The Applications of Radiomics in Precision Diagnosis and Treatment of Oncology: Opportunities and Challenges.

Medical imaging can assess the tumor and its environment in their entirety, which makes it suitable for monitoring the temporal and spatial characteristics of the tumor. Progress in computational methods, especially in artificial intelligence for medical image process and analysis, has converted these images into quantitative and minable data associated with clinical events in oncology management. This concept was first described as radiomics in 2012. Since then, computer scientists, radiologists, and oncologists have gravitated towards this new tool and exploited advanced methodologies to mine the information behind medical images. On the basis of a great quantity of radiographic images and novel computational technologies, researchers developed and validated radiomic models that may improve the accuracy of diagnoses and therapy response assessments. Here, we review the recent methodological developments in radiomics, including data acquisition, tumor segmentation, feature extraction, and modelling, as well as the rapidly developing deep learning technology. Moreover, we outline the main applications of radiomics in diagnosis, treatment planning and evaluations in the field of oncology with the aim of developing quantitative and personalized medicine. Finally, we discuss the challenges in the field of radiomics and the scope and clinical applicability of these methods.

Introduction

Multimodality medical images contain a great deal of valuable information reflecting the development and progression of cancer. Advancements in data mining and machine learning make it possible to extract many quantitative features and convert the rapidly increasing number of medical images into minable data. This comprehensive method used to analyze medical images is known as radiomics. The concept of radiomics was first proposed in 2012 1, 2, and has since attracted the attention of researchers throughout the world (Figure 1).

Figure 1

Publication statistics of radiomics since 2012. The number of publications is going straight up. Abbreviations: CT, Computed Tomography; MRI, Magnetic Resonance Imaging; PET, Positron Emission Tomography

Initially, radiomics refers to extracting large amounts of high-dimensional quantitative features from multimodality medical images such as those of computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET) and ultrasonography (US), then the mining of correlations between these features and the diagnosis/ prognosis of cancer 3, 4. These correlations are revealed with imaging analyses, and tumors can then be decoded into different imaging phenotypes 5. With complimentary information from clinical reports, treatment responses, and genomic/proteomic assays, radiomics may reflect the global outlook of cancer 6. Compared with previous ways that process medical images as pictures for visual inspection, radiomics introduces a new way to mine the information contained in the medical images. Radiomics has been rapidly developed toward clinical application 7-9 in the hope that it will advance precision diagnostics and cancer treatment.

Radiomics may provide quantitative and objective support for decisions surrounding cancer detection and treatment 10. This is accomplished with obtaining quantitative information from medical images, combining imaging features with clinical information, genomic information and other information, and mining these data to detect radiomic biomarkers. Radiomics incorporates a series of computational technologies, and the methodologies used in radiomics are usually oriented for clinical problems. Great progress has been made within the field of radiomics regarding technology that meets clinical requirements and benefits diagnostics and cancer treatments.

Here, we will review recent developments in the methodologies in radiomics, mainly involving feature detection and model construction. We will also discuss the scope and the challenges associated with radiomics for precision diagnostics and cancer treatment. Finally, we will offer our opinion on the future steps required for better acceptance of radiomics.

Development of Methodology and Technology in Radiomics

Radiomics involves several essential steps (see pipeline of radiomics in Figure 2 and Figure 3), including data acquisition and preprocessing, tumor segmentation, feature extraction, knowledge discovery, and modeling. A series of advancements have been made in each of these areas.

Figure 2

The initially radiomics pipeline with medical images. Reproduced with permission from 4. (a) Example CT images of patients with lung cancer. (b) Strategy of radiomic analysis.

Figure 3

The radiomics pipeline of Modelling with manually defined features and Deep learning. For Modelling with manually defined features, it includes the main steps: data acquisition and preprocessing, tumor segmentation, feature extraction and selection, and modeling. For deep learning, it is an end-to-end method without separate steps of feature extraction, feature selection and modelling. Trained model from both two methods should be validated with new dataset, and then could be applied. Abbreviations: AUC, area under the receiver operating characteristic curve; C-index, concordance index; DFS, disease-free survival; PFS, progression-free survival; OS, overall survival

Data acquisition and preprocessing

Radiomics usually begins with data acquisition. It relies on a large volume of medical images and corresponding clinical data to reveal the correlation that exists between them. Radiomics was first proposed using CT images 1 and soon after was applied in the analysis of MR images 2. There were also studies based on PET 11, 12 and US 13 images utilizing similar strategy. Medical images used for radiomic analysis are collected from different hospitals or data centers; thus, these images are usually obtained using different parameters and protocols and reconstructed with different software. The differences may bring unexpected influences on the radiomic model 14. The influence of different image acquisition parameters on the reproducibility and quality of the extracted features in different modalities including CT, PET, MRI and US as well as the preprocessing of multi-modality images have been discussed in several recent literatures and a systematic review published by Larue et al. in 2017 15.

For CT images, differences in CT scanners, voxel size and reconstruction kernels are the main source of influence 16-24. Shafiq-ul-Hassan et al. 16 suggested that resampling the pixel size could minimize the dependency of radiomic features on voxel size. In addition, applying normalization on the gray level and voxel size in radiomic studies could improve the robustness of features 17. Zhao et al. 19 found that smooth and sharp reconstruction kernels should not co-occur in a study, as combining data that includes both two kernels could lead to uncontrollable discrepancies. Another recent study by Mackin et al. found that variations in tube current do not significantly affect the radiomic features 23. A most recent study comprehensively demonstrated that using a controlled protocol can reduce variability in radiomic features, while resampling image thickness cannot reduce the adverse effect derived from thickness variations 24.

There are also several studies focused on the influence of image acquisition on features extraction for PET images. In 2010, Galavis et al. has shown that most of the quantitative textural features presented large variations due to the changes in the grid size, reconstruction algorithm and number of iterations 25. Recent results from Yan et al. 26 and Veld et al. 27 further confirmed this finding. Recently, Gallivanone et al 28 found that reconstruction settings largely affects the stability of radiomic features, and only 26% were stable when different parameters were considered in this anthropomorphic phantom study. Respiratory motion is another factor that should be concerned. Oliver et al. 29 and Grootjans et al. 30 showed that respiratory motion significantly affects the quantitative textural features reflecting tumor heterogeneity with PET. In addition, they found that standard uptake value (SUV) discretization could also impact the results of textural features detecting. Compare to three-dimensional (3D)-PET, respiratory-gated (4D) PET can reduce motion blurring and generate more robust features 29, 31, 32.

