Radiomics Applications in Renal Tumor Assessment: A Comprehensive Review of the Literature

Radiomics texture analysis offers objective image information that could otherwise not be obtained by radiologists′ subjective radiological interpretation. We investigated radiomics applications in renal tumor assessment and provide a comprehensive review. A detailed search of original articles was performed using the PubMed-MEDLINE database until 20 March 2020 to identify English literature relevant to radiomics applications in renal tumor assessment. In total, 42 articles were included in the analysis and divided into four main categories: renal mass differentiation, nuclear grade prediction, gene expression-based molecular signatures, and patient outcome prediction. The main area of research involves accurately differentiating benign and malignant renal masses, specifically between renal cell carcinoma (RCC) subtypes and from angiomyolipoma without visible fat and oncocytoma. Nuclear grade prediction may enhance proper patient selection for risk-stratified treatment. Radiomics-predicted gene mutations may serve as surrogate biomarkers for high-risk disease, while predicting patients’ responses to targeted therapies and their outcomes will help develop personalized treatment algorithms. Studies generally reported the superiority of radiomics over expert radiological interpretation. Radiomics provides an alternative to subjective image interpretation for improving renal tumor diagnostic accuracy. Further incorporation of clinical and imaging data into radiomics algorithms will augment tumor prediction accuracy and enhance individualized medicine.


Introduction
In 2018, renal cell carcinoma (RCC) accounted for 403,300 newly diagnosed cancer cases and 175,100 deaths worldwide [1]. In the United States alone, RCC is the sixth most common cancer in men and eighth in women, accounting for 5% and 3% of all newly diagnosed cases annually, respectively [2]. An increasing incidental detection of renal masses with cross-sectional imaging led to the diagnosis of more asymptomatic, small, and clinically localized renal masses. Small renal masses (SRMs), defined as ≤4 cm in diameter, account for more than 50% of all renal masses, approximately 10-30% of which result in benign histology [3,4]. The diagnosis of SRMs carries the risk of subjecting patients to unnecessary procedures and overtreating lesions that may not progress.
Renal tumor biopsy (RTB) provides a means for tissue sampling to assist in tumor histological and subtype diagnosis for risk-stratified management [4]. Although it shows high diagnostic accuracy for RCC, RTB is an invasive procedure and it is criticized for its inability to sample tumors at multiple sites and distinguish tumor histologic subtypes and nuclear grade [5]. Due to its improved soft-tissue contrast, magnetic resonance imaging (MRI) outperforms computed tomography (CT) in the evaluation of indeterminate renal masses [6], local invasion, and intravascular extension [7]. However, the differentiation of benign lesions, particularly oncocytoma and angiomyolipoma without visible fat (AMLwvf), from RCC can be challenging due to subjective radiological image interpretation [8]. Radiomics is a term that encompasses various techniques for the extraction of quantitative features from medical images to improve diagnostic, prognostic, and predictive image interpretation accuracy. It is essentially the conversion of images into metadata for subsequent mining to improve clinical decision-making algorithms [9]. It assists physicians in identifying complex image patterns that are not visible to the naked eye by using artificial intelligence (AI) methods [10]. Recent advancements in AI, specifically in machine and deep learning, accelerated the application of radiomics to medical imaging as a new beacon to guide clinical decisions.

Study Aims
The primary objective was to perform a comprehensive review of the literature on current radiomics applications in renal mass assessment. The secondary objective was to provide a narrative summary of articles evaluating the accuracy of radiomics for distinguishing benign and malignant tumors, predicting nuclear grade, obtaining gene expression-based biomarkers, and prognosticating RCC patients.

Literature Search
A comprehensive search of English language literature was conducted using the PubMed-MEDLINE database up to 20 March 2020. To capture recent trends in radiomics applications, the search was limited to articles published within the last five years. The search strategy included the following broad terms in isolation or combination: "kidney cancer", "renal cell carcinoma", "machine learning", "deep learning", and "radiomics". We repeated searches on all newly identified articles until no further relevant articles were found. Titles and abstracts of articles that were identified by the keyword search were retrieved for full-text evaluation. Potentially eligible articles were independently screened against the study selection criteria by two authors (R.S. and M.B.).

