Radiomics and imaging genomics in precision medicine
Article information
Abstract
“Radiomics,” a field of study in which high-throughput data is extracted and large amounts of advanced quantitative imaging features are analyzed from medical images, and “imaging genomics,” the field of study of high-throughput methods of associating imaging features with genomic data, has gathered academic interest. However, a radiomics and imaging genomics approach in the oncology world is still in its very early stages and many problems remain to be solved. In this review, we will look through the steps of radiomics and imaging genomics in oncology, specifically addressing potential applications in each organ and focusing on technical issues.
INTRODUCTION
Medical imaging such as computed tomography (CT), positron emission tomography (PET), or magnetic resonance imaging (MRI) is mandatory in the diagnosis, staging, treatment planning, postoperative surveillance, and response evaluation in the routine management of cancer. Although these conventional modalities provide important information on cancer phenotypes, yet a great deal of genetic and prognostic information remains unrevealed.
Recently, there is universal understanding that genomic heterogeneity exists among and even within tumors and that those differences can play an important role in determining the likelihood of a clinical response to treatment with particular agents [1-4]. In other words, the success of precision medicine requires a clear understanding of each patient’s tumoral heterogeneity and individual situation.
Here, “radiomics,” a field of study in which high-throughput data is extracted and large amounts of advanced quantitative imaging features are analyzed from medical images, and “imaging genomics,” the field of study of high-throughput methods of associating imaging features with genomic data, has gathered academic interest. In other words, investigators have suggested that the hidden information embedded in medical images may become utilized through these robust approaches. Indeed, several recent studies employing radiomics and imaging genomics have been found to be useful in quantifying overall tumor spatial complexity and identifying the tumor subregions that drive disease transformation, progression, and drug resistance [5-9]. In this review, we will look through all steps of radiomics and imaging genomics in oncology, specifically addressing potential applications in each organ and focusing on technical issues.
Thorax
Lung
Two recent investigations support the importance of intratumor subregional partitioning using multiparametric images [7,10]. In one study, researchers successfully divided a tumor into necrotic regions and viable regions by incorporating 18F-fluorodeoxyglucose (18F-FDG) PET and diffusion-weighted MRI, which showed good agreement with histology [7]. In the other study, researchers identified clinically relevant, high-risk subregions in lung cancer using intratumor partitioning of 18F FDG-PET and CT images [10].
Overall, many studies have shown that textural features are associated with tumor stage, metastasis, response, survival, and metagenes in lung cancer [11-16]; thereby, providing evidence that textural features show substantial promise as prognostic indicators in thoracic oncology. Tables 1, 2 demonstrate the current literature about radiomics and imaging genomics in the field of clinical oncology [16-111].
Radiomics studies of clinical oncology published in literature
Imaging genomics studies of clinical oncology published in literature
In parallel with the 2011 The International Association for the Study of Lung Cancer (IASLC)/The American Thoracic Society (ATS)/The European Respiratory Society (ERS) classification for lung adenocarcinomas, an extensive volume of literature has covered the subset of subsolid nodules, which correlates with the spectrum of lung adenocarcinoma. Of particular importance is the significance of the presence and degree of a pathologically invasive portion, namely the thickening of alveolar septa and increased cellularity [112,113]. Although approximately half of pure ground-glass opacity (GGO) nodules have been reported to have a pathologically invasive component, discrimination between the invasive and non-invasive proportions remains challenging in pure GGO lesions because of limited visual perception and subjective analysis of conventional CT scans [114,115]. Several investigators have demonstrated that quantification and feature extraction of GGO lesions (using numerical values) can find small pathologically invasive components, which are reflected at the medical imaging voxel level and otherwise not visually detectable [116-118]. Entropy or a high attenuation value, such as the 75th percentile CT attenuation value from histograms, has been reported as a significant differentiation factor for invasive adenocarcinomas [118]. Furthermore, the 97.5th percentile CT attenuation value and the slope of CT attenuation values have been suggested as predictors for future CT attenuation changes and the growth rate of pure GGO lesions [119]. Overall, lung cancer-specific (GGO-related) radiomic features could provide additional information about tumor invasiveness and progression from other indolent or non-invasive lesions and even predict tumor growth (Fig. 1).
