Radiomics and Magnetic Resonance Imaging of Rectal Cancer: From Engineering to Clinical Practice

While cross-sectional imaging has seen continuous progress and plays an undiscussed pivotal role in the diagnostic management and treatment planning of patients with rectal cancer, a largely unmet need remains for improved staging accuracy, assessment of treatment response and prediction of individual patient outcome. Moreover, the increasing availability of target therapies has called for developing reliable diagnostic tools for identifying potential responders and optimizing overall treatment strategy on a personalized basis. Radiomics has emerged as a promising, still fully evolving research topic, which could harness the power of modern computer technology to generate quantitative information from imaging datasets based on advanced data-driven biomathematical models, potentially providing an added value to conventional imaging for improved patient management. The present study aimed to illustrate the contribution that current radiomics methods applied to magnetic resonance imaging can offer to managing patients with rectal cancer.


Introduction
With over 1.8 million new cases reported each year, rectal cancer (RC) is the third most common cancer in men and the second in women, as well as the fourth leading cause of death globally. Its overall incidence is higher in the industrialized world, but it is also rapidly increasing in developing countries [1][2][3]. Locally advanced rectal cancer (LARC) is defined as a tumor penetrating the entire bowel wall (stages 2, T3/T4N0) and/or with involvement of regional lymph nodes (stage 3, any T N1/N2); the standard of care for patients with LARC currently includes neoadjuvant chemoradiotherapy (nCRT) followed by total mesorectal excision [4][5][6].
Magnetic resonance imaging (MRI) is the most accurate imaging modality for both RC primary staging and restaging after treatment. In particular, it plays an essential role in the local staging of the disease both before and after nCRT, in that it can provide comprehensive The aim of this paper was to give an overview of the biophysical background, technical issues and potential applications of radiomics in the management of patients with RC. To this purpose, we searched PubMed (https://pubmed.ncbi.nlm.nih.gov/, accessed on 8 April 2021) for English language papers (including original research and review articles) with keywords, including "radiomics", "rectal cancer", "rectal MRI", "staging", "neoadjuvant (chemoradiation therapy)", "surgery", "prognosis", "outcome", "downstaging", "response" and "lymph node/nodal". For technical aspects, we also searched PubMed and IEEE Xplore (https://ieeexplore.ieee.org/Xplore/home.jsp, accessed on 8 April 2021) for English language papers with keywords, including "features standardization", "deep learning architecture", "machine-learning algorithms", "feature selection", "classifiers", "validation techniques" and "deep learning/machine learning studies design".

Radiomics Workflow
The radiomics workflow is composed of various steps, which are summarized in Figure 1 and are discussed below. The aim of this paper was to give an overview of the biophysical background, technical issues and potential applications of radiomics in the management of patients with RC. To this purpose, we searched PubMed (https://pubmed.ncbi.nlm.nih.gov/, accessed on 8 April 2021) for English language papers (including original research and review articles) with keywords, including "radiomics", "rectal cancer", "rectal MRI", "staging", "neoadjuvant (chemoradiation therapy)", "surgery", "prognosis", "outcome", "downstaging", "response" and "lymph node/nodal". For technical aspects, we also searched PubMed and IEEE Xplore (https://ieeexplore.ieee.org/Xplore/home.jsp, accessed on 8 April 2021) for English language papers with keywords, including "features standardization", "deep learning architecture", "machine-learning algorithms", "feature selection", "classifiers", "validation techniques" and "deep learning/machine learning studies design".

Radiomics Workflow
The radiomics workflow is composed of various steps, which are summarized in Figure 1 and are discussed below.