Inclusion and Exclusion Criteria
Two authors (R.S. and M.B.) individually determined inclusion/exclusion of all articles retrieved in full text, and discrepancies were resolved through discussion by a third reviewer (A.M.). Studies that met the following criteria were included: (a) renal tumor radiomics-based analysis; (b) articles written in English; (c) peer-reviewed publications; (d) methodology documented in replicable detail. After article selection and according to the eligibility criteria, the following types of studies were excluded: articles not related to renal tumors, articles not written in English, review articles, editorials, and replies from author. In accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) criteria, Figure 1 was included to delineate our article selection process. After full-text evaluation, data were independently extracted by the authors for further assessment of qualitative and quantitative evidence synthesis.

Data Extraction and Quality Assessment
Data extraction from each study was independently conducted by two authors (R.S. and M.B.). As proposed by the PRISMA guidelines, we used the population, intervention, comparator, outcomes, and study design approach to specify the eligibility criteria. The studies were considered eligible if patients with renal masses (population) were evaluated with CT/MRI/positron-emission tomography (PET)-based radiomics data (intervention), compared to radiologists' subjective image assessment (comparator) or as a single-arm study group, to investigate the accuracy for benign and malignant tumor differentiation, to predict nuclear grade, identify molecular biomarkers, and determine patient outcomes (outcome). After full-text evaluation, data were independently extracted by the authors for further assessment of qualitative and quantitative evidence synthesis. The following information was extracted from each study: name of author, journal and year of publication, imaging method, number of patients per study, mean age, mean lesion diameter, radiomics method, texture features extracted, and outcome prediction accuracy.

Characteristics of Included Studies
A total of 94 articles were identified from the search query. Overall, 36 reports were excluded after title and abstract review, for the following reasons: 10 not related to urology, 16 not related to

Data Extraction and Quality Assessment
Data extraction from each study was independently conducted by two authors (R.S. and M.B.). As proposed by the PRISMA guidelines, we used the population, intervention, comparator, outcomes, and study design approach to specify the eligibility criteria. The studies were considered eligible if patients with renal masses (population) were evaluated with CT/MRI/positron-emission tomography (PET)-based radiomics data (intervention), compared to radiologists' subjective image assessment (comparator) or as a single-arm study group, to investigate the accuracy for benign and malignant tumor differentiation, to predict nuclear grade, identify molecular biomarkers, and determine patient outcomes (outcome). After full-text evaluation, data were independently extracted by the authors for further assessment of qualitative and quantitative evidence synthesis. The following information was extracted from each study: name of author, journal and year of publication, imaging method, number of patients per study, mean age, mean lesion diameter, radiomics method, texture features extracted, and outcome prediction accuracy.

Characteristics of Included Studies
A total of 94 articles were identified from the search query. Overall, 36 reports were excluded after title and abstract review, for the following reasons: 10 not related to urology, 16 not related to renal cell carcinoma, five review articles, four articles not in English, and one reply to authors. After full-text evaluation, 19 articles were excluded because they did not meet the study´s objectives. Finally, we carefully read the full text of each of the remaining 39 articles which were included in the review. The flow chart of the study selection process is shown in Figure 1. The 39 included studies were divided into four main categories according to study objectives: 19 for tumor differentiation, 11 for nuclear grade prediction, six for gene expression-based molecular biomarkers, and three for patient outcome prediction. Tumor differentiation was further subclassified into four for malignant vs. benign renal mass diagnosis, seven for RCC vs. angiomyolipoma without visible fat, four for RCC vs. oncocytoma, and four for RCC subtypes.