The effects of MRI field strength, imaging protocols, and manufacturers on radiomic features have been investigated by several studies using either living samples or physical phantoms 33-38. For instance, Jirak et al 34 and Mayerhoefer et al 35 suggested that some features were sensitive to MR acquisition parameters including TR, TE and bandwidth, which was confirmed by a recent simulation study 39. Savio et al. 36 investigated the effect of slice thickness on texture-based model in differentiating plaque from normal tissue, and showed that the moderate differences caused by thickness does not affect the classification results. The research conducted by Waugh et al. 40 showed that differences of acquisition parameters may not significantly affect the ability of texture analysis in classification task. Moreover, some mature algorithms and tools could help to remove unwanted low-frequency intensity nonuniformity 41, 42 and perform image intensity normalization 43-45 in order to reduce the difference caused by different acquisition parameters of brain MRI.

US-derived quantitative features have been proved to be strongly correlated with breast biologic characteristics 46, gestational age 47, neonatal respiratory morbidity 48. They can also be used to detecting malignant tumors from benign tumors in thyroid 49, 50 and breast 51, 52. In a review in terms of ultrasound image processing methods for carotid plaque morphology analysis 53, Kyriacou et al. demonstrated that a standard acquisition protocol, image normalizing method helps to generate reproducible measurements. So far, there are still few studies concerning on the reproducibility and stability of quantitative features obtained from US in oncology.

Considering the influence of different image acquisition parameters, researchers should pay more attention on the differences of imaging protocols and provide essential parameters to achieve reproducebility and comparability with other radiomic studies 10. As the importance of using standard imaging protocols should be taken into account 6, researchers should develop specific methods aiming at one disease compared with others.

Tumor segmentation

In radiomic analysis, it is the entire primary tumor that is usually analyzed. Tumor segmentation determines which region will be analyzed further, so this becomes a fundamental step in radiomics. Segmentation includes manual, semiautomatic, and automatic segmentation methods. However, regardless of the method used, there will be certain challenges. First, there is no golden standard for tumor segmentation 2, and the manual segmentation is quite boring and time consuming. Radiologists have to go through every slice of the imaging sequence to delineate the regions of interest, and a mask for a large tumor may contain dozens of slices. Second, there exist numerous morphological variations as tumors are very different from geometric objects, and the variations are usually difficult to model. Third, tumor margins could be blurred by the partial volume effect and not well defined in medical images. Furthermore, segmentation methods especially their reproducibility and reliability are very important. For instance, inter-reader variability of the interpretation for in-plane nodule boundaries could be very large 54. Several researches have provided that features can be changed with different delineation. For example, van Velden et al. found 25 features were sensitive to the change in delineation using two baseline whole-body PET/CT scans from eleven NSCLC patients 27. Also Leijenaar et al. found that 71% of radiomic features were stable in a test-retest cohort of 11 NSCLC patients using a semiautomatic segmentation, and 91% of radiomic features were stable in an inter-observer cohort of 23 NSCLC patients with manual segmentation 12. Since the features are extracted based on the segmented tumors, segmentation methods could ensure that the radiomic features are reproducible and reliable.

Intra-class correlation coefficients (ICC) are usually used to evaluate the inter- and intra-reader agreement. Several studies suggested that ROIs should be detected with an acceptable ICC, so that the further extracted features can be used 7, 8. Several automatic or semi-automatic segmentation methods have been developed to minimize manual cost and increase reproducibility of tumor segmentation. Parmar et al. 55 compared radiomic features computed by a semiautomatic region growing volumetric segmentation algorithm and manual segmentation with NSCLC tumors in twenty patients; they found that the radiomic features generated with the semiautomatic segmentation had significantly higher reproducibility and more robustness compared with the manual segmentation. Heye et al. 56 compared the inter- and intra-observer variability with manual segmentation versus that with semiautomatic segmentation (“random walker”), preformed with MR Oncotreat software using 15 DCE MR studies. They concluded that the semiautomatic segmentation can provide a significant reduction in inter-observer variability compared with the manual method. These studies demonstrated the potential of semiautomatic segmentation in consistency and repeatability. In addition, Christ et al. 57 proposed a method for medical image segmentation integrating marker controlled watershed segmentation and clustering algorithm, which achieved better segmentation than the conservative watershed algorithm in MR images was achieved. Hatt et al. 58 summarized some PET automatic segmentation algorithms and concluded that the segmentation with manual and visual verification show higher reproducibility and reliability than simple threshold-based approaches.

Until now, in morphological imaging (CT, US or MRI), most radiomic studies have used manually segmented ROIs for analysis even though it is quite time consuming and entails large inter-reader variability. In PET or SPECT imaging, several automatic or semiautomatic segmentation algorithms were performed. Although segmentation methods for medical images have been a hot spot of concern for a long time, there is still a long way to detect a fully automated application 59-61.

Feature extraction

Radiomics was initially defined as the extraction of high-throughput features from images. Quantitative imaging features are an important characteristic in radiomics because they bridge medical images and the clinical end point. Typically, there are two main categories in imaging features: manually defined features (including semantics and non-semantics) and deep learning features (Figure 4).

Figure 4

The main features we used for radiomic analysis could be divided into three parts: Empirical features, Statistical features and Deep learning features. All these features could be visualized and interpreted with physical meanings. However, what we should do further is to unravel their physiological significance.

Manually defined Features

Semantic features could be defined as the qualitatively described empirical features proposed by radiologists. These features cannot be described with efficient mathematical expression, but they are useful in clinical settings and in radiomic analysis. For example, Liu et al. suggested that semantic features of CT images were correlated with epidermal-growth-factor-receptor (EGFR) status in patients with lung cancer 62.

Non-semantic features could be defined as quantitatively described imaging features with mathematical expressions. For the current radiomic studies, the mostly used non-semantic features are usually obtained from predefined ROIs. First group contains the shape features describing the tumors, including volume, surface area, compactness, and etc. Second group contains a large number of first-, second-, and high-order features constructed from the intensity of each voxel in the ROIs 6. First-order features describe the distribution of the intensity within the ROI, including maximum, median, minimum, and entropy. Second-order features describe the statistical relationships between voxels, such as textural features related with the representative of tumor heterogeneity. High-order features usually referred to statistical features calculated on matrices that consider relationships between three or more pixels. In addition, wavelet features, referred to as multiresolution image scaling features, and fractals, referred to as model-based features, were also used a lot in radiomic studies 63-65. Other computational imaging features such as local binary patterns (LBP) 66 and scale-invariant feature transform (SIFT) 67 can also be used in radiomic analysis.

Deep Learning Features

Deep learning represents a class of algorithms that use the stacked neural network structure 68. A deep learning model generally includes the following computational components: convolution, pooling, activation, full connection and batch normalization. All the computational components are defined as layers, and they are stacked layer-by-layer. The most important components of the deep-learning model are the convolutional layer and fully connected layer that can adaptively learn features from data. Therefore, the outputs of the convolutional layer or the fully connected layer can be defined as deep-learning features.