Image Acquisition, Segmentation and Feature Extraction
Sharing common imaging protocols across centers is of paramount importance to minimize any potential bias, affecting measurements (e.g., magnet strength, MR coil, the field of view, spatial resolution, various MR sequence parameters, etc.) [30]. However, in a clinical setting, it might be difficult to use standardized protocols on a multicenter basis, especially in the case of retrospective studies [31]. Different solutions have been proposed to reduce the influence of acquisition protocols on radiomics studies. One approach is based on leveraging radiomics variability as a feature selection tool by using test-retest analysis and eliminating radiomics features with a high variability [23,32]. However, this approach has the drawback of not evaluating the diagnostic performance of each feature. In fact, it has been demonstrated that features robust to parameter variations commonly resulted in a robust, but not necessarily high, diagnostic performance [33], whereas highly variable radiomics features could maintain a high diagnostic performance [34].
A different strategy is to apply post-reconstruction batch harmonization techniques to minimize feature variabilities across centers, such as global scaling (min-max scaling), z-standardization, the ratio with the signal intensity of healthy tissue, which is not affected

Image Acquisition, Segmentation and Feature Extraction
Sharing common imaging protocols across centers is of paramount importance to minimize any potential bias, affecting measurements (e.g., magnet strength, MR coil, the field of view, spatial resolution, various MR sequence parameters, etc.) [30]. However, in a clinical setting, it might be difficult to use standardized protocols on a multicenter basis, especially in the case of retrospective studies [31]. Different solutions have been proposed to reduce the influence of acquisition protocols on radiomics studies. One approach is based on leveraging radiomics variability as a feature selection tool by using test-retest analysis and eliminating radiomics features with a high variability [23,32]. However, this approach has the drawback of not evaluating the diagnostic performance of each feature. In fact, it has been demonstrated that features robust to parameter variations commonly resulted in a robust, but not necessarily high, diagnostic performance [33], whereas highly variable radiomics features could maintain a high diagnostic performance [34].
A different strategy is to apply post-reconstruction batch harmonization techniques to minimize feature variabilities across centers, such as global scaling (min-max scaling), z-standardization, the ratio with the signal intensity of healthy tissue, which is not affected by the disease, histogram-matching (where intensity histograms are transformed to match a reference intensity histogram) or the ComBat harmonization method [35]. Each of these techniques has its own strengths and weaknesses, which could affect both the values of the radiomics features and the prognostic value of radiomics-based classification methods differently [35]. Therefore, a thorough assessment of the most appropriate technique to be used is advised when developing radiomics models for multicenter studies.
Once all image datasets have been acquired, the tissues of interest must be outlined by tracing regions of interest (ROIs) inside or around them to differentiate them from neighboring structures (e.g., tumor vs. normal tissue) ( Figure 2). This task is called segmentation and can be manual, semiautomatic or fully automatic. Manual and semiautomatic methods are time-consuming tasks, hence difficult to implement in daily radiological and clinical practice; they also suffer from high intra-and interobserver variability [20]. Several studies have demonstrated that differences among segmentations can strongly affect the values of radiomics features and, consequently, the robustness of the radiomics model [36][37][38]. Therefore, when developing classifiers based on segmentation of ROIs, it is essential to determine the variability of radiomics features concerning the tumor delineation process and to implement the appropriate strategies to deal with this issue, e.g., by eliminating those features whose variability is higher than their prognostic value. by the disease, histogram-matching (where intensity histograms are transformed to match a reference intensity histogram) or the ComBat harmonization method [35]. Each of these techniques has its own strengths and weaknesses, which could affect both the values of the radiomics features and the prognostic value of radiomics-based classification methods differently [35]. Therefore, a thorough assessment of the most appropriate technique to be used is advised when developing radiomics models for multicenter studies. Once all image datasets have been acquired, the tissues of interest must be outlined by tracing regions of interest (ROIs) inside or around them to differentiate them from neighboring structures (e.g., tumor vs. normal tissue) ( Figure 2). This task is called segmentation and can be manual, semiautomatic or fully automatic. Manual and semiautomatic methods are time-consuming tasks, hence difficult to implement in daily radiological and clinical practice; they also suffer from high intra-and interobserver variability [20]. Several studies have demonstrated that differences among segmentations can strongly affect the values of radiomics features and, consequently, the robustness of the radiomics model [36][37][38]. Therefore, when developing classifiers based on segmentation of ROIs, it is essential to determine the variability of radiomics features concerning the tumor delineation process and to implement the appropriate strategies to deal with this issue, e.g., by eliminating those features whose variability is higher than their prognostic value.