Renal Mass Differentiation
The adoption and widespread use of cross-sectional imaging increased the detection rate of renal masses [11]. However, the diagnostic migration toward early disease stages was not translated into improved cancer-specific mortality for localized RCC, indicating that many incidental renal masses are unlikely to progress and pose a threat to patients in the long term [12]. These findings contributed to the growing awareness in the overdiagnosis and overtreatment of these tumors.
The decision whether to treat patients with renal masses is often made without a definitive histologic diagnosis. Therefore, treatment decision-making and management are heavily reliant on clinical imaging. Although RTB can be used to differentiate malignant from benign histology, its role is selective and limited in its ability to discern between aggressive and indolent disease. Several studies showed an inverse correlation between renal mass radiographic size and the probability of harboring benign or low-grade pathology in surgical specimens [13]. Although the likelihood of malignancy increases with mass size, current imaging modalities and subjective radiological image interpretation cannot reliably distinguish malignant tumors from certain benign tumors, such as oncocytomas and AMLwvf [14].
Studies showed that renal masses including RCC subtypes, AMLwvf, and oncocytomas can be portrayed with distinct gray-level imaging patterns that are traceable by radiomic analysis [15]. Radiomics includes a number of approaches designed to convert medical images to quantitative, minable, and high-dimensional data [3]. Machine (ML) and deep learning (DL) algorithms are used to automatically extract and analyze histogram, texture, and shape information from imaging data which may not be evident to the naked eye. Given the limitations of conventional medical imaging, there is increased interest to apply radiomics in oncological imaging as a tool to obtain diagnostic, predictive, and prognostic information from routine clinical imaging [16]. However, despite its extensive use in research and favorable results linking CT/MRI texture features to renal mass characterization, the routine use of radiomics in clinical practice is yet to be seen. For imaging markers, including texture-based metrics, to bridge the translational gap between an experimental research tool and a clinically applicable diagnostic algorithm, its technical and biological validity, biological validity, qualification, and cost-effectiveness need to firstly be established [3]. and was identified as a promising CDS tool for renal mass identification [17]. Similarly, Erdim et al. compared eight ML algorithms to construct a prediction model for renal mass diagnosis based on CECT imaging from both benign and malignant lesions. The random forest model showed the highest accuracy (90.5%) and AUC (0.92) and was a suitable method to distinguish benign and solid renal masses in nine-tenths of patients [18]. Sun et al. compared the performance of radiologic radiomic ML models and expert-level radiologists to differentiate benign from malignant solid renal masses using CECT examinations. The radiomic ML model yielded overall higher performance values in terms of sensitivity and specificity for differentiating clear cell renal cell carcinoma (ccRCC) from papillary RCC (pRCC) and chromophobe RCC (chRCC), ccRCC from AMLwvf and oncocytoma, and pRCC and chRCC from AMLwvf and oncocytoma [15]. Similarly, Xi et al. developed a DL model based on MRI data and compared it to four expert radiologists to distinguish benign renal tumors from RCC. The DL model showed overall higher accuracy, sensitivity, and specificity and was comparable to expert diagnostic opinion [19]. Table 1 summarizes articles investigating the differentiation of benign from malignant lesions using radiomics strategies.

Angiomyolipoma (AMLwvf) vs. RCC Subtypes
Angiomyolipoma (AML), the most common benign solid renal tumor, is composed of varying amounts of dysmorphic blood vessels, smooth muscle components, and mature adipose tissue [35]. Its diagnosis relies on the identification of macroscopic fat that is normally easily detected by CECT or MRI. However, up to 5% of AMLs lack macroscopic fat, a condition known as AML without visible fat (AMLwvf), making it challenging to differentiate them from RCC by conventional imaging [36]. Several studies described certain imaging features that are highly suggestive of AMLwvf and analyzed them with radiomics models [20][21][22][23]25,26].
By applying artificial neural network (ANN) classifiers, Yan et al. showed that texture analysis (TA) may be a reliable quantitative strategy to differentiate between AMLwvf, ccRCC, and pRCC with an accuracy in the range of 90.7-100% based on three-phasic CECT scan images [20]. Other investigations employed similar strategies, although with ML-based TA from CECT images, and they reported higher accuracy (93.9%) and AUC (0.955) [21,24]. Moreover, Cui et al. proposed an automatic computer identification system to differentiate AMLwvf from all RCC subtypes from whole-tumor CECT images using an over-sampling technique to increase the sample volume of AMLwvf [23]. They showed that morphological interpretation by radiologists achieved overall lower performance differentiating AMLwvf from all RCC subtypes. While some studies investigated deep feature classification, by using ML and ANN-based classifiers, to improve texture features to further enhance the differentiation between AMLwvf and RCC [22], others built radiomics nomograms from these texture features, showing that they outperform other models using solely clinical factors [25]. Interestingly, Ma et al. not only showed that whole-tumor radiomics-based CECT analysis is superior to conventional CECT analysis, but also that the unenhanced phase showed higher AUC than the corticomedullar phase (CMP) and nephrographic phase (NP) [26]. These findings are encouraging for the future acquisition of imaging studies without the need of contrast agents to spare patients with renal masses the associated radiation burden of multi-phasic CECT. Table 1 summarizes studies investigating the differentiation of AMLwvf from RCC subtypes.  Whole-tumor radiomics-based CT analysis was superior to conventional CT analysis. Cyst degeneration, pseudocapsule, and sum rad-score were the most significant factors. Unenhanced phase radiomics showed higher AUC than CMP and NP groups. ML-based classifiers refer to the insertion of a new observation into the appropriate category among others that were based on trained datasets of known observations. Support vector machines (SVM) are supervised learning methods for classification that learn the optimal difference between features of each class. Random forest (RF) is a supervised learning method for a classification that is based on decision trees. AMLwvf, angiomyolipoma without visible fat; ANN, artificial neural network; AUC, area under curve; ccRCC, clear cell renal cell carcinoma; CECT, contrast-enhanced computed tomography; chRCC, chromophobe RCC; CMP, corticomedullar phase; CNN, convoluted neural network; DL, deep learning; GLCM, gray-level co-occurrence matrix; GLGM, gray-level gradient matrix; GLRL, gray-level run-length; LR, logistic regression; ML, machine learning; MRI, magnetic resonance imaging; NA, not available; NP, nephrographic phase; PET, positron-emission tomography; PPV, positive predictive value; pRCC, papillary RCC; RF, random forest; RFGB, relational functional grading boosting; ROI, regions of interest; SMV, support vector machine; SRM, small renal mass; TA, texture analysis.

