Radiomics in Oncology: A Practical Guide

Abstract

Radiomics refers to the extraction of mineable data from medical imaging and has been applied within oncology to improve diagnosis, prognostication, and clinical decision support, with the goal of delivering precision medicine. The authors provide a practical approach for successfully implementing a radiomic workflow from planning and conceptualization through manuscript writing. Applications in oncology typically are either classification tasks that involve computing the probability of a sample belonging to a category, such as benign versus malignant, or prediction of clinical events with a time-to-event analysis, such as overall survival. The radiomic workflow is multidisciplinary, involving radiologists and data and imaging scientists, and follows a stepwise process involving tumor segmentation, image preprocessing, feature extraction, model development, and validation. Images are curated and processed before segmentation, which can be performed on tumors, tumor subregions, or peritumoral zones. Extracted features typically describe the distribution of signal intensities and spatial relationship of pixels within a region of interest. To improve model performance and reduce overfitting, redundant and nonreproducible features are removed. Validation is essential to estimate model performance in new data and can be performed iteratively on samples of the dataset (cross-validation) or on a separate hold-out dataset by using internal or external data. A variety of noncommercial and commercial radiomic software applications can be used. Guidelines and artificial intelligence checklists are useful when planning and writing up radiomic studies. Although interest in the field continues to grow, radiologists should be familiar with potential pitfalls to ensure that meaningful conclusions can be drawn. Online supplemental material is available for this article. Published under a CC BY 4.0 license.

Created with BioRender.com

Figure 1.

Graph shows the number of publications per year since 2000 that contain the terms radiomics and texture analysis in PubMed (www.pubmed.gov). Since first being coined in 2012, the term radiomics in the literature has demonstrated an exponential increase, numbering over 1500 publications in 2020 alone. The term radiomics has overtaken texture analysis in publications in PubMed, indicating a shift toward radiomics as the preferred term in the research literature.

Figure 2.

Variations in tumor heterogeneity from less to more heterogeneous are demonstrated in these abdominal masses. (A) Axial T2-weighted MR image in a 40-year-old woman shows a large unilocular cystic lesion in the pancreas that appears to have uniform high signal intensity (SI), with only minor nonenhancing peripheral septa and a smooth border. This appearance is typical for a mucinous cystadenoma. After surgical resection, no invasive malignancy was found. (B) Axial CT image shows a partly heterogeneous mass in the left kidney, which appears well defined and contains predominantly homogeneous bland-appearing tissue with streaks of vascularity. This was found to be a spindle-cell sarcoma after surgical resection. (C, D) Axial nonenhanced (C) and contrast-enhanced (D) CT images of a fibrolamellar hepatocellular carcinoma clearly show the heterogeneous nature of this malignant tumor, with irregular vascular enhancing tissue surrounding a less-vascular central component. Contrast-enhanced imaging is often used in radiomic analyses and is useful to help highlight vascularity and spatial heterogeneity, a determinant of tumor behavior and resistance to therapy that is not readily apparent without contrast material.

Figure 3.

As demonstrated in this diagram, the study design arises by considering the interaction of multiple criteria or activities, including patient population, study endpoint, available imaging and/or clinical data, radiomic feature extraction methodology, and appropriate modeling and validation strategy.

Figure 4.

(A) Overview of a typical radiomic workflow that embodies the study design and details the steps involved in taking clinical and imaging inputs all the way through to the study endpoint. (B) Details of each stage should be clearly reported to allow meaningful interpretation, discussion, and critique of the study findings. The workflow used in Doran et al (15) is illustrated. The authors investigated the utility of radiomics from multivendor multi-parametric MRI in prediction of lymph node status in patients with breast cancer. AUC = area under the curve, DICOM = Digital Imaging and Communication in Medicine, GLCM = gray-level co-occurrence matrix, GLRLM = gray-level run-length matrix, GTDM = gray-tone difference matrix, ICC = intra-class correlation coefficient, OHIF = Open Health Imaging Foundation, PACS = picture archiving and communication system, RIS = radiology information system, ROC = receiver operator characteristic, RFE = recursive feature elimination, SVM = support vector machine, 2D = two dimensional, XNAT = eXtensible Neuroimaging Archive Toolkit.

Figure 5.

Pictorial overview of the feature classes used in most radiomic studies. Shape or morphologic features can be computed in 2D or 3D views, with 3D analysis being the recommended approach for most studies. First-order features are computed from the distribution of SIs within the ROI and include features such as the mean, median, and mode, which describe the central tendency of the data, and other features such as percentiles, skewness, kurtosis, and entropy, which describe the symmetry and heterogeneity of the distribution. Texture or second-order features consider the joint statistics of two or more voxels, so that in the coarse texture example, neighboring pairs of pixels are likely to have similar gray levels, whereas in the fine texture example, neighboring pixel values are independent. In radiologic images, the statistical dependencies between neighbors can be more complex than in these simple examples, and so features derived from the GLCM, gray-level run-length matrix (GLRLM), and other metrics can be effective for quantifying image texture. Filtering the images to emphasize edges, different length scales, or different gray levels can be used before computing texture features with the aim of sensitizing the features to a wider range of biologic correlates.

Figure 6.

SI discretization involves assigning pixels within a given SI range to a single value or bin and is used before calculation of second-order features. In this diagram, the SI histogram is derived from an ROI encompassing a hepatic tumor with varying bin size (or bin width). Increasing the bin size or decreasing the number of bins may cause loss of image detail but reduces noise, whereas decreasing the bin size or increasing the number of bins preserves image detail at the expense of image noise. The choice of image modality and SI range will define the method of discretization.

Figure 7.

Example 2D classification tasks show the impact of under- and overfitting. In the case of underfitting, the linear model fits a straight line and does not have the capacity to capture the nonlinear (curved) nature of the decision boundary, and so its classification performance on both the training and the test data will be suboptimal. In the case of overfitting, the model is insufficiently constrained and tends to generate a complex decision boundary that is overly influenced by noise. In this case, the performance in the training data will be good but will worsen when evaluated on independent test data. Many machine learning models have tuning parameters that can be adjusted to give models at both ends of this spectrum, and so optimizing the tuning parameters (typically using cross-validation techniques) is necessary to produce a well-fitted model.