The typical manner of extracting deep-learning features is to train a deep learning model using the clinical outcomes, and then extract the outputs of several layers as deep learning features 69. For example, after training a convolutional neural network (CNN) for EGFR mutation prediction in lung cancer, the results of the final fully connected layer was extracted to obtain deep learning features for EGFR status prediction 70. Another method of extracting deep-learning features is to use an unsupervised auto-encoder. When there is not enough data to train a supervised deep learning model, an auto-encoder provides an alternative method to extract deep learning features 71-73. A typical auto-encoder includes an encoder network and a decoder network. The result of the final convolutional layer is defined as a deep learning feature.

When compared with manually defined features, deep learning features are more specific to clinical outcomes and data. Most manually defined features describe general information about a tumor (e.g., shape, intensity, and texture), but may lack specificity regarding clinical outcomes. For example, when designing manually defined features, clinical outcomes are not considered in most situations, therefore, the manually defined features may not be suitable for clinical labels. In contrast, deep learning features are learned from data directly, and therefore, the deep learning features can be adapted to specific datasets and related to clinical outcomes more naturally. Generally, deep learning features are extracted from convolutional layers in the deep learning model. Each convolutional layer includes hundreds of convolutional filters that are defined as deep learning features. The deep learning features can describe multi-level tumor information from low-level visual characteristics to high-level abstract features. For example, in recurrence analysis of high-grade ovarian cancer, a convolutional auto encoder was designed to extract deep learning features from CT image. As illustrated in Fig. 3, features in the first convolutional layer describe simple tumor intensity information, and features in the second layer describe tumor edge information. Features in a deeper layer (the 3rd layer) describe complex tumor shape information. As the convolutional layer goes deeper, high-level abstract features are extracted, which are not interpretable visually, but tend to be associated with the recurrence status of the tumor. Without the need to explicitly specify feature formulas, deep learning features naturally include much of the same information as in manually defined features such as the intensity information (the 1st layer), edge information (the 2nd feature) and shape information (the 3rd). Since deep learning features are learned from data directly, they have a relationship with the recurrence status of the tumor (Fig. 3b). In addition, deep learning features can be designed flexibly since different CNN structures provide different deep learning features, and there is no need carefully consider design feature formulas.

Knowledge discovery and modeling

Generally, thousands of quantitative imaging features could be obtained for the tumors. However, too many features may include redundant information and cause overfitting. Thus, feature selection should be performed to preserve the relevant features. A prediction model can then be constructed with the selected key features according to the clinical endpoint.

Feature Selection

The most commonly used feature selection methods in radiomic studies can be divided into three broad categories: filter, wrapper, and embedding 74.

Filter methods usually evaluate the features without involving the model; therefore, these are also called independent methods. There two main kinds of filter methods: univariate and multivariate methods. Univariate filters mostly rank features based on quality (i.e., on the Chi-squared test or Mann-Whitney U test). Multivariate filters are composed of rankers and subset selectors like those used in correlation-based feature selection. With wrapper methods, features are generated first and then evaluated using the model. This means all wrapper methods are considered multivariate and subset-selector methods. With embedded methods, a feature subset is proposed and evaluated during the construction of the model.

Most radiomic studies use filters for coarse selection. For example, several studies used ICCs to detect stable features 4, 73, 75, and others also ranked the features using Student's t-test 8 or the Mann-Whitney U test 76. Another commonly used feature-selection method in radiomic studies is the least absolute shrinkage and selection operator (LASSO), which is an embedded method that can produce selected features and a prediction model together 77-81.

Modelling

The goal of radiomics is to construct a prediction model for clinical outcomes with selected features. Machine learning provides several modelling methods to achieve this goal. Typically, supervised, semi-supervised, and unsupervised learning are fundamental strategies used according to the different levels of available clinical outcomes.

Supervised learning is a learning structure that requires clinical labels to train the model. A supervised learning model includes two steps: training and testing. In the process of model training, training samples with corresponding clinical labels are paired to train the model. Through a pre-defined loss function, the model learns the relationship between the feature and the clinical labels. In the testing phase, the well-trained model is used directly to test its predictive performance. There are many supervised learning methods that have shown great performance in radiomic analysis, such as the support vector machine 8, LASSO-logistic regression 7 and random forest 73. Despite the good performance of many supervised-learning methods, they usually require large amounts of training samples to avoid overfitting.

In many situations, clinical labels may not be enough to train a supervised model, and unsupervised learning can be an alternative. Unsupervised learning methods group data into several clusters according to the similarity between samples. In this process, the clinical label is not necessary for model training. The unsupervised model utilizes a distance measurement to calculate the similarity between samples. The similar training samples are grouped together, while the dissimilar samples are assigned into different groups. These models include many clustering algorithms such as k-means clustering 82, fuzzy clustering 83 and consensus clustering 84. For example, a consensus clustering model was used in a previous study to detect glioblastoma phenotypic subtypes 85, and patients were classified into three groups, each with different molecular pathway activities and a different prognosis.

Generally, supervised learning shows stronger performance than unsupervised learning since the clinical label is considered during the supervised model training. However, insufficiently labelled data limits the performance of supervised learning. As a trade-off between better performance and enough training data, semi-supervised learning could be a good choice. Semi-supervised learning uses a large amount of unlabeled data to mine tumor information and utilizes small amounts of labelled data to build the relationship between features and clinical labels. For predicting the recurrence for high-grade serous ovarian cancer, a semi-supervised deep learning framework was proposed 86. The semi-supervised framework can be divided into an unsupervised feature learning phase and a supervised model training phase. In the unsupervised learning phase, a convolutional auto encoder was built to compress tumor images into low-dimensional features. In this process, the auto encoder tried to extract the intrinsic characteristics of the tumor without a clinical label. Afterwards, a supervised Cox model was built to model the relationships between the features and clinical labels.

Validation

A radiomic model must be validated to show its potential value for clinical application. An independent, external, validated model is considered to be more credible than an internally validated model, as results from independently obtained data are usually more robust. Furthermore, prospectively validated models have the most credibility for radiomic studies. There are dozens of tools to measure the performance of radiomic models. For discrimination analysis, the receiver operating characteristic (ROC) curve is the mostly used method to report the performance of proposed models. Moreover, the area under the curve (AUC), sensitivity, and specificity of the model can also be used to evaluate if the model can predictive the clinical outcome. For survival analysis, the concordance index (C-index) and time-dependent ROC curve are usually used for validation. In addition, calibration is a useful tool for both discrimination analysis and survival analysis, as it can obtain the agreement between the observed clinical outcomes and model predictions 87.