Oncocytoma vs. RCC Subtypes
Renal oncocytoma is a benign renal tumor, accounting for approximately 3-7% of all renal tumors [37]. Although it is widely believed that a central stellate scar is its main characteristic, it was reported present in only 46% of cases and may also appear in 26% of chRCC [38]. Moreover, there are overlapping imaging features between ccRCC and oncocytoma making their differentiation challenging [39]. Therefore, a reliable radiomics method that is non-reliant on the presence of a central scar to characterize oncocytomas would be an invaluable diagnostic tool.
Yu et al. evaluated the utility of TA for the distinction of oncocytoma from RCC subtypes. The ML model's ability to distinguish both ccRCC and pRCC from oncocytoma was excellent with AUC of 0.93 and 0.99, respectively [27]. Li et al. explored the clinical value of radiomics-based approaches, by combining TA with five ML-based models, to differentiate oncocytoma from chRCC. They found that all five classifiers performed well at differentiating between the two with AUC values over 0.85 and concluded that their approach provides valuable preoperative diagnostic accuracy [29]. Other studies used DL approaches such as convolutional neural networks (CNN) to accurately classify chRCC and oncocytoma while also achieving 100% sensitivity in comparison with final pathology results [30]. Coy et al. investigated the diagnostic value and feasibility of a DL-based renal lesion classifier to differentiate ccRCC from oncocytoma in 179 patients with pathologically confirmed renal masses on routine four-phasic CECT [28]. When using the entire tumor volume in combination with tumor mid-slices, the excretory phase showed the best classification performance with 74.4% accuracy, 85.8% sensitivity, and a positive predictive value of 82.5%. Table 1 summarizes articles investigating the distinction between oncocytoma and RCC subtypes.

RCC Subtype Differentiation
According to cell appearance, RCCs can be largely categorized into three major subtypes: ccRCC, pRCC, and chRCC, which constitute more than 90% of all RCCs [40]. The clear cell variant is the most lethal subtype, being most likely to metastasize, whereas the pRCC and chRCC subtypes show better survival rates [41]. Therefore, RCC subtyping is clinically important due to the increased use of novel molecular targeted therapeutic agents.
Kocak et al. developed models to distinguish between the three major RCC subtypes using ML-based quantitative CECT TA. Although texture features derived from CMP performed better than those from unenhanced phase, ML-based TA was relatively poor for distinguishing between the three major RCC subtypes with overall 69.2% accuracy. However, when using CMP images with an ANN boosting algorithm, accuracy improved to 84.6% [32]. Similarly, Han et al. exploited a DL framework to distinguish between RCC subtypes using CECT images. Three-phasic input images were fed to an ANN; its performance was tested with a dataset of 169 biopsy-proven cases and showed an AUC of 0.90 regardless of subtypes [34]. Li et al. developed ML-based radiomics models with CECT images for differentiating ccRCC from non-clear-cell variants and investigated a potential radiogenomics link between imaging features and the von Hippel-Lindau gene mutation. The eight most relevant CMP features were selected to build a ML-based model which achieved a validation AUC of 0.95 and 92.5% accuracy. Five out of the eight all-relevant features were significantly associated with the von Hippel-Lindau gene mutation [31].
Yin et al. combined positron-emission tomography (PET)/MRI-based radiomics as a surrogate biomarker for intratumoral disease risk of molecular subtype ccA and ccB in patients with primary ccRCC [33]. The sparse partial least squares discriminant analysis method was applied to 168 radiomics features selected from 23 specimens of eight patients. Using radiomics features only, the correct classification rate for molecular subtype classification was 86.96%. When combining radiomics features with clinical parameters, mRNA, and microvascular density from each specimen, the best classification rate was 95.65%.

