Role of AI and Radiomic Markers in Early Diagnosis of Renal Cancer and Clinical Outcome Prediction: A Brief Review

Simple Summary Renal cancer (RC) is ranked tenth among all types of cancer in men and women worldwide. Artificial intelligence (AI) and radiomics have allowed the development of AI-based computer-aided diagnostic/prediction (AI-based CAD/CAP) systems for noninvasive and precise diagnosis of RC and prediction of clinical outcome at an early stage. This, in turn, can conserve time, effort, and resources, ultimately benefiting both patients and healthcare providers. This review summarizes the studies from the last decade that used AI and radiomic markers for the early diagnosis of RC and prediction/assessment of clinical outcome/treatment response. Finally, a deep discussion, suggestions, and possible future avenues for improving diagnostic and treatment prediction performance is introduced, which might help fill the research gap. Abstract Globally, renal cancer (RC) is the 10th most common cancer among men and women. The new era of artificial intelligence (AI) and radiomics have allowed the development of AI-based computer-aided diagnostic/prediction (AI-based CAD/CAP) systems, which have shown promise for the diagnosis of RC (i.e., subtyping, grading, and staging) and prediction of clinical outcomes at an early stage. This will absolutely help reduce diagnosis time, enhance diagnostic abilities, reduce invasiveness, and provide guidance for appropriate management procedures to avoid the burden of unresponsive treatment plans. This survey mainly has three primary aims. The first aim is to highlight the most recent technical diagnostic studies developed in the last decade, with their findings and limitations, that have taken the advantages of AI and radiomic markers derived from either computed tomography (CT) or magnetic resonance (MR) images to develop AI-based CAD systems for accurate diagnosis of renal tumors at an early stage. The second aim is to highlight the few studies that have utilized AI and radiomic markers, with their findings and limitations, to predict patients’ clinical outcome/treatment response, including possible recurrence after treatment, overall survival, and progression-free survival in patients with renal tumors. The promising findings of the aforementioned studies motivated us to highlight the optimal AI-based radiomic makers that are correlated with the diagnosis of renal tumors and prediction/assessment of patients’ clinical outcomes. Finally, we conclude with a discussion and possible future avenues for improving diagnostic and treatment prediction performance.

To the best of our knowledge, there is no agreement on the most reliable radiomic markers that can be used to develop a comprehensive AI-based system for the purposes of diagnosing renal tumors and predicting clinical outcome/treatment response simultaneously. After reviewing the state-of-the-art studies that have been developed in the last decade, we highlight the most common radiomic markers that can be correlated with both the diagnosis and treatment response prediction of renal tumors, potentially opening the door for future investigation and development of comprehensive diagnostic and predictive radiomic/AI-based systems.

Computed Tomography (CT) Studies
In the differentiation of benign and malignant renal tumors, Hodgdon et al. and Yang et al. [14,27] discovered that first-and second-order texture markers of unenhanced CT (Phase 1) yielded an accuracy range of 82% to 91% and an area under the curve (AUC) range of 0. 73 [28,[52][53][54] found that first-and second-order texture markers of multiphasic CECT, along with SVM classifiers, achieved an accuracy range of 72% to 94% and an AUC range of 0.75 to 0.97. Yan et al. [55], Ma et al. [56], and Tang et al. [57] achieved comparable results with texture markers from multiphasic CECT. For instance, Yan et al. [55] employed artificial neural networks (ANNs) in conjunction with texture markers and attained 97% accuracy on a relatively small, unbalanced dataset (N = 50). Ma et al. [56] and Tang et al. [57] reported an AUC range of 0.67 to 0.93 using logistic regression (LR) classifiers in combination with texture markers. An expanded study by Ma et al. [56] found high accuracy using the nephrographic phase (Phase 3) of CECT with an AUC range of 0.74 to 0.89.
Nassiri et al. [58] extracted higher-order texture markers and shape markers from Phase 3 CECT and attained an accuracy range of 74% to 79% and an AUC range of 0.77 to 0.84 using Adaboost and random forests (RF) classifiers. Yap et al. [59] used the same markers extracted from multiphasic CECT and reported an AUC range of 0.65 to 0.75. Without the need for higher-order markers, Uhlig et al. [36] yielded an 84% accuracy and an AUC of 0.83 employing an RF classifier. Coy et al. [60] yielded the highest diagnostic accuracy of 74% using deep learning (DL) on Phase 4 (delayed phase) of CECT. Entropy as a first-order texture marker was extracted from unenhanced CT by Kim et al. [61] and was employed to distinguish between RCC and benign cysts using logistic regression (LR), achieving an AUC of 0.92.