Applications of Radiomics in Diagnosis and Treatment of Oncology

Since radiomics was proposed, rapidly increasing radiomic studies have been published to improve the diagnosis and treatment of cancer. More and more studies demonstrate the value of radiomic features as a complimentary tool to decision making of clinicians. Here, we will provide a brief introduction to radiomic studies, focusing on the clinical problems during the whole process of diagnosis and treatment in several typical cancers including brain tumors, head-and-neck cancer, breast cancer, lung cancer, prostate cancer, colorectal cancer, liver cancer and gastric cancer (Figure 5).

Figure 5

Scope of radiomics for diagnosis and treatment evaluation, suggesting the potential directions radiomics could be applied for. Abbreviations: EGFR: epidermal growth factor receptor; IDH: isocitrate dehydrogenase; KRAS: Kirsten rat sarcoma viral oncogene homolog; MGMT: O-6-methylguanine-DNA methyltransferase

Brain Tumors

Malignant brain tumors can be divided into primary tumors which started within the brain (mainly referred to as gliomas, meningiomas, and chordomas), and secondary tumors that spread from elsewhere, known as brain-metastasis tumors 88. Current brain-tumor imaging protocols incorporate CT, multi-parametric MRI, and sometimes PET. As use of radiomics could extract large amounts of quantitative imaging features and capture intratumoral and intertumoral heterogeneity, it makes a radiomic analysis could assess imaging phenotypes that may influence the diagnosis and treatment evaluation of brain tumors.

Diagnosis

Although standard brain imaging can provide results enough for tumor grading, improve was needed for newly proposed imaging protocols. Bai et al. first utilized diffusion kurtosis MRI for the grading of gliomas and detected good performance in the grading of gliomas 89. Moreover, radiomics also attracted attention for the prediction of molecular subtypes of brain tumors, as the precision diagnosis of gene-expression patterns could potentially enhance decision making for targeted therapies 90. Isocitrate dehydrogenase (IDH) 91, O6-methylguanine-DNA methyltransferase (MGMT) 92, 93, 1p/19q co-deletion 94, EGFR expression level 95, Ki-67 expression level 96, p53 status 97, and ATRX mutation 98 have been the main focus in prediction studies of molecular subtype in brain tumors. Radiomic analysis based on multi-parametric MRI aided the prediction of molecular characteristics. More importantly, imaging phenotypes were shown to be associated with molecular pathway activities that may determine the type of targeted therapy 85.

Treatment evaluation and Prognosis

There has been considerable interest in treatment evaluation and prognosis for brain tumors to identify imaging phenotypes that may predict the treatment response in patients with glioblastoma. Two recent studies investigating the responses to bevacizumab treatment in recurrent glioblastoma patients suggested the potential of radiomics to predict different response to the treatment 99, 100. These studies both detected the potential of radiomics to aid in cancer treatment decision-making at a low cost. However, a bevacizumab-naïve control group was needed for these studies to confirm the predictive value. In addition, radiomics provided a new option for determining prognosis for brain tumors. Recent studies have suggested that features detected from MRI and PET were significantly associated with the survival of patients with gliomas 101, 102. In the future, a model could be constructed using radiomics to improve both treatment planning and prognosis.

Head- and-neck cancer

Head-and-neck cancer is one of the cancers that radiomics has also been widely applied 103.

Ren et al. built a MRI based radiomic signature to predict the stage of head and neck cancer preoperatively 104. They found that the radiomic signature from contrast-enhanced T1-weighted MR images and T2-weighted MR images had good performance in discriminating different stages. Leijenaar et al found that CT radiomic signature could predict HPV (p16) status in oropharyngeal squamous cell carcinoma 105. Zhou et al proposed a 3-dimensional deep learning model to predict lymph node metastasis in nasopharyngeal cancer (NPC) 106. Chen et al. evaluated the association between Tumor PD-1 expression and Immunohistochemical biomarkers or radiomic features from PET imaging in NPC 107. Crispin-Ortuzar et al predicted hypoxia status using PET/CT radiomics in NPC patients 108.

Zhang et al. proposed that multi-parametric MRI based radiomics could be a novel prognostic factor in advanced NPC 109. They collected 118 advanced NPC patients and found that radiomic signature achieved significantly improved performance evaluating PFS compared with the TNM staging system. Furthermore, Zhang et al. evaluated 6 different feature selection methods and 9 different classifiers for prediction of local failure and distant failure in advanced NPC 110. They found that Random Forest performed best among the 9 methods. Wang et al. 111 investigated the value of radiomic signatures in prediction of early response to induction chemotherapy in NPC patients. They found that radiomic signature had a good performance in predicting early response to induction chemotherapy. Lu and Lv et al. evaluated the robustness of radiomic features obtained from different PET images in NPC patients 112, 113. Wu et al. performed survival prediction in high-grade Osteosarcoma using radiomics of diagnostic CT 114. They found that radiomic nomogram, which combines a radiomic signature and clinical factors, showed better calibration and classification capacity than a model with only clinical factors. Elhalawani et al investigated radiomic signatures for local recurrence in oropharyngeal cancer 115. A radiomic signature consisting of 2 radiomic features showed robust discrimination ability of recurrence. Vallières et al used radiomics for assessing tumor failure in patients with NPC from 4 centers 116. Gabryś et al predicted the xerostomia in NPC patients after radiotherapy 117. They compared the performance of 6 feature selection, 7 classification, and 10 data balancing methods.

Breast cancer

Breast cancer is the cancer with the highest incidence in women worldwide. As it is a known heterogeneous disease, the precise diagnosis and early prediction of response to the treatment are hotspots of the clinical practice and research in breast cancer. Radiomics, combining multi-modality imaging data and clinical information, is now used widely in breast cancer research.

There are several diagnostic imaging modalities for breast cancer, including US, PET/CT, mammography, and MRI. Thus, radiomic researches in breast cancer also covered all these modalities. Several previous studies utilized radiomics for the prediction of breast cancer subtype or the status of ER, PR, Ki67 and HER2 with US 46, mammography 118, PET/CT 119, 120, and MRI 121, 122. Specifically, Antropova and his colleague proposed a feature fusion method that could combine US, MRI and mammography together for better diagnosis of breast cancer 118. In addition of the characterization of breast cancer, radiomics could also provide a non-invasive approach for prediction of sentinel lymph node metastasis 123.

Most radiomic studies in breast cancer have focused on the evaluation of response to therapy. Chan et al. 124 developed an automated method to predict treatment failure in early breast cancer patients with a pretreatment MRI scan. Most other studies attempted to obtain a radiomic biomarker for pathological complete response (pCR) to neoadjuvant chemotherapy, which is a hot point in breast cancer research. Braman et al found that intra- and peri-tumoral features of DCE-MRI could contribute to pretreatment prediction of pCR 125. Other studies also showed that T1WI, T2WI and DWI could also help to detect pCR 76, 126. While several studies approved the potential of MRI as a biomarker for detecting pCR, Tran et al pointed out that diffuse optical spectroscopic could also be a useful tool for pCR prediction 127.