Various radiomic features, such as mesh-based shape, histogram, gray-level co-occurrence matrix (GLCM), intensity size zone matrix (ISZM), two-dimensional (2D) joint histogram, surface rendering for sigmoid feature, quantification of spiculation and lobulation, fractal analysis, and subregional partitioning, can be extracted from the computed tomography (CT) and positron emission tomography (PET) images of the tumor. The radiomics features are then compared with pathological and clinical data.
Breast
This part of the review will be focused on radiomics and imaging genomic researches in breast imaging using MRI texture analysis. Radiomic research has been applied to detect microcalcifications [120], differentiate benign from malignant lesions [121-123], and distinguish between breast cancer subtypes [124,125]. James et al. [120] hypothesized the magnetic susceptibility of microcalcifications leads to directional blurring effects which can be detected by statistical image processing. In their results, their method could detect localized blurring with high diagnostic performance. Regarding the differentiation between benign and malignancy, several studies have found that texture features may differ between them. In the breast two-dimensional co-occurrence matrix features of dynamic contrast-enhanced (DCE) MRI images and signal enhancement ratio maps, three-dimensional and four-dimensional features may be feasible in distinguishing between benign and malignant breast lesions [121-123]. Holli et al. [124] have investigated to differentiate invasive lobular carcinoma (ILC) and invasive ductal carcinoma (IDC) by using different texture methods. In this study, co-occurrence matrix features were significantly different between ILC and IDC, allowing differentiation between these two histological subtypes. Further, these features were superior to the other texture methods applied including histogram analysis, run-length matrix, autoregressive model, and wavelet transform [124].
Regarding texture analysis of breast MR images, this technique has been applied to predict treatment response [126]. Parikh et al. [126] evaluated whether changes in MRI texture features can predict pathologic complete response (pCR) to neoadjuvant chemotherapy. In their study conducted in 36 consecutive primary breast cancer patients, an increase in T2-weighted MRI uniformity and a decrease in T2-weighted MRI entropy after neoadjuvant chemotherapy may be helpful in earlier predicting pCR than tumor size change.
Regarding relationship between patients’ outcome in patients treated with neoadjuvant chemotherapy and texture features, Pickles et al. [127] showed that higher entropy in DCE-MR images were associated with poorer outcomes. In preoperative setting, Kim et al. [128] evaluated the relationship between MRI texture features and survival outcomes in 203 patients with primary breast cancer. They only used histogram-based uniformity and entropy in T2-weighted images and contrast-enhanced T1 subtraction images. In multivariate analysis, lower T1 entropy and higher T2 entropy were significantly associated with worse outcomes. They concluded patients with breast cancers that appeared more heterogeneous on T2-weighted images (higher entropy) and those that appeared less heterogeneous on contrast-enhanced T1-weighted subtraction images (lower entropy) showed worse outcome.
Abdomen
In the abdominal cancers as well, radiomic approaches are very promising to find imaging biomarker for predicting molecular subtyping related to patients’ prognosis, to optimize the treatment including selection of chemotherapeutic agent, and to predict the treatment response. Radiology is comprehensive for the treatment of tumor and provides anatomic and morphologic details which are available from CT and MRI. Previously, these details, so called imaging traits, were considered as a single entity, and part of them were generally poorly understood and often ignored. Recently, the recognition of the imaging traits is being highlighted because it may provide consequent information enabling prediction of tumor response to management and prognosis [129]. Especially, given the objective methods to evaluate various imaging methods such as texture analysis which measures objectively the heterogeneity of the lesions by quantifying the patterns of pixel intensities were improved [130], clinical usefulness of radiomics is being expected more and more. Texture analysis, a novel technique, measures objectively the heterogeneity of tumors by quantification of the spatial pattern of pixel intensities on cross-sectional imaging.