Automatic Segmentation via Deep Learning Algorithms
Fully automated segmentation methods have the potential of overcoming the aforementioned limitations; however, the intrinsic complexity of rectal MR images, together with differences in signal intensity among different patients and different image acquisition protocols and MR scanners, can form substantial hurdles, which have long prevented developing radiomics as a tool for supporting the diagnosis. Newer deep learning (DL) systems can outperform conventional pattern recognition algorithms in segmenting RC lesions from clinical MR images [39][40][41][42][43].
While DL (also known as "deep structured learning") represents a subgroup of machine learning techniques, it differs from the latter in that learning is fulfilled directly from data and not from ad-hoc parameters chosen beforehand.
More specifically, when ML algorithms are developed, it is mandatory to decide which quantitative features need to be extracted and used to train the network based on a priori knowledge of the problem. Conversely, DL methods are directly fed with images and automatically determine which characteristics of the images are the most important for solving the problem. In DL methods, the training phase can be either supervised (i.e., when the data provided and used from the network to learn [training] contain information regarding the class they belong to which had previously been defined by an expert) or unsupervised (i.e., when data do not contain information regarding the class they belong to).
The core of DL is in artificial neural networks, the main components of which are illustrated in Figure 3.
While DL (also known as "deep structured learning") represents a subgroup of machine learning techniques, it differs from the latter in that learning is fulfilled directly from data and not from ad-hoc parameters chosen beforehand.
More specifically, when ML algorithms are developed, it is mandatory to decide which quantitative features need to be extracted and used to train the network based on a priori knowledge of the problem. Conversely, DL methods are directly fed with images and automatically determine which characteristics of the images are the most important for solving the problem. In DL methods, the training phase can be either supervised (i.e., when the data provided and used from the network to learn [training] contain information regarding the class they belong to which had previously been defined by an expert) or unsupervised (i.e., when data do not contain information regarding the class they belong to).
The core of DL is in artificial neural networks, the main components of which are illustrated in Figure 3. Briefly, an input image is divided into feature maps (i.e., convolutional layers) representing different characteristics of the same image. Each convolutional layer detects different levels of features from simple patterns, such as edges and gradients, to more abstract features. These feature maps are then fed into a net, which processes each of these features and activates those neurons, which are related to the recognition of a specific pattern. Multiple architectures of neural networks have been proposed, which differ in the following characteristics: • Activation function of neurons, i.e., how different inputs can either activate or not a specific neuron, defined by variables, such as "weights" and "bias"; • Architecture, defining the number of neurons and their interconnections; Briefly, an input image is divided into feature maps (i.e., convolutional layers) representing different characteristics of the same image. Each convolutional layer detects different levels of features from simple patterns, such as edges and gradients, to more abstract features. These feature maps are then fed into a net, which processes each of these features and activates those neurons, which are related to the recognition of a specific pattern. Multiple architectures of neural networks have been proposed, which differ in the following characteristics: • Activation function of neurons, i.e., how different inputs can either activate or not a specific neuron, defined by variables, such as "weights" and "bias"; • Architecture, defining the number of neurons and their interconnections; • Learning algorithm, i.e., the implemented mathematical function, which allows neurons to learn.
Learning consists of varying the weights and bias values of each neuron within the network so that the system output matches the true output. The first artificial neural network was the Perceptron, which was introduced by Rosenblatt in 1958 [44]. Its structure is quite simple in that it is made up of a single neuron and can categorize the input into two classes. The Perceptron was followed by developing several additional neural networks, of which the best-known are convolutional neural networks (CNNs). Convolutional neural networks can learn directly from two-or three-dimensional images based on applying specific filters (i.e., convolutional layers in Figure 3), which extrapolate a series of specific characteristics useful for data classification. Starting from this type of network, the neural network U-net (which was specifically created for medical imaging and is characterized by a U-shaped structure) was implemented for the first time [45]. The biggest improvement brought by the U-net consists of being able to segment areas of clinical interest within the image (e.g., tumor tissue, blood vessels, etc.) instead of classifying an input image into two different classes (e.g., similar to a CNN).
To date, there has been an ever-increasing number of studies showing systems based on neural networks for the segmentation of RC, given the greater interest in such technolo-gies and their potential medical applications. The best results related to the application of CNNs have been illustrated by Trebeschi et al., who implemented a CNN-based system using multiparametric rectal MRI [46]. This system accepts T2-weighted and DW images (b-values equal to 0 and 1000 s/mm 2 ) as input and is capable of classifying each voxel (3D region of the image) as neoplastic or healthy tissue (Figure 4a). The main disadvantage of this system is that its performance decreases considerably when input MR images have been acquired with a wider field of view than that of the learning dataset since portions of healthy tissue (such as testicles or subcutaneous adipose tissue) are classified as tumor tissue (Figure 4b). Huang et al. implemented a U-net, which allowed successful tumor localization on T2-weighted images [47]; however, it required a long computational time, which negated any substantial practical advantage than manual segmentation. The same authors showed the possibility of training a neural network, even using an unbalanced dataset, such as a hospital dataset in which the percentage of patients with a given disease is lower than that of healthy individuals [48]. This was possible by "weighting" the effects of pathological cases more and nonpathological cases less on the network.  Panic et al. [49] implemented an innovative system based on CNNs for automatic segmentation, which allowed overcoming the aforementioned limitations. The segmentation mask was obtained by classifying all pixel groups (ROI 3 × 3) forming the input image. The classification was carried out using three CNNs, each of them classifying the same acquired ROI from three different MRI sequences. Considering the different information provided by each individual sequence, it was possible to obtain a final classification of the region.
Current research is yielding encouraging results; however, some efforts are required Panic et al. [49] implemented an innovative system based on CNNs for automatic segmentation, which allowed overcoming the aforementioned limitations. The segmentation mask was obtained by classifying all pixel groups (ROI 3 × 3) forming the input image. The classification was carried out using three CNNs, each of them classifying the same acquired ROI from three different MRI sequences. Considering the different information provided by each individual sequence, it was possible to obtain a final classification of the region.
Current research is yielding encouraging results; however, some efforts are required to bridge the gap between research and clinical needs. First, most research algorithms are validated on internal datasets, which means that protocols and images are quite similar. Validation on larger and, especially, multicentric databases are of key importance. Furthermore, these algorithms are most often not integrated into a single graphical user interface, which can be used by an inexperienced user. Therefore, it would be a challenging but extremely important task to devise a tool, which could be run by anyone via his/her favorite reporting platform.