Nuclear Grade Prediction
The International Society of Urological Pathology (ISUP) and Fuhrman grading systems categorize patients' risk based on pathological features, discriminating between high-and low-risk tumors, which correlate with prognosis. The heterogeneity of tumor images quantitatively assessed by radiomics may enhance proper patient selection for surveillance avoiding RTB, and in decision-making between nephron-sparing and radical approaches. RTB carries risks of invasiveness and sampling bias; therefore, it is reasonable to search for accurate, safe, and non-invasive methods. Moreover, 16% of nuclear grade upstaging was reported on surgical pathology, resulting in an accuracy between 51.5% and 75.9% for grade discrimination [13].
Several retrospective studies evaluated the performance of radiomics for predicting RCC nuclear grade and showed satisfactory accuracy ( Table 2). Two studies assessed the accuracy of non-enhanced-CT scan images to differentiate between low-and high-grade tumors. An 85.1% accuracy was reported using ML-based TA, similar to the 81.5% accuracy reported when using ANN algorithms [42,43].

High-grade (Fuhrman 3-4) tumor detection
10 high-grade The best performance was found in Entropy   Classifiers in ML refer to the insertion of a new observation into the appropriate category among others that were based on trained datasets of known observations. Support vector machines are supervised learning methods for classification that learn the optimal difference between features of each class. Random forest is a supervised learning method for classification that is based on decision trees. CatBoost is a gradient boosting decision library based on decision trees. k-Nearest neighbor in non-parametric statistic algorithm for classification. SMOTE is synthetic minority over-sampling technique and serves as a classification model. ANN, artificial neural network; ccRCC, clear cell renal cell carcinoma; CECT, contrast enhanced compute tomography; CMP, corticomedullary phase; CT, computed tomography; GLCOM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GRLM, gray-level run-length matrix; ISUP, International Society of Urological Pathology; LASSO, the least absolute shrinkage and selection operator; LR logistic regression; ML, machine learning; MLP, multilayer perceptron; MRI, magnetic resonance imaging; NCCT: non-contrast enhanced compute tomography; NP, nephrographic phase; ROC, receiver operating characteristics; SMV, support vector machine; SSF, spatial scaling factors; TA, texture analysis.
Moreover, most studies built models with texture features retrieved from different phases of CECT scans. Models combining texture features from several phases outperformed those using single phase-based TA. Lin et al. used ML classifiers to extract texture features from every CT scan's phase (pre-contrast phase, CMP, and NP) and compared each phase alone to three-phase CT images. The model combing all phases achieved the best accuracy (74%), positive predictive value (91%), negative predictive value (59%), and AUC (0.87). The top-ranked models were reported by He et al. with a predictive mean value of 92.5% ± 1.83% using ANN-based radiomics. The best accuracy (94.1% ± 1.14%) was achieved by combining texture features from conventional image which were calculated from manually selected regions of interest (ROI), such as mean attenuation, parenchyma attenuation and absolute enhance attenuation, and CMP [45,[48][49][50][51].
MRI has the advantage of fine soft-tissue characterization while avoiding radiation exposure; therefore, Goyal et al. aimed to identify texture features from multiphasic MRI in a small cohort of patients (n = 29) and found several features that provided excellent radiomics performance, where the top three features held AUCs >0.82 [47]. Likewise, Cui et al. performed TA from multiphasic CECTs (n = 347) and MRIs (n = 93) with an external validation of patients (n = 20) who underwent both modalities. Interestingly, ML classifiers based on all-sequence MRI images (71% to 73% in internal and 64% to 74% in external validation) and all-phase CT images (77% to 79% in internal and 61% to 69% in external validation) had significant increases in accuracy [52].
The creation of risk assessment scores for personalized medicine is promising, and radiomics is a step forward in such a pathway by adding TA-based scores that might improve prediction accuracy. A predictive score beyond textures features was reported by Ding et al., integrating ML-based texture and non-texture features. Overall, a round-shape was a good discriminator of high-grade tumor (AUC: 0.723 (0.632-0.803)) similar to TA features; however, a model including both texture and non-texture features did not outperform the TA-based model's accuracy [44].
Similarly, image TA beyond the tumor was investigated by Gill et al. who compared TA from juxtatumoral perinephric fat between low-and high-grade tumors and found that the majority of features were significantly different, and the gray-level co-occurrence matrix had the best accuracy (AUC 0.746 (0.63-0.86)) [46].
The clinical utility of radiomics to discriminate between low-and high-grade tumors is yet to be determined; however, high accuracies were reported especially when using multi-phase images. MRI was not proven to be better for discrimination, but this was based only on small patient samples. Moreover, including texture feature analysis from tissues beyond the tumor and adding other quantitative data based on the standard radiological features commonly used might improve radiomics performance. Unifying radiomics to personalized medicine algorithms using clinical features and even genomic data might help build more robust prediction scores.