Also in the abdomen, some of researchers started to utilize variable imaging modalities as well as conventional CT or MRI for radiogenomic researches although most of them are pilot studies. Metabolic imaging by PET-CT and hyperpolarized 13C labeling MRI can be also applied to predict high-grade malignancy and to give an early indication of tumor response [131,132]. Recent MRI techniques including diffusion-weighted imaging and hepatobiliary phase imaging after gadoxetic acid administration has been studied the relationship with histologic and clinical phenotypes including microvascular invasion in hepatocellular carcinoma (HCC) and patients’ prognosis in intrahepatic cholangiocarcinoma (ICC) [133,134].
Nevertheless, we should overcome some important hurdles against radiomics in the abdominal field: first, it is not easy to obtain volumetric data for abdominal tumors because the tumor boundary is indistinct from the normal tissue or adjacent organs compared with the tumor in the lung. For generalized data, acquisition of volumetric data with automatic or semiautomatic manner is necessary [135]. Second, in the case of tumor arising from the hollow viscus, the boundary is more complicated. The shape of tumor on the imaging study might be different from that on the pathologic specimen. Because an intestinal tumor is growing with bowel wall, the lumen of involved bowel may be at the center of the tumor. Therefore, segmentation of adenocarcinoma in the stomach or colon is not easy.
In this part, feasible imaging biomarkers for the abdominal cancers will be addressed and the application of radiomics in the abdominal diseases will be introduced.
Hepatocellular carcinoma
HCC is the most common primary cancer of the liver and the second most common cause of cancer-related death. HCC is known as a silent killer which displays minimal symptoms in the early stage of disease and often rarely induce remission despite of the treatment at detection because of the current lack of specific biomarkers. Current staging systems, such as Barcelona Clinic Liver Cancer (BCLC) staging system, do not consider the molecular characteristics of the tumor, even the various etiology of the tumor. Reflecting the varied etiology, HCCs show extreme genetic heterogeneity. And the variability in the prognosis of individuals with HCC suggests that HCC may consist of several distinct biologic phenotypes, which result from activation of different oncogenic pathways during carcinogenesis or from a different cell of origin. In principle, any of the components of a signaling pathway may undergo mutation, although in practice more frequently susceptible genes emerge from genetic screens. Tumor protein P53 (TP53) and β-catenin are the most frequently mutated genes and are associated with a prognosis [136,137]. The other hand, the transcriptional characteristics of HCC can provide insight into the cellular origin of the tumor, and individuals with HCC who shared a gene expression pattern with fetal hepatoblasts had a poor prognosis. Activation of activator protein 1 (AP-1) transcription factors might have key roles in tumor development [138].
Intrahepatic cholangiocarcinoma
In the ICC, an aggressive primary liver cancer, epidermal growth factor receptor (EGFR), vascular endothelial growth factor (VEGF), and other angiogenic promotors are frequently over-expressed [139,140]. According to a study about molecular profiling of cholangiocarcinoma, V-Ki-ras2 Kirsten rat sarcoma viral oncogene homolog (KRAS), phosphatidylinositol 3-kinase catalytic 110-KD alpha (PIK3CA), mesenchymal-epithelial transition factor (MET), EGFR, proto-oncogene B-Raf (BRAF), and neuroblastoma rat sarcoma viral oncogene homolog (NRAS) oncogenic mutation were frequently identified in a quarter of ICC patients [141]. These molecular variabilities of ICC cause the expression of microvascular phenotypes related to aggressiveness and tumor size.
Colorectal cancer and hepatic metastasis
Compared with the liver, texture analyses in the tumor arising from the gastrointestinal tract including colorectal tumors are relatively fewer because the complexity of image data processing including objective (automatic or semi-automatic) tumor segmentation. Some studies endorsed the analysis of the largest cross section of the tumor rather than the whole tumor, but whole tumor analysis is more representative of tumor heterogeneity in colorectal cancer [142]. According to a study about assessment of primary colorectal cancer using whole-tumor texture analysis, entropy, kurtosis, standard deviation, homogeneity, and skewness might be related to 5-year overall survival of the patients [143]. Unlike from other organs, greater homogeneity at a fine-texture level were associated with a poorer prognosis, leading us to hypothesize that these might be tumors with greater cell packing and more uniform distribution of vascularization and contrast enhancement. In terms of hepatic metastasis, there are several studies focused on hepatic texture in patients with colorectal cancer. In the several studies, increased entropy might be related to the presence of metastasis [144] or poor prognosis after chemotherapy [145,146], but tumor size or volume seemed to be not a predictor of good responders. Therefore, texture analysis could be a good alternative for existing scales for evaluation of tumor response after treatment such as World Health Organization criteria and Response Evaluation Criteria In Solid Tumors (RECIST) criteria.