Feature Selection, Model Construction and Statistical Analysis
Once tumor segmentation has been carried out, quantitative features must be extracted, which characterize the selected tissue(s). The main features extracted are divided into firstorder statistics features, which are derived from gray-level histograms, and second-order statistics features, which are computed from matrices relating each pixel of the image to those in its surroundings, e.g., using gray-level co-occurrence matrices (GLCM) or graylevel run-length matrices (GLRLM). It is essential that those steps be standardized as much as possible to avoid any bias, which may occur due to the inhomogeneity of the source data, especially in the perspective of performing multicentric studies and of future validation for clinical use [50].
A large number of variables is often extracted, resulting in the generation of features, which are highly correlated among themselves and could hence lead to overfitting (i.e., error due to over-adaptation of the prediction model to the input data), thereby requiring a prior selection of extracted features. Filter feature selection algorithms rank features according to the relationship between each feature and the output (e.g., using correlation, area under the receiver operating characteristic (ROC) curve, chi-squared test, etc.); ranking scores are then used to choose those input variables, which will be entered into the model. Conversely, wrapper feature selection methods are based on a specific machine-learning algorithm to evaluate the performance of different feature subsets. A greedy search approach is followed by evaluating all the possible combinations of features against a performance metric. For regression, this metric can be p, R-squared or adjusted R-squared values, whereas, for classification, it can be accuracy, precision, recall, f1-score, etc. Finally, the algorithm selects the combination of features yielding optimal results for the specified machine-learning algorithm. Intrinsic methods feature selection during the model building process (e.g., least absolute shrinkage and selection operator (LASSO), Elastic Net, etc.) [51].
Once the best performing features have been selected, patient clinical, biological and genetic data should ideally be incorporated into the model-building to create a complete clinical decision support system. Several classification methods can be used depending on the number of patients available and the clinical purpose (e.g., revelation and segmentation, characterization). Different classification methods have been proposed, which differ in learning approach (i.e., supervised or unsupervised) and type of correlation with the output (e.g., linear, polynomial, Gaussian, etc.). The most common classifiers are logistic regression, k-nearest neighbor, naïve Bayes classifier, support vector machines, decision tree, neural network and deep learning [52]. A recent study has concluded that there was no single best classifier across all datasets [53]; hence, the performance of different classifiers should be assessed to obtain more reliable, consistent and generalizable models.
After training the model on the data available (training set) and optimizing it on the second group of patients (testing set), its performance must be validated on a dataset Having an external validation dataset can be very difficult since the images require annotation, implying that an experienced radiologist would review each image and define the ground truth together with the oncologist and/or the pathologist. Moreover, transferring images between different centers can be problematic due to technical and privacy-related issues. A successful image exchange program can be transformational for both care providers and patients as it would allow training and validating machine learning models involving a massive scale population, thus driving better health outcomes. However, in pilot or preliminary studies, it is always mandatory to carry out validation if having an external dataset is not feasible. The latter can be accomplished using different strategies, such as k-fold, bootstrap, random subsampling or a nested approach. The biggest issue with internal validation is the potential leakage of the feature selection algorithm into the whole data, which might lead to overly optimistic results. For creating such unseen data sets, although the hold-out technique seems to be the most appropriate internal validation method, there is also a nested cross-validation technique, which is primarily used for this purpose and might give similar estimates to an independent validation [54].

