Imaging tumour heterogeneity
Session type: Symposia
It is becoming increasingly appreciated that malignant cancers can be characterised by genetic plasticity within highly selective local microenvironments. This combination promotes somatic evolution and the emergences of clades of cells in spatially explicit micro-habitats. Malignancy can be defined by these habitats, which increase the probability that cancers will develop therapy resistant phenotypes. Heterogeneity across cancers has been known at the nuclear level for years and is recognised as a strong predictor of poor prognosis. For most advanced cancers and most patients, response to therapy is fleeting, owing to the inevitable evolution and proliferation of a resistant population
Induction of genomic alterations and localised selection by heritable and/or environmental factors will result in phenotypic heterogeneity. Phenotypic and physiologic heterogeneity can be viewed radiographically, wherein non-uniform patterns of enhancement or attenuation can be associated with poor outcome. In order to systematically address this issue, we have created a database structure that can be populated with images, as well as quantitative image feature data that can be mined in combination with patient outcomes and genetic data from biopsies. This enterprise is termed ‘Radiomics', which allows real-time data analyses and association of features with prognostic, diagnostic and predictive models. The goal of radiomics is to convert images to mineable data, with high fidelity and high throughput. The radiomics enterprise can be divided into a pipeline of processes with definable inputs and outputs, each with its own challenges that need to be overcome. Each of these steps must be developed de novo and, as such, poses discrete challenges that have to be met. Even though this field is in its infancy, meaningful classifier models have been generated in detecting and diagnosing a number of cancer subtypes.
To date, the radiomics effort has focused on agnostic and semantic image features, which quantify indescribable and describable features, respectively. An example of an agnostic feature is image ‘entropy', determined by the randomness of pixel intensities in a near-neighbourhood. Examples of a sematic feature are ‘spiculated', ‘spherical', ‘central necrosis'' which are commonly used as descriptors for tumour anatomy. These have all been shown to have high prognostic value in non-small cell lung cancer (NSCLC) and are being used to classify indeterminate lung nodules in lung screening CTs. More recently, we have been combining orthogonal MR images (e.g. STIR, Diffusion and Contrast enhanced T1) to develop data cubes for each voxel, which can then be clustered using fuzzy logic to identify specific sub-tumoural ‘habitats'; each with their own unique combination of perfusion, lipid/water ratio and cellular density. We hypothesise that these habitats describe specific sub-tumoural regions associated with genetic clades, and hence may inform the application of targeted therapy.
Thank you to Olya Grove and Robert A Gatenby (H. Lee Moffitt Cancer Center and Research Institute, Tampa, USA) who also contributed to this work.