Recent years have witnessed growth in the medical imaging field. The large number of medical images has brought new
opportunities to researchers: a) they can give more robust and reliable medical image analysis reports
based on a large amount of images; and b) they can develop more advanced methods that rely on a large
amount of samples, such as data mining and machine learning, which can be introduced into the field of
medical image analysis. The fast evolution of medical imaging has fostered a comprehensive analysis method
for medical images called radiomics. Radiomics generally refers to the extraction and analysis of large
amounts of advanced quantitative imaging features with high throughput from medical images obtained using
computed tomography (CT), positron emission tomography (PET) or magnetic resonance imaging (MRI) (Kumar,
Gu et al. 2012, Lambin, Rios-Velazquez et al. 2012, Aerts, Velazquez et al. 2014, Gillies, Kinahan et
al. 2015). In comparison with the traditional practice of treating medical images as pictures intended
solely for visual inspection, the aim of radiomics’ analysis is the transformation of medical images
into minable data combined with other patient clinical information. Using sophisticated bioinformatics
tools, researchers are able to develop models that may potentially improve diagnostic, prognostic, and
predictive accuracy (Gillies, Kinahan et al. 2015).
Radiomics’ analysis is mainly performed on medical images such as CT images where tumors are clearly visible. Tumors
may be significantly different in terms of their properties, e.g., their shape, size, and composition.
These characteristics are visible in CT images and most of them can be extracted using their mathematical
expression. The radiomics’ process is similar to a conventional CAD process in that they both extract
features from medical images and performance feature analysis to facilitate the decision making of radiologists.
However, the difference is that CAD usually focuses on the detection and diagnosis of lesions, while
radiomics tries to make the most of the images to provide more comprehensive information based on the
mined features. For example, researchers recently found that radiomics’ features were highly correlated
with lung cancer prognosis. By mining numerous features, they discovered that a subset of lung tumors
shared four characteristics related to intra-tumor heterogeneity, and these cancers grew and proliferated
faster than others, leading to worse patient outcomes (Aerts, Velazquez et al. 2014).
Radiomics also provides a complementary method of gene analysis. The hypothesis behind this is that gene expression
or mutation is reflected in the tumor phenotype (Gatenby, Grove et al. 2013). In the traditional gene
analysis process, tumor tissue is sampled from a certain location of the tumor. However, a gene mutation
may occur in other parts of the tumor that are not sampled. Thus, a traditional gene analysis may have
sampling errors (Gillies, Kinahan et al. 2015). Radiomics’ features are extracted from the whole tumor
image, which contains much more complete information compared to a specific area of the tumor. It is
expected that radiomics will complement traditional gene analysis.