Radiomics for the Personalized Management of RC Patients: Current Evidence and Perspectives
The classification models illustrated above lend themselves to providing insights into tissue structure, which can be useful to characterize tumor heterogeneity, predict patient response to therapy and overall survival, and assess the relationship between radiomic and genomic characteristics, under the assumption that imaging reflects not only the macroscopic features of tissues but also their cellular and molecular properties [55]. For the purpose of RC management, radiomics data can be obtained, which complement and/or refine the diagnostic information achieved by conventional MR protocols and image analysis.
In the future, radiomics can be expected to be an additional tool for RC patient management with the ability to support clinicians and overcoming the main challenges of conventional imaging, such as the overstaging of early RC and lack of accuracy in detecting nodal metastases. Those applications could drastically change patient management and therapeutic options, and more patients may take advantage of local excision if properly staged as T1 [28,29]. Moreover, the possibility of providing data, which go beyond the visual assessment and conventional qualitative and quantitative analysis of MR images makes radiomics a promising tool for predicting outcome before surgery and evaluating the response to nCRT (potentially allowing the selection of respondents from nonrespondents before the beginning of the completion of potentially toxic treatments), guiding towards the optimal therapeutic approach on an individualized basis [25,36,56,57]. Of note, in contrast to conventional biopsy (which samples only a selected portion of the tumor), radiomics analysis allows gathering information from the entire tumor, thus taking into account tissue heterogeneity, which may have an impact on lesion characterization and treatment planning (e.g., due to the presence of multiple tumor clones and differences in the tumor microenvironment, which may affect tumor sensitivity and resistance to nCRT or molecular target agents) [58].