Tanaka et al. [62] implemented a DL pipeline using the Inception-V3 convolutional neural network (CNN) and reported an accuracy range of 41% to 88% and an AUC range of 0.49 to 0.85, favoring Phase 2 CECT over other contrast phases. Li et al. [63] successfully distinguished benign ONC from malignant chrRCC by employing first-and second-order texture markers of multiphasic CECT along with an SVM classifier, resulting in 95% accuracy and an AUC of 0.85. The authors suggested that phases 2 and 3 outperformed other contrast phases for the specific task. In subsequent studies with larger datasets [64,65], they discovered that incorporating clinical factors improved the overall diagnostic accuracy. Meanwhile, Zabihollahy et al. [66] employed 2D and 3D CNNs in conjunction with semiautomated and automated tumor segmentation methods, reporting an accuracy range of 77% to 84%.
For RCC subtyping, Deng et al. and Yu et al. [34,67] demonstrated the efficacy of first-order texture markers, namely mean, standard deviation (STD), kurtosis, skewness, entropy, and median of Phase 3 CECT. Deng et al. [34] achieved 47% accuracy and an AUC between 0.80 and 0.84 using LR, while Yu et al. [67] yielded an AUC ranging from 0.86 to 0.92 using SVM. Furthermore, Shehata et al. [68] integrated shape, texture, and functional radiomic markers and obtained an accuracy between 79% and 98%, sensitivity from 0.89 to 0.95, and specificity from 0.91 to 1.00 using a multilayer perceptron artificial neural network (MLP-ANN). They also identified Phase 3 as the most useful phase for RCC subtyping. Zhang et al. [35] concurred with the significance of these markers but extracted them from Phase 2, achieving an accuracy of 78% to 87% and an AUC of 0.94 to 0.96 using an SVM classifier.
Verghase et al. [69] demonstrated that multiphasic CECT first-, second-, and higherorder texture markers are extremely important. They performed statistical analysis using stepwise LR and achieved an AUC range of 0.80 and 0.98. Uhlig et al. [36,70], in two subsequent studies, suggested first-and second-order texture markers, as well as shape markers obtained from Phase 3 CECT, demonstrating an accuracy range of 54% to 92% and an AUC of 0.45 to 0.85 using XGBoost and RF classification models. Finally, Chen et al. [71] promoted second-order texture markers of Phase 3 CECT and used LR to obtain an accuracy range of 82% to 88% and an AUC range of 0.86 to 0.90.
For RCC grading and staging purposes, Feng et al. [37] found that first-order texture markers such as entropy, STD, and kurtosis extracted from CECT are statistically significant. They reported an accuracy ranging from 70% to 79% and an AUC between 0.74 and 0.83. Shu et al. [38] found that first-and second-order texture markers, along with shape markers derived from CECT phases 2 and 3, could serve as valuable radiomic markers. Using the LR classification model, they yielded an accuracy of 72% to 78% and an AUC of 0.77 to 0.82. The same group extended their study by including a slightly larger cohort, excluding shape markers, and replacing the LR classifier with SVM and RF classifiers, resulting in an enhanced diagnostic accuracy of 92% to 94% and an AUC of 0.96 to 0.98. Two studies [72,73] focused on extracting second-order texture markers from CECT phases 2 and 3.
Ding et al. [72] achieved an AUC ≥ 0.67 after using LR classifiers, whereas Yin et al. [73] used SVM instead and yielded an enhanced AUC of 0.86. Second-and higher-order textures of Phase 3 CECT were reported particularly useful by Bektas et al. [74]. They relied on the evidence of achieving 85% accuracy and an AUC of 0.86 upon employing SVM classifiers. Lin et al. [75] determined that multiphasic CECT first-and second-order texture markers are useful in identifying renal tumors with 74% accuracy and an AUC of 0.87 using gradient boosting decision tree classifiers. Momenian et al. [76] posited that first-order texture markers in Phase 2 CECT have the potential to grade ccRCC tumors, achieving 97% accuracy using RF classifiers. Lai et al. [77] reported that first-order texture markers and shape markers of unenhanced CT can sufficiently classify ccRCC tumors using a Bagging classifier, resulting in an AUC of 0.75.