STEPS OF RADIOMICS
Image acquisition
The first step in the radiomics algorithm begins with image acquisition (Fig. 2). However, image acquisition parameters including radiation dose, scanning protocol, reconstruction algorithm, and slice thickness vary widely in routine clinical practice. Therefore, comparison of features extracted from different methods of image acquisition becomes more challenging. Furthermore, several radiomics features were reported to be sensitive according to variations in section thickness, pixel size, and reconstruction parameters [147,148]. On the other hand, Yan et al. [149] successfully identified several features which remained stable despite different PET image reconstruction settings. Variability issue concerning methods of image acquisition needs to be further investigated.
Radiomics is defined as the processing of radiological imaging data including sequential steps of image acquisition, region of interest (ROI) segmentation, and multiple feature extraction.
Segmentation
Accurate identification of the tumor volume is mandatory for radiomics feature extraction. In most cases, segmentation of the tumor is feasible; however, in certain cases it may be challenging due to indistinct tumor margins [150,151]. For example, in the spectrum of lung adenocarcinoma, GGO is always an issue as it may represent the tumor itself or surrounding hemorrhage and inflammation. Among the variable methods of tumor segmentation, automated or semi-automated methods have been reported to be superior to manual methods for segmenting the tumor [150,152].
Feature extraction
From the identified tumor region, multiple quantitative image features as well as traditional qualitative (semantic) features can be extracted; thus, is the main body of radiomics in oncology. Both quantitative and qualitative (semantic) features have shown some potential for precision medicine in oncology, and these features are continuously being refined and developed with evolving research [17,117,153].
Currently available quantitative radiomic features can be divided into four major classes: (1) morphological, (2) statistical, (3) regional, and (4) model-based. Morphological features are the most basic and provide information about the shape and physical characteristics of a tumor. Statistical features, which are calculated using statistical methods, can be further classified into 1st-order statistical (histogram) features and higher-order statistical (texture) features. These features describe the distribution or spatial arrangement of voxel values within the tumor. Regional features can quantify beyond the immediate neighborhood and represent intratumor clonal heterogeneity. Model-based features are extracted using mathematical approaches, such as the fractal model. Overall, each category yields various quantitative parameters that reflect specific aspects of a tumor.
Feature selection
With the emergence of precision medicine, developing radiomics features as a biomarker of oncological outcome has become an issue. In this context, a major advantage of radiomics studies is that numerous features which may carry potential as future biomarkers can be extracted from a single tumor region. However, for clinical application, these numerous radiomics features need to be reduced to a number of practical usage, in other words, a selection process for choosing the most prognostic and useful radiomics features is needed. In a large study involving a total of 440 radiomics features, according to the different feature selection method and classification method, considerable variability in predictive performance was reported [18].
IMAGING GENOMICS
Radiomics integrating genomic profiles is called imaging genomics. Imaging genomics researches have become an increasingly important research direction due to its potential to improve disease diagnosis, prognosis, and treatment choice [10,11]. As genomic profiling of tumor is generally obtained through invasive procedures such as surgery or biopsy, genomics obtained from noninvasive imaging studies routinely performed in daily practice has the merit. Imaging genomics refers to the relationship between the imaging characteristics of a disease (i.e., the imaging phenotype or radiophenotype), and its gene expression patterns, gene mutations, and other genome-related characteristics [12,13]. The primary goals of imaging genomics research are to improve our knowledge of tumor biology and to develop imaging surrogates for genetic testing [13-15].
