The emerging field of cancer radiomics uses statistical analyses — increasingly, employing machine learning or artificial intelligence — to extract quantitative data about tumor volume, morphology, texture, and microvasculature invisible to the human eye from medical images to predict tumor growth and treatment response, a patient’s risk of tumor recurrence, local spread and metastasis, and prognosis. By providing more accurate information about tumor biology, the hope is that radiomics will lead to improved and earlier diagnosis, more precise predictions and prognoses, and ultimately, more individually tailored, adaptive and effective treatment plans for patients.
Radiomics seeks to convert 2- or 3-dimensional standard-of-care CT, MRI, or PET scans into mineable data for clinical decision making.1 A relatively new and investigational field, radiomics has captured the imagination of researchers and cancer clinicians because of its potential to provide far more detailed and accurate inferences and predictions about tumor biology and behavior, including responses to treatment, than can be achieved with expert visual image interpretation.1-6 Radiomic techniques can extract detailed quantitative information at a point in time — or over time, from repeated imaging examination data — about tumor volume, morphology, “texture,” microvasculature, receptor status, molecular type, anatomic spatial distributions of signal intensities, and change in these variables. Texture is a mathematical representation of the spatial variation of pixel intensity values within an image.
Imaging data can be subjected to sophisticated statistical analysis and computer modeling to identify radiomic signatures that correlate with clinical outcomes, to inform treatment plans and prognosis.1-6
Radiomic signatures might even improve the accuracy and precision of diagnostic imaging, allowing the detection of malignancies before they are detected by visual interpretation of images by expert radiologists.7 Already, radiomic signatures have shown promise in predicting treatment response and toxicity among patients undergoing radiotherapy and immunotherapy.4,8
Currently, radiomics research is largely conducted retrospectively with standard-of-care imaging studies. But in the future, imaging parameters and timing might be tailored specifically to optimize radiomic analysis.
How It’s Done
Radiomics workflow involves 5 steps that yield information for clinical decision making6: image acquisition, region of interest (ROI) segmentation, image postprocessing, feature extraction, and data analysis.
After regions of interest are identified in acquired images, they are segmented: image data are preprocessed to reduce noise or enhance contrast, ensuring the image quality and removing imaging artifacts.6 Organ and tumor boundaries and contours are then identified and delineated. Image segmentation increasingly involves varying degrees of computer automation and machine learning, with the goal of reducing human bias and error.2,6 But fully automated radiomic image segmentation is not yet reliable. Machine learning can be idiosyncratic; training these systems with one set of images does not always yield algorithms that work with other datasets.2
After segmentation, image postprocessing ensures that image pixel intensities, voxel (3-dimensional pixel) spacing, and other imaging data parameters are standardized to allow for feature extraction.2 Outliers are filtered to normalize image data.2,6
Once imaging data has been segmented and postprocessed, automated radiomic feature extraction is undertaken. Here, too, complex machine learning is playing an increasingly important role. Pixel intensity levels and other variables are used to extract meaningful information about shapes, orientations, pixel intensity distributions, and textures with which radiomic signatures or models can be identified.2,6 Computer texture analysis has been one epicenter of research because it can help quantify details of tumor biology invisible to visual image assessment by the human eye.2,6
Mathematical filters are employed to aid in the selection of extracted features for inclusion in these models; they identify and exclude features that cannot be reliably reproduced during extraction throughout repeated extractions from the same data.2 Signatures can reflect single image examination features or changes between image datasets acquired over time. Once extracted features are filtered, they are analyzed for meaningful correlations with clinical data.2,6