Luo et al. [78] achieved 81% accuracy and an AUC of 0.87 using RF classifiers on the derived first-order texture markers and shape markers from CECT phases 1 and 4, while first-, second-, and higher-order texture markers obtained from unenhanced CT were suggested by Yi et al. [79] to sufficiently grade ccRCCs, using an SVM classification model, with 90% accuracy and an AUC of 0.91. In line with Yi et al. [79], He et al. [80] agreed on the marker types, while disagreeing on the CECT phases from which they should be derived, instead recommending phases 2 and 3 of CECT and attaining an accuracy range of 91% to 94% using ANNs. Xu et al. [81] utilized an ensemble of various DL networks on 2D regions of interest (ROIs) of Phase 2 CECT and achieved 82% accuracy and an AUC of 0.88. Demirjian et al. [39] carried out a comprehensive investigation focused on both the grading and staging of ccRCC tumors. For grading, they employed multiple secondorder texture markers in conjunction with the mean intensity, which served as a first-order texture marker, extracted from CECT. For the staging process, they relied exclusively on second-order texture markers. In both cases, they utilized RF classification models and attained AUC values of 0.73 and 0.77 for grading and staging, respectively. Table 1 summarizes the above-mentioned AI-based CAD systems from the last decade that utilized CECT imaging in terms of the following attributes: study, main goal, data, radiomics, methods, results, and findings. Studies with the same main goal are grouped together for comparison purposes.
To sum up, the AI-based CAD systems that utilized CECT images demonstrated promising findings in the early diagnosis of RCC. These systems have effectively differentiated malignant from benign tumors with an accuracy range of 41% to 98% and an AUC range of 0.49 to 0.97, classified RCC tumor subtypes with an accuracy range of 47% to 92% and an AUC range of 0.49 to 0.92, and graded and staged RCC tumors with an accuracy range of 70% to 97% and an AUC range of 0.67 to 0.98. Entropy, a first-order texture marker, has frequently been identified as a crucial radiomic marker extractable from multiphasic CECT. Phases 2 and 3, namely the arterial phase/corticomedullary phase and portal venous/nephrographic phase, have been the most commonly used and recommended. Furthermore, machine learning classifiers such as LR, RF, SVM, and ANN have yielded the best classification results. While CECT has proven sufficient in RCC diagnosis, it is not the preferred modality when radiation exposure is contraindicated (e.g., in pregnant or pediatric patients). This has prompted researchers to explore the capabilities of alternative imaging modalities, such as MRI, to avoid radiation exposure whenever possible. Our search within the last decade revealed a limited number of studies on this topic, which we discuss in detail below.

Magnetic Resonance Imaging (MRI) Studies
In the differentiation of benign and malignant renal tumors, Xu et al. [29] investigated the potential of DL and ML using T2-weighted MRI and DW-MRI. Their study included a total of 217 patients with renal tumors, allocating 173 patients to the training set and 44 patients to the testing set. Following manual identification of ROIs, the investigators used three distinct DL ResNet-18 models and three separate handcrafted-based RF models, incorporating a total of 96 radiomic markers. The first model used T2-weighted imaging, the second model used DW-MRI, and the third model combined both modalities. The ResNet-18 models demonstrated accuracies of 77%, 80%, and 81.3%, while the handcrafted RF models attained accuracies of 77%, 71%, and 82%. Oostenburgge et al. [30] conducted a study to evaluate texture markers derived from 3D ADC maps of DW-MR images for distinguishing benign ONC from malignant RCC. The dataset comprised 39 renal tumors, including 32 RCCs and 7 ONCs. The authors found that entropy, STD, tumor volume, and gender demonstrated statistical significance among the different tumor groups. By integrating these markers, they achieved an AUC of 0.91 with 86% sensitivity and 84% specificity using the LR classification model. Furthermore, they discovered that entropy and the 25th percentile were statistically significant when comparing healthy cortical regions with tumor tissue.   Li et al. [31] included 92 DW-MRI renal tumors, with malignant tumors including (ccRCC, N = 38), (paRCC, N = 16), and (chrRCC, N = 18) and benign tumors comprising (AML, N = 13) and (ONC, N = 7). The authors constructed 3D ADC maps and calculated 10 distinct first-order texture markers. Following statistical analysis to identify significant markers, they evaluated diagnostic performance using ROC analysis. They reported that the mean, median, 75th percentile, 90th percentile, STD, and ADC entropy of malignant tumors were significantly higher than those of benign tumors. They reported 80% sensitivity, 86.1% specificity, and an AUC of 0.85%. Razik et al. [23]  Nikpanah et al. [92] explored the potential of deep CNNs along with T2-weighted MRI and multiphasic CEMRI for differentiating ccRCC from ONC. Their study included 74 patients with a total of 243 renal masses, comprising 203 ccRCC and 40 ONC tumors. The researchers placed 2D ROIs around the tumors and input them into an AlexNet CNN model, achieving 91% accuracy and an AUC of 0.9. Arita et al. [93] analyzed the texture of 3D ADC maps derived from DW-MRI to distinguish between benign AML and malignant nccRCC. They encompassed a training dataset of 67 tumors (AML = 46 and nccRCC = 21) and a validation dataset of 39 tumors (AML = 24 and nccRCC = 15). A total of 45 texture markers were extracted, and the long-zone high gray-level emphasis, as a second-order texture marker, was reported as the most dominant marker for identifying AML. Their RF classification model yielded an AUC of 0.82, which was comparable to the radiologic assessment.