Staging
In a retrospective study involving a total of 152 patients with RC, who underwent surgery alone (i.e., without any neoadjuvant therapy), Ma et al. found that radiomics analysis carried out on high-resolution T2-weighted 3-Tesla MR images using a support vector machine and a random forest algorithm yielded an area under the receiver operating characteristic curve (AUC) as high as 0.862 (sensitivity 83.3%, specificity 85.0%) and 0.746 (sensitivity 79.3%, specificity 72.2%) in predicting the degree of differentiation, and the T-and N-stages of the lesions, respectively [59]. Lu et al. reported that the texture features obtained from sagittal fat-suppression combined with transverse T2-weighted images in patients undergoing preoperative MRI may be valuable in selecting an optimal treatment strategy, allowing differentiation of T1/2 from T3/4 tumors with an AUC of 0.740 [60] (Figure 5). Similarly, Sun et al. [28]   Lymph node assessment before and after treatment using conventional MRI can be challenging, spurring interest in developing advanced techniques, which could bridge this gap. Liu et al. reported that DWI-based texture analysis could predict local invasion depth (stage pT1-2 versus pT3-4) and nodal status (pN0 versus pN1-2) of RC with significant differences occurring between pN0 and pN1-2 tumors concerning the mean apparent diffusion coefficient (ADCmean), whereas ADCmax and entropy were independent predictors of positive nodal status [61]. Li et al. [29] recently built and validated a combined radiomic-clinical nomogram for predicting nodal metastases in the preoperative clinical setting of patients with colorectal cancer, achieving good accuracy, AUC, sensitivity and specificity (73.7%, 0.750, 60.2% and 84.3%, respectively).
Other studies have shown the potential of MRI texture features for providing valuable information in identifying the status of RC lymph node invasion [62,63]. A combined model derived from radiomic signatures and restaging results in patients with LARC undergoing preoperative and posttreatment MRI could predict positive lymph node status with a negative predictive value of 87.8% and 100% in posttreatment MRI T3-T4 and T1-T2 tumors, respectively [64].

Assessment of Treatment Response
In a study by Yi et al., MRI-based radiomics was able to predict response to nCRT in patients with LARC from conventional T2-weighted MR images, allowing prediction of complete pathological response, good response (as defined by the Dowrak/Rödel system [65]) and downstaging with AUCs of 0.91, 0.90 and 0.93, respectively [66] (Figure 6). Lymph node assessment before and after treatment using conventional MRI can be challenging, spurring interest in developing advanced techniques, which could bridge this gap. Liu et al. reported that DWI-based texture analysis could predict local invasion depth (stage pT1-2 versus pT3-4) and nodal status (pN0 versus pN1-2) of RC with significant differences occurring between pN0 and pN1-2 tumors concerning the mean apparent diffusion coefficient (ADC mean ), whereas ADC max and entropy were independent predictors of positive nodal status [61]. Li et al. [29] recently built and validated a combined radiomic-clinical nomogram for predicting nodal metastases in the preoperative clinical setting of patients with colorectal cancer, achieving good accuracy, AUC, sensitivity and specificity (73.7%, 0.750, 60.2% and 84.3%, respectively).
Other studies have shown the potential of MRI texture features for providing valuable information in identifying the status of RC lymph node invasion [62,63]. A combined model derived from radiomic signatures and restaging results in patients with LARC undergoing preoperative and posttreatment MRI could predict positive lymph node status with a negative predictive value of 87.8% and 100% in posttreatment MRI T3-T4 and T1-T2 tumors, respectively [64].