Lung
For lung cancers, significant genomic heterogeneity components that affect the likelihood of metastasis and predict response to therapy have been established [154,155]. Furthermore, genomic analysis is now essential for appropriate therapeutic planning in this era of precision medicine for advanced lung cancers with distinct tumor subregions. Accordingly, there have been several attempts to explore tumor genomics by applying a radiomic approach. Nevertheless, imaging genomics, the link between genomics and radiomic phenotyping in lung cancer, is still poorly understood.
Preliminary data have associated radiomic features from CT and PET scans in non-small cell lung cancer with each other to predict metagenes with an acceptable accuracy of 65% to 86%, among which tumor size, edge shape, and sharpness ranked highest for prognostic significance [16]. In one study, the authors performed a detailed analysis of features from 18F-FDG PET in patients with early-stage lung cancer [156]. Multiple features of PET tracer uptake correlated with signatures associated with major oncogenomic alterations in lung cancer [156,157]. According to another recent study, the combination of radiomic features and clinical information successfully predicted oncogenic fusion genes in lung cancer [19]. In general, researchers have shown promising results in using radiomics to identify radiographic tumor phenotypes that favored specific genetic expressions [16,17,19,156,158].
Breast
The published work to date has usually focused on determination of breast cancer molecular subtypes, or correlation with recurrence scores. These early efforts appeared to have great potential and have established a strong basework for future larger-scale research endeavors which will hopefully validate the implementation of breast MRI imaging genomics into clinical practice.
The most popular topic for breast MRI imaging genomics is breast cancer molecular subtypes [20]. Gene expression profiling has made stratification of breast cancers possible into four major molecular subtypes (luminal A, luminal B, human epidermal growth factor receptor-2 [HER2], and basal like) [159,160]. These different molecular subtypes have been regarded as important because each subtype are supposed to show different patterns of disease expression, response to therapy, and prognosis [161-163]. The most common molecular subtype, luminal A typically concurs with the best prognosis [159], while luminal B subtype shows good response to radiation therapy and has intermediate survival [164], in contrast to HER2 and basal subtypes, which display good response to chemotherapy but have the worst overall survival [161]. Based on prior results, oncologists take advantage of these molecular subtypes when making decisions about systemic treatment in daily practice [165].
Usual way to determine molecular subtype is based on immunohistochemistry (IHC) patterns of estrogen receptor (ER), progesterone receptor (PR), HER2, and Ki-67 expression [165]. These IHC findings are replaced expensive genetic tests and used as surrogate marker [165-167]. Agreement between IHC surrogate markers and genetic testing ranges from 41% to 100% and IHC surrogate markers have been shown to be less robust about predicting outcomes [168]. Therefore, more accurate means of classifying molecular subtypes are needed and imaging genomics is regarded as strong candidate.
There are two published articles that have attempted to build models based on imaging features to predict molecular subtype [20,125]. Waugh et al. [125] in a study of 148 cancers and 73 test sets, used texture analysis derived from 220 imaging features to evaluate surrogate molecular subtypes. Unfortunately, the authors were only able to display a classification accuracy of 57.2% with an area under the receiver operating characteristic (AUC) curve of 0.754. Nevertheless, the authors identified that entropy features, which refer to internal pixel distribution patterns that are representative of growth patterns, were the best features to discriminate among breast cancer subtypes. They conclude that their study may have been underpowered to assess the performance of a model due to the small number of features. Grimm et al. [20] used 56 imaging features, including morphologic, texture, and dynamic features, to evaluate surrogate molecular subtypes in 275 breast cancers. At multivariate analysis, their results showed a strong association between the collective imaging features and both luminal A (P=0.0007) and luminal B (P=0.0063) breast cancers.