There are also radiomic studies focusing on the prognosis of breast cancer. Park et al 128 have developed a radiomic signature combining MRI features and clinical information for individualized estimation of disease free survival in breast cancer patients.

Lung cancer

Lung cancer is the most harmful cancer, and its prevalence continues to increase worldwide 129. Radiomics has been widely used in the diagnosis, treatment evaluation, and prognosis in lung cancer 130, 131.

One of the most important diagnostic applications of radiomics is screening of lung cancer. Kumar et al performed a classification between malignant and benign lesions with diagnostic data from the LIDC-IDRI dataset 132. The proposed method achieved a sensitivity of 79.06% and a specificity of 76.11%. Shen et al. proposed a deep learning model based on CT images and achieved better prediction results for malignant lung nodule compared with previous methods 133. Maldonado et al. developed a noninvasive radiomic method to discriminate solid and ground glass opacities using high resolution CT 134. Besides diagnostic CT, researches on low-dose CT were also performed for screening malignant lung nodules. Carter et al. 135 performed a screening study on patients with low-dose CT screening detected lung cancer in National Lung Cancer Screening Trial (NLST) dataset. Accuracies of 80% and 79% were found for predicting nodules that will develop cancerous in one or two years, respectively. Liu et al. also utilized a low-dose CT cohort from the United States and radiological image traits of lung nodule to predict malignancy and achieved impressing results 136.

Accurate preoperative TNM staging is also important for treatment decisions of lung cancer. Aerts et al. extracted 440 CT radiomic features from 1,019 patients and found that many of the radiomic features were correlated with the overall stage, T stage, N stage, and M stage of lung cancer and head-and-neck cancer 4, 137. Coroller et al. used CT radiomic signature to predict distant metastasis (M staging) on a United States cohort of 182 pathologically confirmed lung adenocarcinoma 138 and Zhou et al. combined CT radiomic features and clinical risk factors to identify the distant metastasis based on a Chinese cohort including 348 lung cancer patients 139. The results from all of these studies showed good performance of radiomics on preoperative M staging. Wu et al. used PET/CT radiomic features to predict distant metastasis with 101 early-stage NSCLC patients and they found additional value of PET in M staging 140.

Besides screening and staging, radiomics has been applied to predict the gene mutation or pathological type in lung cancer. Liu et al. and Zhang et al. all found that EGFR mutation was associated with CT radiomic features 141-143. Rios et al. developed CT radiomic signatures to distinguish somatic mutations on a dataset of 763 lung adenocarcinoma patients 73. They found that a CT radiomic signature that is related with radiographic heterogeneity can successfully predict EGFR status. Wu et al. used radiomics to predict the histologic type of lung cancer, adenocarcinoma or squamous carcinoma, on two NSCLC cohorts from Netherlands 144. Zhu et al. performed a similar study on a Chinese cohort and also found that radiomic signature could serve as a diagnostic factor for histologic subtype classification of NSCLC 145. Fan et al. 146 developed a radiomic signature for the preoperative discrimination of lung invasive adenocarcinoma manifesting as a ground-glass nodule.

Treatment response evaluation is essential in lung cancer due to its value in treatment decision. Mattonen et al. found that radiomic signatures could predict the recurrence after Stereotactic Ablative Radiation Therapy (SART) in Lung Cancer 147, 148. Fave et al. used delta-radiomic features to predict outcomes in Stage III NSCLC patients during the radiation therapy 149. Their results suggest the change of radiomic features due to radiation therapy would be indicators of tumor response. Coroller et al. found that pretreatment CT radiomic features could predict pathological response after neoadjuvant chemoradiation in patients with advanced NSCLC 150. Aerts et al. found that CT radiomic feature before treatment is able to predict EGFR mutation status in NSCLC and associated with gefitinib response 151. Cook et al. used heterogeneity features in PET images to evaluate the treatment response of Erlotinib in NSCLC 152. Song et al. found that CT radiomic signature could predict progression-free survival after TKI therapy 153.

Besides treatment response, many radiomic studies have focused on prognosis. Aerts et al. found the clusters of radiomic features were associated with the prognosis of lung cancer 4. Song et al. also demonstrated a relationship between CT radiomic features and overall survival in patients with NSCLC 154. Balagurunathan et al. tested the reproducibility and prognosis of quantitative radiomic features from CT images, and many features were associated with the prognosis of lung cancer 75. Coroller et al. found that radiomic features are prognostic for both distant metastasis and survival 138. Huang et al. found that the radiomic signature was associated with DFS 155.

Colorectal cancer

Colorectal cancer is a common and fatal type of cancer. The incidence and deaths of colorectal cancer are increasing every year, and many questions remain regarding the three different stages of diagnosis, treatment evaluation, and prognosis.

Accurate identification of the extent of lymph-node metastasis in colorectal cancer patients is critical for the determination of treatment strategies. If it is confirmed that the patient has no lymph-node metastasis before surgery, there is no need to perform lymph-node dissection during surgery 156-158. Huang et al. performed radiomic analysis on 526 patients with colorectal cancer 7. Based on preoperative CT images of the portal vein, the radiomic features were extracted, and a nomogram was established. The C-index was 0.736 on the training set and 0.778 on the validation set. Moreover, the authors proposed an easy-to-use nomogram based on the radiomic model to help its application (Figure 6). Another relevant application of radiomics in colorectal cancer is the prediction of KRAS/NRAS/ BRAF mutation. The National Comprehensive Cancer Network (NCCN) guidelines for colorectal cancer recommend that patients should screen for KRAS/ NRAS/BRAF mutations, because these mutations are indicative of poor response to cetuximab and panitumumab 159-162. Therefore, determination of the KRAS/NRAS/BRAF mutation status before or during treatment is essential to predict treatment efficacy and achieve a personalized diagnosis. Yang et al.163 extracted 346 radiomic features from the preoperative CT images of 117 patients with colorectal cancer for a radiomic analysis, achieving an AUC of 0.869.

Figure 6

Developed radiomics nomogram for prediction of pCR to NCRT in rectal cancer. Reproduced with permission from 8.

Patients with rectal cancer usually undergo neoadjuvant chemoradiotherapy before surgery 164, 165, but it is not known whether pCR is achieved before neoadjuvant chemoradiotherapy. This information is only obtained after surgery. Approximately 15-27% of locally advanced patients will achieve pCR 166, 167, and studies have shown that these patients do not need surgery 168-170. Based on information from 48 patients, Ke et al. 171 constructed an artificial neural network based model, and the AUC was 0.71-0.79. Liu et al. applied radiomics to identify patients with locally advanced rectal cancer who achieved PCR before surgery 8. The study enrolled 222 patients with locally advanced rectal cancer. T2WI and DWI were used to extract 2252 radiomic features. The AUC was 0.976 (95% confidence interval [CI], 0.9185-0.9711), and the predicted PCR accuracy was 94.3% (95% CI, 91.9%-97.1%). Based on 114 patients (18 PCR) Natally et al. extracted radiomic features on T2WI and made predictions using random forest classification; the AUC was 0.93, sensitivity 100%, and specificity 91% 172.