Gunduz et al. [94] used texture analysis of ADC maps for distinguishing benign ONC from malignant chrRCC in a small cohort of 14 patients (ONC = 6 and chrRCC = 8). The study identified six texture markers, with five being second-order (run variance, shortrun emphasis, normalized run-length nonuniformity, run percentage, long-run emphasis) and one being first-order (square root of mean ADC). They achieved 87.5% sensitivity and 83% specificity using ROC analysis. Matsumoto et al. [32] explored texture analysis on DW-MRI for differentiating between AMLs and ccRCCs. Their study consisted of two datasets. The first dataset comprised 83 tumors (AML = 18 and ccRCC = 65) that were used for the development of the diagnostic model, while, the second dataset included 39 tumors (AML = 13 and ccRCC = 17), serving as external validation. From the ADC maps, they extracted 39 texture markers and employed an RF model to determine the importance of these markers. They identified the mean ADC value as a significant firstorder texture marker and both long-run low gray-level enhancement and gray-level run emphasis as dominant second-order texture markers in the diagnostic process, achieving an AUC of 0.87.
For RCC subtyping and grading, Goyal et al. [22] examined the power of texture markers derived from multiparametric MRI techniques, including T1-weighted MRI, T2weighted MRI, multiphasic CEMRI, and DW-MRI. Using a total of 34 renal masses, consisting of 29 ccRCCs (low-grade = 19, high-grade = 10) and 5 nccRCCs, they placed 2D ROIs on the maximum viable tumor area. Then, they extracted multiple first-order texture markers, including mean, entropy, STD, skewness, MPP, and kurtosis, from each MRI sequence for further investigation. The authors suggested that various first-order texture markers derived from multiparametric MRIs could serve as valuable diagnostic tools for both subtyping and grading renal tumors. Sun et al. [40] explored the power of texture analysis on susceptibility-weighted magnetic resonance imaging (SW-MRI) to grade ccRCCs. The study encompassed 45 patients, comprising 29 low-grade and 16 high-grade ccRCC tumors. The total number of derived texture markers was reduced from 396 to 10. Using multivariable logistic regression, the authors constructed a diagnostic model which produced 77.3% accuracy, 80.5% sensitivity, and 71.4% specificity.
Chen et al. [41] aimed at grading ccRCC using Phase 2 CEMRI. Their study included 99 tumors, with 61 low-grade and 38 high-grade cases. They placed 2D ROIs, then extracted and analyzed various first-, second-, and higher-order texture markers. Using RF importance analysis, six texture markers, namely entropy, sum of entropy, kurtosis, horizontal gray-level nonuniformity, gray-level nonuniformity, and run-length nonuniformity, were selected. Subsequently, they achieved 86.2% accuracy, 72.7% sensitivity, 94.4% specificity, and an AUC of 0.76 on the validation dataset (N = 29) using a modeled MLP-ANN classifier. Despite high specificity, the sensitivity was relatively low, which could be attributed to class imbalance. The utility of various radiomic markers, including shape markers and first-and second-order texture markers, extracted from T2-weighted and multiphasic CEMRI was explored by Choi et al. [95] to grade ccRCC. Their study encompassed 364 renal tumors, of which 272 were low-grade and 92 were high-grade. Their RF classification model demonstrated 98% accuracy, 72% sensitivity, 95% specificity, and an AUC of 0.89. Although the overall diagnostic performance was satisfactory, the relatively low sensitivity was likely attributable to data imbalance.
