiomic features to determine molecular phenotype of gliomas.62 Zhang et al.63 extracted 15 optimal radiomic features (n = 152) using SVM-recursive feature elimination (SVM-RFE) that could detect IDH mutation with accuracy of 82.2%. Han et al.103 Interestingly, multiple radiogenomic correlation experiments have revealed strong associations of imaging phenotypes with pathways that are implicated in extracellular matrix destruction, cell invasion and metabolism.

Furthermore, radiomics offers an opportunity to perform an analysis on complete tumour that could mitigate the limitation of sampling errors and inability of complete molecular and histopathological assessment by neuropathologists given the lack of tumour sample.104,105,106 With quantitative mutation values rather than binary designations, radiomics can help neuro-oncologists and neurosurgeons make personalised therapy decisions and reliably predict response to therapies.

Limitations

A major feature limiting radiomic quantification is poor reproducibility secondary to variability and lack of consistency attributed to the absence of standardised acquisition parameters and radiomic approaches.107 The accuracy of radiomic signatures typically varies when tested on different datasets. Multiple studies have addressed impact of different acquisition parameters on textural analysis. Magnet strength, flip angles, different spatial/matrix size, TR/TE variations in T1WI and T2WI, and different scanner platforms can affect texture features.108,109,110,111,112 Molina et al.112 found that no textural measures were robust under dynamic range changes, entropy was the only robust feature under spatial resolution changes. Buch et al.109 concluded that some of the features were more robust and some of the features were more susceptible to different acquisition parameters, necessitating the need for standardised MRI techniques for textural analysis. Furthermore, variation in usability of textural analysis software add complexity to standardisation and reproducibility. Multiple studies used indigenous software with varying algorithms making reproducibility and repeatability of these studies almost impossible. Future studies are needed to assess accuracy of these results from different type of software to help with standardisation.

Scarcity of publicly available databases with annotated radiological studies for specific clinical domains limit capability of researchers to conduct large sample size studies. Small sample size and a high number of prediction variables often leads to overfitting, a major limitation in machine learning models. To prevent overfitting, it is recommended to have sample size 6–10 times larger than the analysed variables or conducting analysis with a few preselected robust variables only. Collaboration among research universities is required to create professionally annotated standardised datasets for larger cohort studies which can be split into training, testing, and validation datasets to avoid overfitting. This would also allow the researchers to test their algorithms on external cohorts and validating robustness of their solutions. A recent development towards achieving this is the use of federated learning which facilitates multi-institutional validation of machine learning models without explicit sharing of data using a distributed framework.113

Variability in the selection of appropriate regions of interest for feature extraction can affect certain radiomic attributes, such as shape-based measures. There are no existing guidelines for radiologists to report quantitative imaging features, making huge existing image repositories inaccessible for curation. For generating high-quality data with segmented and annotated appropriate regions of interest, radiologists need to be integral part of data quantification and curation.114

The lack of routinely acquired gene expression profiles and tissue sampling errors impose limitations to the application of radiogenomics in current clinical workflow.115 It is difficult for a single institute to create a large imaging database with auxiliary data such as genomic profile, demographics, treatment information and their outcomes. The Cancer Genome Atlas (TCGA) has made cancer datasets publicly available with a comprehensive catalogue of genomic profiles to address this issue. The clinical translation of radiogenomics is also hindered by spatial and temporal heterogeneity within a given brain tumour. However, ability of radiomics to perform an analysis on complete tumour might address this limitation.

Deep learning algorithms that facilitate automated feature learning have recently shown great promise in tasks ranging from tumour segmentation119 Test–retest experimental settings have also been widely proposed to facilitate selection of stable and robust radiomic measures. In one of the first studies of this kind, the repeatability of CT radiomics was ascertained in a “coffee-break” test–retest setting with scans obtained from the same scanner within an interval of 15 min.120 Similar settings are warranted for brain imaging to identify suitable radiomic features for clinical applications. The Cancer Imaging Archive hosts imaging datasets of brain tumour collections (HGGs and LGGs), among other cancers, obtained from several institutions. Such datasets have been widely used by the research community to develop and validate radiomics and radiogenomics tools.

One obvious deficiency in virtually all the retrospective radiogenomic studies is lack of information regarding the location of the biopsy sample vis-à-vis pixels in the patient’s images.121 Image-localised biopsies and subsequent imaging-pathology co-registration are essential steps in mitigating biases associated with locating biopsy region on MRI. Hu et al.122 have previously co-registered MRI scans and corresponding texture maps with biopsy locations to study regional genetic variation with spatially matched imaging descriptors. In a follow-up study, Hu et al.123 have proposed a Gaussian process and transductive learning based probabilistic model to quantify spatial uncertainty in radiogenomic pipelines. The sparse availability of ground truth labels in radiogenomic models can be modelled as a ‘weak supervision’ or ‘incomplete supervision’ task. Multi-instance learning techniques may be implemented to address this limitation.124

The field of radiomics promises to elevate the role of medical imaging by enabling objective tumour characterisation. In oncology, radiomics can provide prognostic information non-invasively via biomarker utilisation.19 In neurosurgery, it can be used for improved pre- and post-operative treatment planning. Emerging companies are now providing software that delivers web-based radiological analysis with the PACS viewing system.125 Support from academic institutions, such as the American College of Radiology, is growing. Such efforts can readily facilitate the transition of radiomic research to clinical practice.

Conclusions

Radiomics is not intended to replace radiologists in the future, but rather improve disease diagnosis and characterisation with greater precision. It is imperative for the future of neuroradiology, neurosurgery and neuro-oncology to utilise advances in radiomics and radiogenomics in order to provide less invasive and tumour-specific precision treatment strategies and to ultimately optimise patient care. In order for this field to continue evolving and make its way into clinical practice, it is vital to develop more standardised and reproducible methods of data interpretation, maintain publicly available databases of radiological studies, and conduct prospective large-scale multi-institutional clinical trials. In the future, the fields of radiomics and radiogenomics promise to improve the utility of already available imaging modalities and channel them towards personalised medicine.