Although treatment may improve survival, it has not proven to improve the DFS of patients with locally advanced rectal cancer 13, 173, 174. Distant metastasis is the main cause of patient treatment failure. For high-risk patients, increased systemic therapy reduces the risk of metastasis and improves survival. Therefore, pre-operative risk stratification may help to select individualized treatment strategies and improve the prognosis of patients. Based on 108 patients with locally advanced rectal cancer who underwent total mesorectal excision 175, Meng et al. extracted 485 radiomic features; the C- index of the radiomic signature was 0.767 (95% CI, 0.72-0.86), and the 3-year time-dependent AUC was 0.827. Combined with clinical features, the C-index and AUC of the Cox model were improved: 0.788 (95% CI, 0.72-0.86) and 0.837, respectively.

Identification of patients who need more aggressive treatment and follow-up is critical. Lovinfosse et al. explored the significance of PET/CT-based radiomic features for prognosis. The study extracted the SUVmax, SUVmean, metabolic tumor volume, total lesion glycolysis, and texture features, established univariate and multivariate Cox models, and analyzed the prognostic performance of the features 176.

Prostate cancer

Prostate cancer (PCa) is the most common cancer affecting men in Western societies 177. Multi-parametric MRI (mp-MRI) plays important role in diagnosis and treatment of PCa because of the advantage of high soft tissue contrast.

The non-PCa condition such as prostatitis and benign prostatic hyperplasia is easily confused with PCa. Ginsburg SB et al utilized mp-MRI based radiomic model for differentiation between PCa and non-PCa 178. The result showed that the validation model achieved relatively good performance with high AUC using T2 WI, ADC, and T2 WI and ADC features for PCa detection. PCa exhibits an extraordinary variable biological behavior. According to NCCN Prostate Cancer, version 3, the identification of aggressive PCa is vital for clinic decision. Some radiomics models based on mp-MRI were developed and validated to identify PCa aggressiveness in patients 179, 180. The results showed that radiomic models could be helpful for identifying the presence and absence of PCa aggressiveness compared with PIRADS v2.0 assessment, an mp-MRI based subjective assessment method. The pathological Gleason score (GS) of PCa has clinical implications for the treatment and prognosis. Chaddad et al. achieved the best performance for GS = 3 + 4, and 64.76% for GS ≥ 4 + 3 to predict the GS of PCa patients via radiomic features based on mp-MRI181.

In addition, radiomics is also used for treatment evaluation of PCa. Many researchers successfully utilized radiomic features derived from pretreatment mp-MRI to predict biochemical recurrence of PCa after radical prostatectomy, radical radiotherapy or external beam radiotherapy 182, 183.

Liver and Gastric cancer

As medical imaging is playing a more and more important role in the management of liver cancer and gastric cancer, there are more opportunities to utilize quantitative imaging analysis in the clinical setting for diagnosis and treatment evaluation.

Radiomics is mainly used in cases of liver cancer for noninvasive prediction of microvascular invasion (MiVI). Two research teams have verified the validity of radiomics for preoperative MiVI prediction 184, 185. Bakr et al. extracted quantitative radiomic and semantic imaging features from the triphasic CT scans of 28 treatment-naïve hepatocellular carcinoma (HCC) patients 184. Delta radiomic features were designed as the absolute difference and ratio computed from all pairs of the three imaging phases. The proposed imaging features from arterial and venous phases were regarded as potential biomarkers for MiVI prediction in HCC. Jie et al. further used a cohort of 304 patients with hepatitis B virus-related (HBV) HCCs to validate the usefulness of imaging biomarkers for MiVI prediction 185. The radiomics-based nomogram showed satisfactory predictive accuracy for MiVI in HBV-related HCC.

For gastric cancer (GC), Ma et al. evaluated the value of CT-based radiomic signature to discriminate Borrmann type IV GC from primary gastric lymphoma (PGL) 186. Ba-Ssalamah et al. used CT texture features to discriminate GC, lymphoma, and gastrointestinal stromal tumors 187. Liu et al. explored the application of radiomics in predicting pathologic differentiation for the type of GC 188. Liu et al. used radiomic analysis to predict T stage, N stage, overall stage, and perineural invasion in patients with GC 189-191.

Radiomics-based methods for treatment evaluation and prognostic prediction in liver cancer rely on two aspects: overall survival (OS) and recurrence predictions. Zhou et al. proved the prognostic power of radiomic signatures for early recurrence of HCC with partial liver resection 192. The LASSO was used to build the radiomic signature. The AUC of the combined model was 0.836, which shows it performed better than clinical predictors. Akai et al. investigated the use of random forest along with radiomic analysis to predict DFS and OS for HCCs with resection as treatment 193. Furthermore, Cozzi et al. proved the radiomic signature could be used as a robust biomarker for local control and survival prediction 194. Compacity and Barcelona Clinic Liver Cancer (BCLC) stage were finally chosen as effective covariates for prognosis prediction. All these studies indicated that radiomic signatures are potential biomarkers for prognostic prediction in liver cancer research.

Radiomics is also a useful tool for treatment evaluation and prognosis in GC. Giganti et al. found that radiomic texture features had predictive ability for treatment response for neoadjuvant therapy in patients with GC 195. Hyun et al. found that textural features on CT images were associated with better prognosis in patients with HER2-positive gastric cancer who received target therapy 196. Giganti et al. investigated the association between CT based radiomic features and overall survival in patients with GC 197. Texture features were significantly associated with the prognosis of GC.

Challenges and Prospects for Radiomics

In recent years, we have witnessed the progress of radiomics in methodologies and clinical applications. Its potential has been revealed in helping clinical experts to uncover cancer characteristics that fail to be appreciated by naked eyes. In particular, radiomics not only helps radiologists to make precision diagnoses, but also provides oncologists a useful tool for treatment planning and response evaluation 131, 198. Based on the progress outlined here, further improvements in radiomics will advance clinical care for cancer patients. Next, we will discuss the challenges and opportunities for further radiomics studies.

Big clinical data and data sharing

A reliable conclusion is based on enough data, as is radiomics. Big and standardized clinical data will make radiomics clinically applicable. However, the radiomic researchers need better access to big data, as medical images are dispersed in different hospitals or data centers, and the size of the patient population is the basis of large amounts of medical images. Another challenge is that the collection of medical images is time consuming. Thus, data sharing among institutes and hospitals around the country or even around the world is important for radiomics, although it presents complex logistical problems. The Cancer Imaging Archive (TCIA) provides a good example of data sharing with a large portion of clinical information 199, and it is still growing with contributions from different institutes and hospitals. Researchers could develop new methods and tools to validate their hypotheses with this large dataset 59, 79. The Quantitative Imaging Network (QIN) is another useful project for data integration that may promote radiomic studies 200. However, high quality datasets are still needed for radiomic researches 201.