Assessment of Treatment Response
In a study by Yi et al., MRI-based radiomics was able to predict response to nCRT in patients with LARC from conventional T2-weighted MR images, allowing prediction of complete pathological response, good response (as defined by the Dowrak/Rödel system [65]) and downstaging with AUCs of 0.91, 0.90 and 0.93, respectively [66] (Figure 6). Diagnostics 2021, 11, x FOR PEER REVIEW 11 of 16 Figure 6. Performance of three predictive models in predicting downstaging, pathological complete response (PCR) and good response in LARC patients. The red box highlights the performance score associated with the highest AUC. Adapted from [66].
T2-weighted images are often used because they tend to offer better diagnostic performance in RC staging [67]; however, other studies have incorporated additional sequences conveying functional information, e.g., DW images. Liu et al. built a radiomics model from T2-weighted and DW images obtained in 222 patients with LARC before and after nCRT, enabling individualized prediction of complete pathological response (AUC as high as 0.9756 in the validation cohort) and, thus, potentially aiding the identification of LARC patients, who could avoid surgery after nCRT [68]. More recently, by performing segmentation of T2-weighted and diffusion-weighted MR images (b-800 images and ADC maps), Bulens et al. created models using LASSO regression analysis, which enabled a noninvasive prediction of response to neoadjuvant treatment in patients with LARC undergoing nCRT followed by surgery, with a higher AUC between 0.83 and 0.86 for predicting complete or near-complete pathological response (ypT0-1N0), respectively [57].
An MRI-based radiomic nomogram has been developed by Wang et al., which can accurately differentiate good and poor responders in patients with LARC undergoing nCRT and achieve significant risk stratification concerning progression-free survival [69]. Zhou et al. found that pre-therapeutic, multiparametric MRI radiomic features could predict nonresponse to neoadjuvant therapy in patients with LARC, yielding an AUC of 0.822 [70]. Delta-radiomics signatures obtained in patients with LARC before and after nCRT and surgery have also been shown to be able to successfully predict treatment outcomes and serve as independent prognostic factors [71].

Prediction of Individual Patient Prognosis and Potential Eligibility on Target Therapies
Radiomics analysis may provide information, which could impact overall patient prognosis and potential eligibility for target therapies [72]. A combined model, including the radiomics and clinical features obtained from pretreatment T2-weighted MR images, was able to predict the likelihood of developing distant metastases, thus potentially aiding in tailoring treatment strategies in high-risk patients [73]. Cui et al. developed and validated a T2-weighted image-based radiomics signature for the specific prediction of KRAS mutation status in 304 patients with RC using a support vector machine algorithm providing an AUC of 0.722 in the training dataset. Importantly, a KRAS mutation was not associated with baseline clinical and histopathological features, suggesting that the proposed radiomics signature may be useful for supplementing genomic analysis to determine T2-weighted images are often used because they tend to offer better diagnostic performance in RC staging [67]; however, other studies have incorporated additional sequences conveying functional information, e.g., DW images. Liu et al. built a radiomics model from T2-weighted and DW images obtained in 222 patients with LARC before and after nCRT, enabling individualized prediction of complete pathological response (AUC as high as 0.9756 in the validation cohort) and, thus, potentially aiding the identification of LARC patients, who could avoid surgery after nCRT [68]. More recently, by performing segmentation of T2-weighted and diffusion-weighted MR images (b-800 images and ADC maps), Bulens et al. created models using LASSO regression analysis, which enabled a noninvasive prediction of response to neoadjuvant treatment in patients with LARC undergoing nCRT followed by surgery, with a higher AUC between 0.83 and 0.86 for predicting complete or near-complete pathological response (ypT0-1N0), respectively [57].
An MRI-based radiomic nomogram has been developed by Wang et al., which can accurately differentiate good and poor responders in patients with LARC undergoing nCRT and achieve significant risk stratification concerning progression-free survival [69]. Zhou et al. found that pre-therapeutic, multiparametric MRI radiomic features could predict nonresponse to neoadjuvant therapy in patients with LARC, yielding an AUC of 0.822 [70]. Delta-radiomics signatures obtained in patients with LARC before and after nCRT and surgery have also been shown to be able to successfully predict treatment outcomes and serve as independent prognostic factors [71].