Gene Expression-Based Molecular Biomarkers
Predicting molecular biomarkers by radiomics entails potential clinical utility, as TA-derived scores might achieve appropriate accuracy and serve as surrogate biomarkers. Studies addressing this topic are enlisted in Table 3. Table 3. Summary of studies applying radiomics and gene expression-based models to identify molecular biomarkers.  BRCA1-associated protein (BAP1) gene mutations are found in~15% of ccRCC and are associated with high-grade tumors and poor prognosis as reported in genome-wide association studies. Therefore, Ghosh et al. tested radiomics to predict BAP1 mutation status on ccRCC and found an AUC of 0.71 for features retrieved from the NP phase of CECT scans. Similarly, Kocak et al. aimed to predict the presence of gene PBRM1 mutations by creating an ANN algorithm and ML-based TA from CMP images; they found that the former outperformed the latter, upholding 95% accuracy and 0.987 AUC for PBRM1 mutation status. This tumor suppressor gene's mutation was associated with advanced-stage and higher grade ccRCC, and it was also suggested to influence response rates to immune checkpoint inhibitors [53,54].
Angiogenesis is a largely known pathway in the pathophysiology of RCC, to such an extent that several targeted therapies used in RCC aim to halt its development. Therefore, Yin et al. tested the feasibility to predict tumor vascularity using radiomics features from PET and MRI in a small cohort (n = 9). Tumor vascularity, vascular endothelial growth factor (VEGF) expression, and microvascular density were measured from fresh frozen tissue specimens. A correlation between VEGF expression and radiomics features was not found (p = 0.539); however, the microvascular density did correlate with radiomic features, and the conjunction of image datasets from both modalities, PET/MRI, displayed the highest correlation (r = 0.639, p = 0.044). This proof-of-concept study highlights the possibility to predict pathological findings by radiomics alone, which is of utmost utility when such findings can serve as biomarkers [58].
These data can aid in the construction of nomograms that can exert an influence over clinical decision-making. Thus, some authors paved the path toward personalized medicine by assessing quantitative genomics data as predictors. Park et al. predicted T1 ccRCC aggressiveness with 85% accuracy (AUC: 0.796) by utilizing a model built on deep neural network algorithms and comprising data from FOXC2, PBRM, and BAP1 gene and protein expression [56]. Similarly, Li et al. constructed a gene panel including 15 genes (Table 3) and developed an expression-based prediction score which accurately differentiated between low-and high-risk ccRCC and predicted three-year overall survival (AUC: 0.784). In fact, the authors created a nomogram which included clinical and biochemical data, and which may prove to be a clinically indispensable tool after further validation is obtained [55]. Azuaje et al. conducted a study where data from proteomics, gene expression, and DL-based TA of histology samples were examined to discriminate between ccRCC samples and normal tissue. Interestingly, the image-based data correlated with that of proteomics, and the biological processes and pathways of the proteins involved correlated with extracellular organization activities and immune response. Furthermore, complementary RNA (cRNA) expression and gene expression also strongly correlated with image-based data (r = 0.76). Despite no radiological TA being performed in these studies, the relevant information obtained can be applied to further develop prediction scores comprising quantitative image data [57].