Reproducibility and quality control

Reproducibility and quality control are involved in the entire radiomic process, including data acquisition, feature extraction and selection, modelling, and validation. Every decision in this process affects the result, its reproducibility, and the quality of the research. Moreover, the inclusion and exclusion criteria of patients in radiomic studies are other important factors related that affect reproducibility. Different criteria may result in the availability of different cohorts for analysis, especially for small cohort studies. Therefore, developing unique inclusion and exclusion criteria for similar clinical applications of radiomics should be considered seriously. Disclosure of the details in every step is important; with the development of data sharing and unique criteria of patients' inclusion, models may be validated using an open dataset that could become the standard for comparison of future studies. Recently, the image biomarker standardization initiative (IBSI) towards seeking standardization for improved reproducibility of high-throughput imaging analyses mainly focused on radiomics was proposed 69. This was a highly valuable step towards the improvement of radiomic studies. In addition, the radiomics quality score (RQS) (Figure 7) was proposed to aid in the assessment of radiomics studies 10. Improving research quality will increase the influence of radiomics.

Figure 7

Flowchart depicting the workflow of radiomics and the application of the RQS. Reproduced with permission from 10. Abbreviations: RQS, radiomics quality score.

Interpretability of radiomic features and models

Clinical experts often think of the radiomic model as a black box that can provide good prediction results for particular clinical outcomes 131, 198, which tends to make radiomics a less accepted approach. Improvement of the interpretability of radiomic features and models is urgently needed. For manually defined imaging features, it is easier to correlate them using pathophysiology; for deep learning, it is a little more difficult. Visualization of deep learning features and prediction models could potentially help to solve this problem. Fortunately, there are many researchers investigating methods to accomplish this goal and several tools have been proposed to diminish the black box perception 202.

Conclusions

Radiomics is still a newly proposed and rapidly progressing field that integrates radiology, oncology, and machine learning. With the growth of clinical data and advanced machine-learning methodologies, it is playing an increasingly important role in precision diagnostics and oncology. Further research should focus on the reproducibility and interpretability to further make radiomics more acceptable in the field.