Prediction of Individual Patient Prognosis and Potential Eligibility on Target Therapies
Radiomics analysis may provide information, which could impact overall patient prognosis and potential eligibility for target therapies [72]. A combined model, including the radiomics and clinical features obtained from pretreatment T2-weighted MR images, was able to predict the likelihood of developing distant metastases, thus potentially aiding in tailoring treatment strategies in high-risk patients [73]. Cui et al. developed and validated a T2-weighted image-based radiomics signature for the specific prediction of KRAS mutation status in 304 patients with RC using a support vector machine algorithm providing an AUC of 0.722 in the training dataset. Importantly, a KRAS mutation was not associated with baseline clinical and histopathological features, suggesting that the proposed radiomics signature may be useful for supplementing genomic analysis to determine KRAS expression in RC patients [74]. Using T2-weighted MRI-based texture analysis, Oh et al. identified three imaging features, which could preoperatively differentiate mutant from wild-type KRAS, yielding a sensitivity, specificity and accuracy of 84%, 80%, and 81.7%, respectively [75]. In a study involving 158 patients with pathologically proven RC, who had undergone preoperative MRI, the mean values of six texture parameters (mean, variance, skewness, entropy, gray-level nonuniformity, run-length nonuniformity) were significantly higher in KRAS-mutant patients than KRAS wild-type patients having AUC values of texture features ranging from 0.703 to 0.81, and higher T stage and lower ADC values occurring in KRAS-mutant cancers [76] (Figure 7). Meng et al. developed a radiomic model from MRI datasets of 345 patients with RC taking into account multiple factors, such as lymph node metastasis, tumor differentiation grade, a fraction of Ki-67-positive tumor cells, human epidermal growth factor receptor 2 (HER-2) expression and KRAS-2 gene mutation status, yielding an AUC of 0.699 for signatures evaluating Ki-67 and of 0.697 for an integrated evaluation model incorporating radiomics signature and MRI-reported lymph node status [27]. KRAS expression in RC patients [74]. Using T2-weighted MRI-based texture analysis, Oh et al. identified three imaging features, which could preoperatively differentiate mutant from wild-type KRAS, yielding a sensitivity, specificity and accuracy of 84%, 80%, and 81.7%, respectively [75]. In a study involving 158 patients with pathologically proven RC, who had undergone preoperative MRI, the mean values of six texture parameters (mean, variance, skewness, entropy, gray-level nonuniformity, run-length nonuniformity) were significantly higher in KRAS-mutant patients than KRAS wild-type patients having AUC values of texture features ranging from 0.703 to 0.81, and higher T stage and lower ADC values occurring in KRAS-mutant cancers [76] (Figure 7). Meng   ROC curves corresponding to quantitative texture features derived from T2-weighted MR images for differentiating KRAS mutation status in rectal cancer. Adapted from [76].
Radiomics analysis of T2-weighted images also has the potential of identifying microsatellite instability in the preoperative setting of patients with RC [77,78]. Similar findings have also been obtained by applying radiomics methods to computed tomography (CT) [79] and dual-energy CT datasets [80].

Conclusions
Rectal cancer is a complex disease for which radiomics has the potential of providing an added value to conventional MR imaging, especially in terms of improved staging, evaluation of treatment response and prediction of patient outcome. While radiomics is an area of active research in full development, radiologists should be aware of its potential capabilities, and efforts should be made to maximize standardization of imaging protocols Figure 7. ROC curves corresponding to quantitative texture features derived from T2-weighted MR images for differentiating KRAS mutation status in rectal cancer. Adapted from [76].
Radiomics analysis of T2-weighted images also has the potential of identifying microsatellite instability in the preoperative setting of patients with RC [77,78]. Similar findings have also been obtained by applying radiomics methods to computed tomography (CT) [79] and dual-energy CT datasets [80].

Conclusions
Rectal cancer is a complex disease for which radiomics has the potential of providing an added value to conventional MR imaging, especially in terms of improved staging, evaluation of treatment response and prediction of patient outcome. While radiomics is an area of active research in full development, radiologists should be aware of its potential capabilities, and efforts should be made to maximize standardization of imaging protocols and data collection on a multicenter basis to validate findings on a broad scale for prospective use in clinical practice.
Funding: This research received no external funding.