Hoang et al. [96] explored diagnosing renal RCC using texture analysis of multiphasic MRI. Their study involved 212 renal lesions, of which 96 were normal, 11 were ONC, 87 were cRCC, and 8 were paRCC. These lesions were divided into two halves for training and validation purposes. Following the placement of 2D ROIs, first-order texture markers, including mean, skewness, STD, and kurtosis, were extracted. Using an RF classifier among all phases, Phase 1 CEMRI demonstrated the best diagnostic accuracy of 79.1%. However, integrating texture markers from different phases raised the final diagnostic accuracy to 83.7%. In a following study by the same researchers [33], the utility of multiphasic CEMRI for differentiating benign from malignant renal tumors and distinguishing major subtypes of RCCs was investigated. The study included 140 renal lesions, of which 30 were ONC, 90 were RCC, and 22 were paRCC. After placing 2D ROIs on the slices encompassing the largest cross-section in each contrast phase, multiple first-and second-order texture markers were extracted using histogram analysis, gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size-zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM). Least absolute shrinkage and selection operator (LASSO) regression was then applied to select the optimal markers for classification. The study concluded that first-order texture markers were useful to identify malignancy, while adding second-order texture markers improved the accuracy of subtyping. In terms of classification, they used RF classification models and reported 77.9% accuracy in distinguishing between paRCC and ccRCC, 79.3% accuracy in differentiating between ONC and ccRCC, and 77.9% accuracy in discriminating between ONC and paRCC. Table 2 summarizes the above-mentioned AI-based CAD systems from the last decade that utilized different MRI modalities in terms of the following attributes: study, main goal, data, radiomics, methods, results, and findings. Studies with the same main goal are grouped together for comparison purposes.   To sum up, the AI-based CAD systems that utilized various types of MRIs demonstrated interesting results and findings in the early diagnosis of RCC. These systems achieved an accuracy range of 77% to 91% and an AUC range of 0.82 to 0.91 for differentiating malignant from benign tumors. Furthermore, they attained an accuracy range of 77% to 98% and an AUC range of 0.76 to 0.89 for subtyping and/or grading RCC tumors. First-order texture markers such as entropy, MPP, mean, skewness, and kurtosis have been frequently identified as the most dominant and important radiomic markers derived from multiparametric MRIs. These markers are useful for differentiating between benign and malignant renal tumors. The addition of second-order texture markers derived from GLRLM has also proven valuable. Notably, texture analysis of ADCs derived from DW-MRI was the most commonly used technique among the reviewed MRI studies. Additionally, RF classifiers were chosen by the majority of these studies, yielding superior classification results. In spite of MRIs being useful for identifying malignancy status, subtyping RCCs, and grading RCCs, there is a lack of research investigating the staging of RCCs. Staging is crucial for determining a tumor's spread, size, and location, making it a vital area for future investigation.

AI-Based Prediction Of Clinical Outcome/Treatment Response Studies
A limited number of studies have explored the role of AI and/or radiomics derived from CT and/or MRI in predicting clinical outcomes and treatment responses in RCC patients. Focusing on MRI studies, Bharwani et al. [43] investigated the potential correlation between various radiomic markers extracted from DW-MRI and dynamic contrast-enhanced MRI (DCE-MRI) and the response to neoadjuvant sunitinib therapy, specifically overall survival (OS), in metastatic RCC (mRCC). Their study included 20 mRCC patients who survived after completing three treatment cycles. By placing 3D ROIs on DW-MR images, they calculated tumor volume, constructed ADC maps, and generated ADC histograms. They then determined mean ADC, AUC low (ADC 25th percentile), kurtosis, and skewness before and after treatment. Using 2D ROIs on DCE-MR images, they computed the maximum signal intensity and the wash-in rate before and after treatment. Using the Kaplan-Meier (KM) method, they analyzed OS by dividing mRCC patients into two groups based on the median of the aforementioned descriptive statistics as a cutoff. Their findings revealed that patients with a tumor volume below the median at baseline experienced a prolonged OS. An increase in AUC low of ADCs greater than the median was indicative of reduced OS, whereas a decrease in AUC low suggested a prolonged OS. Moreover, a positive correlation was found between mean ADC in the primary tumor and metastases.