The first commercially available genomic biomarker was 21-gene recurrence score (Oncotype DX, Genomic Health, Redwood City, CA, USA) which guided treatment decisions [169, 170]. Oncotype DX was developed to quantify the likelihood of disease recurrence in patients with early stage invasive breast cancer who were ER-positive and lymph node-negative. Results consists of three categories: low-, intermediate-, or high-risk. Patients at low-risk are thought to derive minimal benefit from the addition of chemotherapy to standard hormonal therapy. The 21-gene recurrence score is included within the treatment guidelines from the National Cancer Care Network and the American Society of Clinical Oncology [171,172]. Several additional commercially available genomic biomarkers have also been designed to predict recurrence of therapeutic response, such as MammaPrint (Agendia, Amsterdam, the Netherlands), Mammostrat (Clarient Diagnostic Services, Aliso Viejo, CA, USA), PAM50 (Prosigna, Seattle, WA, USA), but these tests are newer and not yet widely used clinically. Recently, investigators have explored associations between 21-gene recurrence scores and breast MRI, but still there are no published studies about the newer genomic biomarkers which may provide an opportunity for future investigations [173-175]. In a study of 98 patients who underwent preoperative breast MRI and Oncotype DX recurrence score testing, Sutton et al. [175] reported similar results while investigating 44 morphologic and texture imaging features. At multivariate analysis, kurtosis on the first (P=0.0056) and third (P=0.0005) postcontrast sequences was significantly correlated with recurrence scores. Recently, Li et al. [21] investigated relationship between computer-extracted MRI phenotypes with multigene assays of MammaPrint, Oncotype DX, and PAM50 to evaluate the role of radiomics in assessing the risk of breast cancer recurrence on 84 patients. On multivariate analysis, significant associations between radiomics signatures and multigene assay recurrence scores were reported. Use of radiomics for distinguishing poor and good prognosis demonstrated AUC values of 0.88, 0.76, and 0.68 for MammaPrint, Oncotype Dx, and PAM50 risk of relapse based on subtype, respectively.
Abdomen
Imaging genomics about HCC is a very early stage, but initial result by Segal and his colleagues [176] was promising. On the basis of several different imaging traits, tumors with internal arteries and an absence of hypodense halos were related to increased specific gene expression resulting in increased risk for microvascular invasion (Fig. 3). The presence of internal arteries was also an independent factor for a poor prognosis [176]. Researchers of the previous paper maintained the imaging genomic study about prediction of microvascular invasion of HCC, and they introduced radiogenomic venous invasion (RVI) which is a contrast-enhanced CT biomarker of microvascular invasion derived from a 91-gene HCC gene expression. They revealed that the diagnostic accuracy of RVI was 89%, and positive RVI score was associated with lower overall survival than negative RVI score in the study cohorts [177]. Kitao and his colleague [178] concentrated to HCC with β-catenin mutation. The β-catenin mutation is known that it is associated with the promotion of carcinogenesis and acceleration of bile production with a relatively favorable prognosis. They evaluated gadoxetic acid-enhanced MRI, and explored some parametric variables including contrast-to-noise ratio, apparent diffusion coefficient (ADC) of diffusion-weighted imaging, and enhancement ratio of postcontrast imaging. They concluded HCC with β-catenin mutation predicted by characteristic imaging parameters including high enhancement ratio at gadoxetic acid-enhanced MRI and high ADC at diffusion-weighted imaging had significant positive correlations among phenotypes such as expression of β-catenin, glutamine synthetase, and organic anion transporting polupeptide 1B3 (OATP1B3) [178]. In terms of prognostic consequences, imaging genomics may be useful to decide therapeutic options. The gene expression related to doxorubicin resistance in HCC cells was investigated and some associated imaging traits were examined. Doxorubicin is a chemotherapeutic drug usually used with transcatheter arterial chemoembolization. Among these imaging traits, a poorly defined tumor margin was considered a significantly related factor of the doxorubicin resistance [179].
Imaging traits of hepatocellular carcinoma (HCC) and gene expression. (A) Three imaging traits in HCC: internal arteries, hypodense halo, and texture heterogeneity. (B) Strategy to make an association map between imaging traits and gene expression. Reprinted from Segal et al. [176], with permission from Nature Publishing Group.
