Radiomics in Abdominopelvic Solid-Organ Oncologic Imaging: Current Status

Abstract

Radiomics is the process of extraction of high-throughput quantitative imaging features from medical images. These features represent noninvasive quantitative biomarkers that go beyond the traditional imaging features visible to the human eye. This article first reviews the steps of the radiomics pipeline, including image acquisition, ROI selection and image segmentation, image preprocessing, feature extraction, feature selection, and model development and application. Current evidence for the application of radiomics in abdominopelvic solid-organ cancers is then reviewed. Applications including diagnosis, subtype determination, treatment response assessment, and outcome prediction are explored within the context of hepatobiliary and pancreatic cancer, renal cell carcinoma, prostate cancer, gynecologic cancer, and adrenal masses. This literature review focuses on the strongest available evidence, including systematic reviews, meta-analyses, and large multicenter studies. Limitations of the available literature are highlighted, including marked heterogeneity in radiomics methodology, frequent use of small sample sizes with high risk of overfitting, and lack of prospective design, external validation, and standardized radiomics workflow. Thus, although studies have laid a foundation that supports continued investigation into radiomics models, stronger evidence is needed before clinical adoption.

Keywords: abdomen; features; oncology; radiomics; texture.

Fig. 1—

Schematic of steps involved in typical radiomics pipeline. Process starts with standardization of image acquisition and reconstruction parameters. Selection of lesion(s) or organ(s) of interest is then performed manually, semiautomatically, or fully automatically. Before feature extraction, image preprocessing is performed that often includes normalization and image resampling to ensure consistency of results and improve model training performance. Feature extraction involves extraction of various levels of radiomics features, after which feature selection is made via deep learning algorithms such as convolutional neural networks (CNNs). Final step involves model development and validation and then integration into clinical workflow.

Fig. 2—

ROI selection and image segmentation in 63-year-old woman with hepatocellular carcinoma. Axial contrast-enhanced CT image of abdomen shows heterogeneous mass in right hepatic lobe. Whole-liver segmentation (red outline), lesion segmentation (green outline), and segmentation by bounding box (yellow square) that includes mass and surrounding liver tissue are shown.

Fig. 3—

Effect of image normalization on distribution of pixel values within image in 63-year-old woman with hepatocellular carcinoma. A, Axial contrast-enhanced CT image of abdomen shows heterogeneous mass in right hepatic lobe. B–D, Histograms show distribution of pixel values without normalization (i.e., original pixel values) (B), with normalization to adjust pixel values to range from −1 to 1 (C), and with normalization to adjust pixel values to range from 0 to 1 (D). Normalization accelerates feature extraction and model training and ensures consistency of features across all input images.

Fig. 4—

Principal component analysis (PCA) of 36 radiomics features extracted from 603 head-and-neck squamous cell carcinomas (SCCs). PCA is used to reduce dimension to two features yet preserve clusters of features shown by green, red, and black dots. Dots indicate data obtained from head-and-neck SCCs originating from oral cavity (green), hypopharynx (red), and oropharynx (black). PCA is effective for dimension reduction.

Fig. 5—

Radiomics features correlation heatmap. Heatmap shows underlying correlation coefficient between radiomics features extracted from adrenal cortical carcinomas. This important step displays feature grouping, allowing use of only meaningful extracted features. Highly correlated features should be removed in further analysis to avoid redundancy. 1 = voxel volume, 2 = mesh volume, 3 = sphericity, 4 = elongation, 5 = maximum 2D diameter slice, 6 = surface area, 7 = maximum 2D column, 8 = gray-level cooccurrence matrix (GLCM) joint average, 9 = GLCM joint entropy, 10 = GLCM maximum probability, 11 = GLCM joint energy, 12 = GLCM difference entropy, 13 = GLCM difference variance, 14 = GLCM inverse differential moment, 15 = GLCM autocorrelation, 16 = GLCM maximum correlation coefficient, 17 = GLCM cluster prominence, 18 = GLCM informational measure of correlation 1, 19 = GLCM inverse difference, 20 = gray-level dependence matrix (GLDM) gray-level variance, 21 = GLDM dependence entropy, 22 = GLDM gray-level nonuniformity, 23 = GLDM small dependence high gray-level emphasis, 24 = GLDM emphasis, 25 = GLDM dependence variance, 26 = GLDM low gray-level emphasis.1, 27 = first-order skewness, 28 = first-order median, 29 = first-order robust mean absolute deviation, 30 = first-order total energy, 31 = first-order root mean squared, 32 = first-order minimum, 33 = first-order range, 34 = first-order 10 percentile, 35 = first-order mean, 36 = gray-level run length matrix (GLRLM) gray-level variance, 37 = GLRLM gray-level nonuniformity, 38 = GLRLM gray-level nonuniformity.1, 39 = GLRLM gray-level emphasis, 40 = GLRLM emphasis.1, 41 = GLRLM run percentage, 42 = GLRLM run entropy, 43 = GLRLM run length nonuniformity.1, 44 = gray-level size zone matrix (GLSZM) zone variance, 45 = GLSZM size zone nonuniformity, 46 = GLSZM gray-level nonuniformity.1, 47 = GLSZM small area high gray-level emphasis, 48 = GLSZM large area low gray-level emphasis, 49 = GLSZM high gray-level zone emphasis, 50 = GLSZM low gray-level zone emphasis, 51 = GLSZM small area low gray-level emphasis, 52 = neighboring gray tone difference matrix (NGTDM) complexity, 53 = NGTDM contrast.

Fig. 6—

Use of radiomics features to predict Gleason score of prostate cancer. A, ADC map in 67-year-old man with Gleason score 3 + 3 prostate cancer and PSA level of 6.2 ng/mL. Tumor (arrow) is area of low ADC in right peripheral zone. B, Corresponding parametric map of A shows high gray-level run emphasis (HGRE) texture features. C, ADC map in 60-year-old man with Gleason score 4 + 5 prostate cancer and PSA level of 9 ng/mL shows tumor (arrow) is area of low ADC in right apical transition zone. D, Corresponding parametric map of C shows HGRE texture features. E, Superimposed histograms of pixel-level HGRE texture values show difference between Gleason score 3 + 3 (blue) and Gleason score 4 + 5 (brown) tumors.