Antunes et al. [44] conducted a study aiming to identify the optimal radiomic markers on an integrated positron emission tomography (PET)/MRI that best describe early treatment responses and changes in advanced mRCC patients (N = 2) undergoing sunitinib therapy. They extracted a total of 66 radiomic markers, including raw T2w signal, postprocessed T2w, 30 postprocessed T2w textures, raw ADC map, 30 ADC textures, standard uptake value (SUV), and 2 PET textures. Subsequently, they employed a scoring function to determine the top 25 ranked radiomic markers in the two patients under study. They found that SUV from PET, T2w difference average from T2w, and ADC energy from DW-MRI ADC maps were ranked the highest among the 25 radiomic markers in terms of reproducibility and capturing treatment-related changes or responses. Furthermore, the integration of these radiomic markers resulted in improved prediction performance. However, they acknowledged that their findings are limited due to a small sample size of only two patients and emphasized the need for further investigation in a larger cohort to validate their results.
Reynolds et al. [51] conducted a study to investigate the potential of radiomic markers derived from DW-MRI (N = 12) and DCE-MRI (N = 10) as predictors of early treatment response in RCC patients following stereotactic ablative body radiotherapy (SABR). For shape markers, 3D ROIs from CT images were utilized to contour the tumor, and tumor volume was calculated at baseline and after three different follow-up scans to estimate tumor volume change. For textural markers, ADC maps were derived from 3D ROIs of DW-MR images, and an ADC histogram was constructed. Mean ADC, median, kurtosis, and skewness were subsequently calculated. Employing DCE-MR images, various functional markers were estimated, including mean T onset , mean IRE, mean MaxE, mean T wout, mean IRW, mean K trans , iAUCAC60, % washout voxels, % plateau voxels, % persistent voxels, and % nonenhancing voxels. Spearman rank correlation coefficients (ρ) were computed to compare changes in the aforementioned parameters against the % change in tumor volume. Statistically significant correlations were observed between the change in percentage washout, change in mean IRE, and mean K trans and the change in tumor volume (p < 0.05). Changes in ADC kurtosis also demonstrated statistically significant positive correlations with % tumor volume change (p < 0.05).
For CT studies, Lubner et al. [50] conducted a study to identify radiomic texture markers that can be extracted from phases 1 and 3 of CECT images in RCC patients (N = 157) and might be correlated with histological findings and treatment response. From 2D ROIs, and after applying various texture filters, they extracted six different first-order texture markers: mean, STD, MPP, entropy, skewness, and kurtosis. Their study found that entropy, STD, and MPP were correlated with histologic type, nuclear grade, and clinical outcomes (time to recurrence and OS) in patients with RCC.
Boos et al. [45] assessed the ability of mean and median intensity attenuation, represented by Hounsfield units (HU) and estimated from CECT images, in predicting treatment responses (response, stable, and progression) in patients with RCC tumors (N = 19) who received targeted therapy, specifically vascular endothelial growth factor receptor (VEGFR) tyrosine kinase inhibitors (TKI). After estimating the mean and median HUs, they performed the Wilcoxon signed-rank test to compare the change between the baseline and consecutive post-treatment scans for overall outcome assessment. They concluded that the median HU attenuation shift provided better prediction accuracy (79%) than the mean (74%) and thus is preferable. Moreover, a shift in median <-44 HU indicated a partial response, while a shift in median >-41 HU indicated progression, and therefore, median HU shift correlates well with clinical outcomes in mRCC patients.
Haider et al. [46] conducted a study to highlight potential radiomic predictors of PFS and OS that could be extracted from CECT images in RCC patients (N = 40) undergoing treatment with sunitinib. After placing 2D ROIs on renal tumors, they extracted various first-order texture markers, such as MPP, STD, skewness, kurtosis, and size-normalized STD (nSTD), and entropy as a second-order texture marker. A Cox proportional hazards survival statistical analysis was employed to determine the predictors of both PFS and OS. Their study revealed that nSTD extracted at baseline and after treatment is positively correlated with both OS and PFS, while entropy and tumor size changes are predictors of OS but not PFS.
Mains et al. [47] investigated radiomic functional markers derived from CECT images that could potentially predict OS and PFS in mRCC patients (N = 69). After placing 2D ROIs, they identified seven markers that describe functionality, specifically, blood volume using deconvolution (BV deconv ), blood flow using deconvolution (BF deconv ), standardized perfusion values using deconvolution (SPV deconv ), blood volume using maximum slope (BV max ), standardized perfusion values maximum slope (SPV deconv ), blood volume using the Patlak model (BV patlak ), and permeability surface area product using the Patlak model (PS). They applied various statistical analysis methods on the histogram data of the aforementioned markers and found that medians and modes of BV deconv , BV patlak , and BF deconv are statistically significant (p < 0.05) and have the strongest correlation with clinical outcomes (PFS and OS).