Although the imaging genomic study about ICC is not common, an interesting study using a texture analysis of CT data in patients with ICC was recently published (Fig. 4) [180]. They focused on the relationship between the heterogeneity in tumor enhancement pattern of ICC and a molecular profile based on hypoxia markers, such as VEGF, EGFR, cluster of differentiation 24 (CD24), multidrug resistance-associated protein 1 (MRP-1), hypoxia-inducible factor 1α (HIF-1α), glucose transporter 1 (GLUT1), carbonic anhydrase IX (CA-IX), mouse double minute 2 homolog (MDM2), and P53. On the result, the combination of entropy, correlation, and homogeneity was significantly related to EGFR and CD24 expression, and it might be meaningful imaging textures quantifying visible variations in enhancement. The hypoxic microenvironment and abnormal vasculature derived by these molecules leads to tumor-related angiogenesis which affects local tumor growth and metastasis, which supports that several anti-angiogenic agents such as bevacizumab (anti-VEGF antibody) and cetuximab (anti-EGFR antibody) are used for the patients with advanced ICC [181,182]. Furthermore, CD24 is a cell adhesion molecule associated with chemoresistance capability and poor survival in ICC. Recently, CD24 is considered an emerging target for directed molecular therapy, as decreased invasiveness was observed with CD24 inhibition [183].
Representative texture features of intrahepatic cholangiocarcinoma. (A) Quantitative image phenotypes derived from texture analysis. These features are automatically computed based on the region of interest extracted from computed tomography (CT). (B) Schematic process for making the prediction model of intrahepatic cholangiocarcinoma. Reprinted from Sadot et al. [180]. VEGF, vascular endothelial growth factor; EGFR, epidermal growth factor receptor; CA-IX, carbonic anhydrase IX; HIF-1α, hypoxia-inducible factor 1α; P53, protein p53; MDM2, mouse double minute 2 homolog; CD24, cluster of differentiation 24; MRP-1, multidrug resistance-associated protein 1; GLUT1, glucose transporter 1.
PARTICULAR CONSIDERATIONS REGARDING RADIOMIC APPROACH
Reproducibility of features and study results
Although a large number of radiomics features have shown potential in tumor response and prognosis, reproducibility of radiomics features and study results remain challenging. Unfortunately, several early investigators have reported that many features were often unstable [184-186]. In a study of 219 radiomics features, only 66 features reported intraclass correlation coefficient value of more than 0.90 [184,185]. Fig. 5 depicts the ICC distributions among radiomics features according to color. Hence, validation across different institutions may serve as the solution for reproducibility of features and study results.
Intraclass correlation coefficient values are depicted for each radiomics feature belonging to seven categories. Darker colors have greater reproducibility. Note the overall high correlation of radiomics features. IQR, interquartile range; RMS, root mean square; MPP, mean value of positive pixels; UPP, uniformity of distribution of positive pixels; Max3D, maximum three-dimensional diameter; SVR, surface to volume ratio; GLCM, gray-level co-occurrence matrix; GLCM-S, gray level co-occurrence matrix subsampled; IMC, informational measure of correlation; IMC-S, informational measure of correlation subsampled; ISZM, intensity size zone matrix.
Issues of imaging modality
Special consideration is required to apply radiomics due to MR specific characteristics, intensity inhomogeneity which can significantly affect radiomic feature extraction [23,187]. Thus, before registration of MR images, the necessity of bias field correction by convolving the images with a Gaussian low-pass filter, resulting in uniform intensities across the volume should be inquired [188]. Furthermore, the stability of MRI-based radiomics features has not been investigated, and thus would be a valuable future study.
CONCLUSION
A radiomics and imaging genomics approach in the oncology world is still in its very early stages and many problems remain to be solved. However, in the close future, we believe that radiomics and imaging genomics will play a significant role of performing image genotyping and phenotyping to enhance the role of medical imaging in precision medicine.
Notes
No potential conflict of interestrelevantto this article was reported.
Acknowledgements
We are thankful to Professor Hyunjin Park from School of Electronic and Electrical Engineering and Center for Neuroscience Imaging Research, Sungkyunkwan University, Suwon, Korea, Seung-Hak Lee and Jonghoon Kim from Department of Electronic Electrical and Computer Engineering and Center for Neuroscience Imaging Research, Sungkyunkwan University, Suwon, Korea, who devoted their time and knowledge in technical support to provide graphic figures for this study.