Disease Progression and Patient Outcome Prediction
The timely identification of patients who are at risk of developing worse outcomes is paramount for treatment clinical decision-making. Furthermore, predicting responses to targeted therapies or resistance profiles would lead to personalized medical algorithms, saving time and resources while avoiding adverse effects of unfruitful therapies, thereby also improving survival rates and RCC patients' quality of life. Recently, investigators explored these possibilities, and their findings are presented in Table 4. Antunes et al. reported a pilot study on predicting the response to sunitinib treatment in two patients with metastatic RCC. The authors sought to evaluate the feature differences between test and retest images retrieved from multiphasic MRI and PET scans using fluorothymidine as radiotracer at baseline and three weeks after administering sunitinib. The changes in image texture features were analyzed with ROIs from tumor and normal tissue images to assess any changes in both. Overall, a low variability between test and retest images for both modalities was found. Moreover, 63% difference was found in the top-ranked texture features of tumor regions from pretreatment images and after three weeks of sunitinib treatment, whereas only a 17% difference was found in the normal tissue regions. The authors concluded that MRI and PET texture features can detect early structural and functional changes of treatment response. Naturally, further studies with larger population are warranted [59].
Data extracted from RNA sequencing (RNA-seq) and gene expression profiles were processed and classified using ML-based methods by Singh et al. to discriminate patients with early stages (I-II) and late stages (III-IV) of pRCC. The principal component analysis solely based on gene expression accurately discriminated between normal and tumor samples and between early and late stages. Furthermore, the different algorithms used for most selected features yielded satisfactory accuracy rates, ranging between 82.5% and 88%, along with a precision and recall AUC of 0.69-0.79 and correlation coefficient of 0.60-0.68. Tracing the biological pathways and protein interaction revealed that most expression differences were in processes related to microtubules, chromosomes, and the cell cycle. Interestingly, late stages of pRCC had higher expression of kinetochore and centromere proteins, which are well-known alterations and distinctive neoplastic features leading to chromosome instability [60]. Tabibu et al. classified a TCGA (The Cancer Genome Atlas) dataset into high-and low-risk ccRCC according to a risk index based on tumor and nuclei shape features on histopathology slices constructed by a CNN. After grade discrimination, survival probability was tested, and the high-grade group was associated with a lower survival rate [61]. Clearly, radiomics-based progression prediction in RCC is only in its initial stages; however, if such a goal is reached, it will prove to be a major breakthrough for RCC management.

Limitations/Challenges of Radiomics
Although radiomics is attracting substantial attention in medicine and the field of oncology in general, its real-life implementation in the clinical scenario still faces obstacles. Firstly, the variability in study design, radiomic methods employed, texture features extracted, and recorded endpoints make it difficult to compare any two techniques and to perform quantitative analysis. Secondly, most ML and DL algorithms utilized in these studies were validated with their own dataset; therefore, without external validation, result generalizability and reproducibility cannot be applied to other datasets and populations [62]. Thirdly, repeatability, reproducibility, sample size, statistical power, and standardization are still vital factors to be considered in future investigations [63]. Fourthly, gene-based changes may result in significantly different radiomics features which are not consistent across all imaging phases [64]. Therefore, the link between image tumor properties and tumor biology is not straightforward, correlation does not imply causation, and a statistical relationship between radiomics and genetic footprints or prognosis should be established [65]. Lastly, if dataset heterogeneity is not addressed for the conduction of future studies, bridging the translational leap from an experimental research tool to an essential clinically applicable diagnostic and predictive instrument will be challenging [3].

Future Perspectives
The focus of future research should be constructing larger unified medical registries and databases to further enhance the accuracy of radiomics techniques. The use of improved algorithms should not be limited to large computer analyzing centers but extended to mobile devices and access by cloud services ensuring the protection of users' personal data. For individualized AI-based software to function with robotic platforms across operating rooms worldwide and take intraoperative decisions in real time, dedicated transnational regulatory approvals will have to be accorded. Additionally, there are major concerns regarding the reliability of diagnoses coming from algorithms and the role of programming biases interfering with patient management. Clearly, human intelligence will continue being the cornerstone in shaping future AI advances to ensure that these systems prioritize patient care and to develop mechanisms that prevent undesired use.

Conclusions
By further including data and making models more robust, the predictive precision of radiomics will continue contributing toward personalized medicine. Instant predictive analytics can be obtained from large semi-automated patient datasets and electronic medical records to be used for delivering insights into various disease processes. The diagnostic accuracy, however, is highly dependent on the quality of data and its efficient integration with different sources to generalize it. Shared decision-making will certainly not be replaced by these methods, although they may serve as an important accessory to the information patients obtain from traditional medical care. Undoubtedly, we are beginning a new era in medicine, and the future applications that radiomics may bring are limitless.