Khodabakhshi et al. [48] conducted a study to investigate possible radiomic markers extracted from Phase 2 CECT along with clinical biomarkers for the prediction of OS in RCC patients after partial or radical nephrectomy (N = 210). The 2D ROIs were manually drawn, and a total of 225 radiomic markers were extracted, including 29 shape markers, 50 firstorder texture markers, and 136 second-order texture markers. Additionally, 59 clinical markers were included in the analysis. They employed Cox proportional hazards regression as a marker selection method, which resulted in a reduced set of 11 radiomic markers and 12 clinical markers. Then, they applied the accelerated failure time technique to specify the contribution of the selected markers on OS time. Their study revealed that flatness, area density, and median were the most significant radiomic markers (p < 0.05), while tumor heterogeneity, grade, and stage were the most significant clinical markers (p < 0.05). Therefore, all of these markers combined were significant predictors for OS in RCC patients.
Zhang et al. [49] investigated the prediction potential of radiomic markers extracted from CECT images and clinical markers linked to PFS after partial or radical nephrectomy in ccRCC (N = 175). After manual segmentation of tumors using 3D ROIs, they extracted a total of 428 radiomic markers (107 per CT phase). They then applied the least absolute shrinkage and selection operator with 5-fold cross-validation (LassoCV) to select the dominant markers, resulting in six markers (four shape-based markers and two secondorder texture markers) as follows: least axis length (Phase 2), maximum 2D diameter row (Phase 4), surface volume ratio (Phase 1), maximum 2D diameter slice (Phase 3), size-zone nonuniformity (Phase 2), and complexity (Phase 2). Subsequently, they established a multivariate Cox regression model using a training dataset (N = 125) for PFS prediction and saved the other 50 subjects for validation. This model depended on a weighted sum of the selected markers. In addition, they integrated statistically significant clinical markers (age, clinical stage, and Karnofsky performance status (KPS) score), resulting in a PFS prediction model encompassing both clinical and radiomic markers. After validating their model on the validation dataset (N = 50), they achieved an accuracy of 70%, sensitivity of 58%, specificity of 74%, and an AUC of 0.71. They concluded that radiomic-based markers extracted from CECT, especially Phase 2, demonstrated better prediction performance of PFS in ccRCC patients when combined with clinical markers. Table 3 provides a summary of the aforementioned radiomic-based CAP systems developed in the last decade using different CT and/or MRI modalities. The Table includes the following details: study, main goal, radiomics, methods, results, and findings.
In summary, these few studies investigated the potential of developing radiomicsbased CAP systems utilizing various types of CT and MRI scans, showing promising results in early prediction of treatment response, including overall survival rate, progression-free survival, and time to recurrence. Most of these studies relied on statistical analyses to identify statistically significant radiomic markers correlated with specific clinical outcomes or treatment responses. Histogram measures of ADC maps extracted from DW-MR images, particularly changes in mean ADC, ADC energy, and ADC kurtosis, were significant predictors of clinical outcomes (p < 0.05) [43,44,51]. Additionally, changes in radiomicbased functional parameters extracted from DCE-MR images, namely changes in percentage washout, mean IRE, and mean K trans , demonstrated significant correlations with changes in tumor volume (p < 0.05) and thus are potential indicators for clinical outcomes or treatment responses [51]. First-order texture markers, specifically entropy, STD, and MPP, are correlated with time to recurrence and OS [46,50], while nSTD is positively correlated with both OS and PFS [46]. Functional-based radiomic markers, such as median HU, are potential predictors of partial response and progression [45], while medians and modes of BV deconv , BV patlak, and BF deconv are strong predictors of OS and PFS. Moreover, shapebased radiomic markers, namely flatness and area density, extracted from Phase 2 CECT are strong predictors of OS [48], while least axis length extracted from the same phase of CECT is a potential indicator of PFS [49]. Some clinical markers can be combined with radiomic markers for enhanced prediction performance of PFS, such as age, stage, and KPS score [49], and for improved prediction of OS, such as grade and stage [48]. Despite these promising findings, there is considerable heterogeneity and diversity in the number of patients included in each study, the type of treatment administered before or after nephrectomy (e.g., radiation therapy, targeted therapy, neoadjuvant therapy, etc.), and the final goals or endpoints of these studies (e.g., type of clinical outcome). Furthermore, most of these studies were primarily statistical-based in nature and were not aimed at the development of a comprehensive automated AI-based CAP system. Table 3. Summary of the last decade's CT-and/or MRI-based studies for predicting/assessing patient outcomes (e.g., treatment response, recurrence, and overall survival (OS), and progression-free survival (PFS)).

