Rizzo et al. European Radiology Experimental (2018) 2:36 European Radiology https://doi.org/10.1186/s41747-018-0068-z Experimental NARRATIVEREVIEW Open Access Radiomics: the facts and the challenges of image analysis Stefania Rizzo1* , Francesca Botta2, Sara Raimondi3, Daniela Origgi2, Cristiana Fanciullo4, Alessio Giuseppe Morganti5 and Massimo Bellomi6 Abstract Radiomics is an emerging translational field of research aiming to extract mineable high-dimensional data from clinical images. The radiomic process can be divided into distinct steps with definable inputs and outputs, such as image acquisition and reconstruction, image segmentation, features extraction and qualification, analysis, and model building. Each step needs careful evaluation for the construction of robust and reliable models to be transferred into clinical practice for the purposes of prognosis, non-invasive disease tracking, and evaluation of disease response to treatment. After the definition of texture parameters (shape features; first-, second-, and higher- order features), we briefly discuss the origin of the term radiomics and the methods for selecting the parameters useful for a radiomic approach, including cluster analysis, principal component analysis, random forest, neural network, linear/logistic regression, and other. Reproducibility and clinical value of parameters should be firstly tested with internal cross-validation and then validated on independent external cohorts. This article summarises the major issues regarding this multi-step process, focussing in particular on challenges of the extraction of radiomic features from data sets provided by computed tomography, positron emission tomography, and magnetic resonance imaging. Keywords: Clinical decision-making, Biomarkers, Image processing (computer-assisted), Radiomics, Texture analysis Key points with or without associated gene expression to support evidence-based clinical decision-making [1]. The con- Radiomics is a complex multi-step process aiding cept underlying the process is that both morphological clinical decision-making and outcome prediction and functional clinical images contain qualitative and Manual, automatic, and semi-automatic segmentation quantitative information, which may reflect the under- is challenging because of reproducibility issues lying pathophysiology of a tissue. Radiomics’ analyses Quantitative features are mathematically extracted can be performed in tumour regions, metastatic lesions, by software, with different complexity levels as well as in normal tissues [2]. Reproducibility and clinical value of radiomic The radiomics quantitative features can be calculated features should be firstly tested with internal by dedicated software, which accepts the medical images cross-validation and then validated on independent as an input. Despite many tools developed for this spe- external cohorts cific task being user-friendly in terms of use, and well performing in terms of calculation time, it is still chal- lenging to carefully check the quality of the input data Background and to select the optimal parameters to guarantee a reli- In the new era of precision medicine, radiomics is an able and robust output. emerging translational field of research aiming to find The quality of features extracted, their association with associations between qualitative and quantitative infor- clinical data, and also the model derived from them, can mation extracted from clinical images and clinical data, be affected by the type of image acquisition, postproces- * Correspondence: [email protected] sing, and segmentation. 1Department of Radiology, IEO, European Institute of Oncology, IRCCS, Milan, This article summarises the major issues regarding this IT, Italy multi-step process, focussing in particular on the challenges Full list of author information is available at the end of the article © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Rizzo et al. European Radiology Experimental (2018) 2:36 Page 2 of 8 that the extraction and radiomics’ use of imaging features Considering that many parameters can be tuned by from computed tomography (CT), positron emission tom- the user, hundreds of variables can be generated from a ography (PET), and magnetic resonance imaging (MRI) single image. generates. Most of the abovementioned features are neither ori- ginal nor innovative descriptors. Indeed, the definition Definition and extraction of image features and use of textural features to quantify image properties, Different kind of features can be derived from clinical as well as the use of filters and mathematical transforms images. Qualitative semantic features are commonly to process signals, date back a few decades [6]. There- used in the radiology lexicon to describe lesions [3]. fore, the main innovation of radiomics relies on the – Quantitative features are descriptors extracted from the omics suffix, originally created for molecular biology images by software implementing mathematical algo- disciplines. This refers to the simultaneous use of a large rithms [4]. They exhibit different levels of complexity amount of parameters extracted from a single lesion, and express properties firstly of the lesion shape and the which are mathematically processed with advanced stat- voxel intensity histogram, secondarily of the spatial ar- istical methods under the hypothesis that an appropriate rangement of the intensity values at voxel level (texture). combination of them, along with clinical data, can ex- They can be extracted either directly from the images or press significant tissue properties, useful for diagnosis, after applying different filters or transforms (e.g., wavelet prognosis, or treatment in an individual patient (person- transform). alisation). Additionally, radiomics takes advantage of the Quantitative features are usually categorised into the full use of large data-analysis experience developed by following subgroups: other -omics disciplines, as well as by big-data analytics. Shape features describe the shape of the traced region Some difficulties arise when the user has to choose of interest (ROI) and its geometric properties such as which and how many parameters to extract from the im- volume, maximum diameter along different orthogonal ages. Each tool calculates a different number of features, directions, maximum surface, tumour compactness, and belonging to different categories, and the initial choice sphericity. For example, the surface-to-volume ratio of a may appear somehow arbitrary. Nonetheless, methods spiculated tumour will show higher values than that of a for data analysis strictly depend on the number of input round tumour of similar volume. variables, possibly affecting the final result. One possible First-order statistics features describe the distribution approach is to start from all the features provided by the of individual voxel values without concern for spatial rela- calculation tool, and to perform a preliminary analysis to tionships. These are histogram-based properties reporting select the most repeatable and reproducible parame- the mean, median, maximum, minimum values of the ters; to subsequently reduce them by correlation and voxel intensities on the image, as well as their skewness redundancy analysis [9]. Another approach is to make (asymmetry), kurtosis (flatness), uniformity, and random- an aprioriselection of the features, based on their ness (entropy). mathematical definition, focussing on the parameters Second-order statistics features include the so-called easily interpretable in terms of visual appearance, or textural features [5, 6], which are obtained calculating directly connectable to some biological properties of the statistical inter-relationships between neighbouring the tissue. voxels [7]. They provide a measure of the spatial ar- Alternatively, machine-learning techniques, underlying rangement of the voxel intensities, and hence of the idea that computers may learn from past examples intra-lesion heterogeneity. Such features can be derived and detect hard-to-discern patterns from large and com- from the grey-level co-occurrence matrix (GLCM), plex data sets, are emerging as useful tools that may lead quantifying the incidence of voxels with same intensities to the selection of appropriate features [10–12]. at a predetermined distance along a fixed direction, or from the Grey-level run-length matrix (GLRLM), quanti- Analysis and model building fying consecutive voxels with the same intensity along Many of the extracted features are redundant. Therefore, fixed directions [8]. initial efforts should focus on identifying appropriate Higher-order statistics features are obtained by statistical endpoints with a potential clinical application, to select methods after applying filters or mathematical transforms information useful for a specific purpose. Radiomics’ to the images; for example, with the aim of identifying re- analysis usually includes two main steps: petitive or non-repetitive patterns, suppressing noise, or highlighting details. These include fractal analysis, Min- 1. Dimensionality reduction and feature selection, kowski functionals, wavelet