Table

Specifications of radiomic studies in different cancers

Studies	Study design	No. of patients(training + testing)	No. and type of radiomic features	Statistical analysis	Image Modality	Clinical Characteristics
Brain tumor
Kickingereder et al 88	Retrospective, single center study	79 + 40	6095 (first-order, volume and shape, texture)	Cox regression analysis	MRI	Prognosis
Xi et al 90	Retrospective, single center study	98 + 20	1665 (first-order, size and shape, texture, wavelet)	LASSO, SVM	MRI	Diagnosis
Kickingereder et al 91	Retrospective, single center study	121 + 60	1043 (first-order, size and shape, texture)	Chi-square and Wilcoxon test, LASSO, Cox regression	MRI	Prognosis
Han et al 92	Retrospective, single center study	184 + 93	79 (size and shape, intensity, textural, wavelet)	random forest, U test	MRI	Diagnosis
Li et al 93	Retrospective, single center study	200 + 70	431 (first-order, size and shape, texture, wavelet)	logistic regression, t test, Chi-square test	MRI	Diagnosis
Li et al 95	Retrospective, single center study	180 + 92	431 (first-order, size and shape, texture, wavelet)	t test, Chi-square test, LASSO, SVM	MRI	Diagnosis
Kickingereder et al 97	Retrospective, single center study	112 + 60	4842 (first-order, size and shape, texture, wavelet)	PCA, Cox regression model	MRI	Treatment evaluation
Grossmann et al 98	Retrospective, single center study	57 + 56	65 (first order statistics, size and shape, texture)	PCA, Spearman rank correlation, Wilcoxon test, Cox	MRI	Treatmentevaluation and Prognosis
Papp et al 99	Retrospective, single center study	70	48 (Histogram, shape, texture)	Geometrical Probability Covering Algorithms,	PET	Prognosis
Head and neck cancer
Ren et al 102	Retrospective, single center study	85 + 42	970 (shape, intensity, textural and wavelet)	LASSO	MRI	Diagnosis
Leijenaar et al 103	Retrospective, multi-center study	628 + 150	902 (intensity, shape, texture, Wavelet and Laplacian)	LASSO and multivariable logistic regression	CT	Diagnosis
Zhou et al 104	Retrospective, single center study	122 + 39	257 (intensity, texture, and geometric)	3D CNN	PET/CT	Diagnosis
Chen et al 105	Retrospective, single center study	53	41 (textural features or histograms)	logistic regression analysis	PET/CT	Diagnosis
Zhang et al 107	Retrospective, single center study	88 + 30	970 (shape, intensity, textural and wavelet)	LASSO	MRI	Prognosis
Wang et al 109	Retrospective, single center study	120	591 (texture)	LASSO	MRI	Treatment evaluation
Wu et al 111	Retrospective, single center study	102 + 48	474 (shape, intensity, textural and wavelet)	LASSO	CT	Prognosis
Elhalawani et al 113	Retrospective, single center study	420 + 45	134 (intensity direct, histogram, shape and texture)	multivariate Cox	CT	Prognosis
Vallières et al 114	Retrospective, multi-center study	194 + 106	1615 (intensity, shape, texture)	random forest	PET/CT	Prognosis
Breast cancer
Guo et al 46	Retrospective, single center study	215	463 (morphology, intensity, texture, wavelet)	t test, ICC, LASSO, and SVM	US	Diagnosis
Antropova et al 116	Retrospective, single center study	2060	Deep learning features	CNN, SVM	FFDM, US, MRI	Diagnosis
Antunovic et al 117	Retrospective, single center study	43	20 (size and shape, first-order)	Univariate analysis, hierarchical clustering, exact Fisher's test	PET/CT	Diagnosis
Ha et al 118	Retrospective, single center study	73	109 (texture)	Hierarchical clustering, logistic regression, Cox	PET	Diagnosis and Prognosis
Saha et al 120	Retrospective, single center study	461 + 461	529 (first order, size and shape, texture)	correlation, RF, logistic regression	MRI	Diagnosis
Chan et al 122	Retrospective, single center study	563	8192 (washin and washout intensity values)	PCA, LASSO, logistic regression, K-M survival	MRI	Treatment Evaluation
Braman et al 123	Retrospective, single center study	117	99 (texture)	mRMR, consensus clustering, LDA, DLDA, QDA, naïve Bayes, SVM	MRI	Treatment Evaluation
Chamming et al 76	Retrospective, single center study	85	texture	Mann-Whitney U test, logistic regression	MRI	Treatment Evaluation
Partridge et al 124	Prospective, multi-center study	272	ADC, FTV	Z-test, logistic regression	MRI	Treatment Evaluation
Tran et al 125	Retrospective, single center study	37	40 (texture)	Logistic regression, naïve Bayes, k-NN	DOS	Treatment Evaluation
Park et al 126	Retrospective, single center study	194+100	156 (texture)	Elastic net, Cox Model, K-M analysis	MRI	Prognosis
Lung cancer
Kumar et al 130	Retrospective, multi-center study	38106 + 4234	500 (Deep learning features)	3D CNN	CT	Diagnosis.
Shen et al 131	Retrospective, multi-center study	825 + 275	Deep learning features + 319 (first order, shape and size, textural, and wavelet)	mrmr, SVM, MC-CNN	CT	Diagnosis
Aerts et al 4	Retrospective, multi-center study	474 + 575	440 (intensity, shape, texture and wavelet)	Cluster and Cox	CT	Prognosis and Diagnosis
Parmar et al 135	Retrospective, multi-center study	558 + 320	440 (intensity, shape, texture and wavelet)	Cluster and Cox	CT	Prognosis
Coroller et al 136	Retrospective, single center study	98 + 84	635 (intensity, shape, texture, LoG and Wavelet)	mrmr and Cox	CT	Prognosis
Zhou et al 137	Retrospective, multi-center study	241 + 107	485 (intensity, shape, texture, and wavelet)	concave minimization and SVM	CT	Prognosis
Jia et al 138	Retrospective, single center study	70 + 31	70 (statistical, histogram, morphologic, and texture)	c-means, ICC, linear correlation coefficient, LASSO and Cox	PET	Prognosis
Liu et al 139	Retrospective, single center study	298	219 (size, shape, location, air space, histogram, laws texture and wavelet)	Multiple logistic regression analysis, and SVM	CT	Diagnosis
Zhang et al 141	Retrospective, single center study	140 + 40	485 (shape, intensity, textural and wavelet)	LASSO	CT	Diagnosis
Rios et al 56	Retrospective, multi-center study	353 + 352	440 (intensity, shape and texture)	MRMR, Random forest classifier	CT	Diagnosis
Wu et al 142	Retrospective, multi-center study	198 + 152	440 (intensity, shape and texture)	24 Feature Selection Methods, 3 Classifiers	CT	Diagnosis
Zhu et al 143	Retrospective, single center study	81 + 48	485 (shape, intensity, textural and wavelet)	LASSO	CT	Diagnosis
Fan et al 144	Retrospective, multi-center study	160 + 235	355 (shape, intensity, textural and wavelet)	LASSO	CT	Diagnosis
Fave et al 147	Retrospective, single center study	107	212 (texture, shape)	Cox, AIC	CT	Prognosis
Coroller et al 148	Retrospective, single center study	101 + 26	440 (intensity, shape and texture)	PCA	CT	Treatment evaluation
Aerts et al 149	Retrospective, multi-center study	47	183 (shape, intensity, textural and wavelet)	Coefficient of Variation	CT	Treatment evaluation
Song et al 151	Retrospective, multi-center study	117 + 197	1032 (shape, intensity, textural and wavelet)	LASSO and COX	CT	Treatment evaluation
Song et al 152	Retrospective, single center study	661 + 61	592 (shape, intensity, textural and wavelet)	SVM and COX	CT	Prognosis
Balagurunathan et al 55	Retrospective, single center study	32 + 59	329 (shape, size, and texture)	concordance correlation coefficient, and K-M	CT	Prognosis
Huang et al 155	Retrospective, single center study	141 + 141	132 (textural)	LASSO and COX	CT	Prognosis
Colorectal cancer
Huang et al 7	Retrospective, single center study	326 + 200	150 (histogram and GLCM)	LASSO	CT	Diagnosis
Yang et al 161	Retrospective, single center study	61 + 56	346 (histogram, shape, and texture)	ICC, SVM	CT	Diagnosis
Ke et al 169	Retrospective, single center study	48	103 (histogram, GLCM, shape)	Mann-Whitney test, T test, ANN	MRI	Treatment evaluation
Liu et al 8	Retrospective, single center study	152 + 70	2252 (intensity, shape, wavelet features)	T test, LASSO	MRI	Treatment evaluation
Natally et al 170	Retrospective, single center study	93 + 21	34 (texture, wavelet features)	Wilcoxon rank-sum test, random forest	MRI	Treatment evaluation
Meng et al 173	Retrospective, single center study	54 + 54	485 (Shape, Intensity Texture)	log-rank test, LASSO Cox regression	MRI	Prognosis
Lovinfosze et al 174	Retrospective, single center study	86	(Histogram, texture)	Cox regression analysis	PET/CT	Prognosis
Prostate cancer
Chen et al 176	Retrospective, single center study	266 + 155	781 (first- and second-order, histogram, texture, and Form Factor Parameters)	LASSO	T2WI, ADC	Diagnosis
Wang et al 177	Retrospective, single center study	54	40 (texture features)	RFE-SVM	DCET1WIDWIT2WI	Diagnosis
Algohary et al 178	Retrospective, single center study	30 + 15	308 (first-order statistics, Gabor, Laws, and Haralick)	QDA ,RF,SVM	T2WI, ADC	Diagnosis
Chaddad et al 179	Retrospective, single center study	99	57 (GLCM,JIM)	Random Forest	ADCT2WI	Diagnosis
Shiradkar et al 181	Retrospective, multi-center study	70 + 50	150 (first-order statistics, Gabor, Laws texture)	A machine-learning classifier	T2WI, ADC	Prognosis
Liver and Gastric cancer
Bakr et al 182	Retrospective, single center study	28	4176 (Intensity, Texture, Shape, Edge)	ICCLasso	CT	Diagnosis
Peng et al 183	Retrospective, single center study	184 + 120	980 (Shape, Intensity, texture)	Lasso	CT	Diagnosis
Ma et al 184	Retrospective, single center study	70	485 (shape, intensity, textural and wavelet)	LASSO	CT	Diagnosis
Ba-Ssalamah et al 185	Retrospective, single center study	47	30 (histogram, texture)	LDA, K-NN	CT	Diagnosis
Zhou et al 190	Retrospective, single center study	215	300 (histogram, GLCM)	ICC, LASSO	CT	Prognosis
Akai et al 191	Retrospective, single center study	127	96 (filtration, histogram)	Multivariate CoxRandom Forest	CT	Prognosis
Cozzi et al 192	Retrospective, single center study	138	35 (shape and size, histogram, second and high order)	Cox	CT	Prognosis
Francesco et al 195	Retrospective, single center study	56	107 (first- and second-order Textural, shape and size)	Cox	CT	Prognosis