Study
Main Goal Radiomics Methods

Results Findings
Bharwani et al. [43] To A positive correlation between mean ADC was found between the primary tumor and metastases Antunes et al. [44] To find the optimal radiomic markers on an integrated positron emission tomography (PET)/MRI that best describe early treatment response/changes in advanced mRCC undergoing sunitinib therapy (N = 2)

Discussion and Future Directions
The success of accurate and timely diagnosis of renal tumor malignancy status, specific subtype, and associated grade (I-IV) and stage (I-IV) holds significant clinical importance, as it directly influences the determination of appropriate treatment and management plans. As a result, precise prediction of clinical outcomes or treatment responses, including recurrence rate, overall survival rate, and progression-free survival rate, is essential to avoid the burden of unnecessary treatment strategies. This, in turn, can conserve time, effort, and resources, ultimately benefiting both patients and healthcare providers. Suggested Diagnostic Radiomic Markers:  [68]. In this context, both imaging modalities can be utilized for subtyping and grading purposes. However, MRIs are preferable in cases involving pediatric patients or pregnant women [97] to prevent exposure to ionizing radiation. For staging, a few CT studies demonstrated promising diagnostic performance [39], while MRI studies did not investigate radiological staging.

Suggested Prediction Radiomic Markers:
In terms of treatment response prediction, entropy, mean, skewness, kurtosis, STD, and median have been identified by most CT studies [46,50] as potential radiomic markers for predicting OS and PFS. On the other hand, histogram measures of ADC maps extracted from DW-MR images, specifically changes in mean ADC, ADC energy, and ADC kurtosis, were the most promising predictors of clinical outcome in MRI studies [43,44,51]. To the best of our knowledge, no studies have employed AI, ML, or DL for the purpose of predicting treatment response; rather, they have relied on statistical analyses to identify significant markers correlated with clinical outcome/treatment response. A limited number of studies have depended on morpho-logical or functional markers, which, if integrated, could significantly enhance clinical outcome/treatment response prediction [43,46,48,49,51]. It is worth noting that one study [50] attempted to identify optimal radiomic-based markers correlated with both histological findings and treatment responses. The authors concluded that first-order texture markers, specifically entropy, STD, and MPP extracted from phases 1 and 3 of CECT, were correlated with histological type, nuclear grade, and clinical outcomes (time to recurrence and OS) in patients with RCC. However, this study did not attempt to incorporate these radiomic markers into a comprehensive AI-based CAD/CAP system capable of simultaneously diagnosing RC and predicting treatment response, which is the ultimate goal of this survey. Consequently, the research gap remains and warrants further investigation. Future Directions: While renal cancer diagnosis is a well-established research area, with numerous CT and MRI studies having developed radiomic and AI-based CAD systems for determining malignancy status, subtyping, grading, and staging, some investigations still suffer from low sensitivity or specificity [36,38,40,53,58,60,66,78,82,87,89,93,96]. Consequently, integrating radiomic markers extracted from multiple imaging modalities, such as CT and MRI, may improve diagnostic performance. Furthermore, as radiological-based analysis may not be sufficient for predicting clinical outcome/treatment responses, incorporating histopathological image analysis that captures characteristics such as cell color, shape, size, and staining could enhance prediction capabilities. Identifying robust AI models may reduce subjectivity by pinpointing optimal markers for treatment response prediction purposes. It is worth noting that a new trend in predicting treatment response using radiogenomics has recently emerged in a few studies and requires further investigation [98][99][100][101][102][103].
In conclusion, more investigative studies are still ongoing for both CT and MRI for the purpose of diagnosing renal tumors, as well as predicting clinical outcome/treatment responses for optimal management plans. Progress in the early diagnosis of renal tumors and treatment response prediction depends mainly on the identification of optimal discriminating markers for the intended diagnostic/prediction problems, as well as the development of robust, reproducible, and generalizable AI-based diagnosis/prediction models. By providing these future directions and suggestions, we aim to encourage investigators and researchers to address this research gap and achieve the intended goal of establishing a comprehensive, unified CAD/CAP system that can be reliably used for both renal tumor diagnosis and clinical outcome/treatment response prediction, ultimately leading to improved healthcare outcomes.

Conflicts of Interest:
The authors declare no conflict of interest.

Abbreviations
The following summarizes the list of abbreviations that are used in this survey: