Role of DECT-Based Imaging Biomarkers and Machine Learning to Predict Renal Cell Carcinoma Subtypes

Keywords

clear cell renal cell carcinoma - RCC subtyping - machine learning

Introduction

Clear cell renal cell carcinoma (ccRCC) is the most common histopathologic type of RCC.[1] The histologic type of RCC acts as an independent predictor of distant metastasis and cancer-related death, with ccRCC having a worse prognosis than both chromophobe and papillary RCC.[2] [3] Development of various immunomodulators and targeted therapeutic agents also makes it mandatory to differentiate between subtypes of RCC as different histologic types have varying responses to various agents.[4] [5] While image-guided biopsy or surgical excision can provide the exact histopathology in most cases, they are invasive in nature. Imaging techniques like multiphasic computed tomography (CT) and magnetic resonance imaging (MRI) suffer from high subjectivity and overlapping findings among different RCC subtypes and grades.[6] [7] [8] [9]

Quantitative imaging parameters using dual-energy CT (DECT) and perfusion CT have been previously investigated to distinguish various renal masses, differentiate enhancing masses from cysts, and predict the histology and grade of RCCs.[10] [11] [12] [13] [14] In the era of precision medicine, machine learning (ML) models can be trained to differentiate benign and malignant tumors, predict histologic type and nuclear grading, genetic or molecular signatures, and predict prognosis, using the plethora of quantitative markers available on imaging. There are few studies on ML using DECT in renal masses, and most of them focus on the grading and prognostication of ccRCC. Few studies have evaluated the use of ML to differentiate ccRCC from non-ccRCC.[15] [16] [17] [18]

In this study, our aim was to evaluate the performance of DECT-based imaging markers to predict the histologic subtype of RCC as well as grade of ccRCC and to assess whether the accuracy could be improved using ML algorithms.

Materials and Methods

Dataset Retrieval

Ethical approval was waived due to the retrospective and observational study design. We searched our hospital's pathology report database, from January 2017 to November 2022, for histopathological nephrectomy specimens with proven RCC. These histopathological reports served as reference standards to classify RCCs into ccRCCs and non-ccRCCs and to further grade ccRCCs using the Fuhrman classification system. Dual-energy abdominal CT images of these patients were retrieved from our hospital's picture archiving and communication system (PACS) using their hospital identification numbers. Patients without a baseline preoperative scan in the PACS were excluded. Ultimately, 112 RCC patients who had a preoperative CT abdomen performed at our hospital and subsequently underwent surgery were included in the analysis. Of these, 25 were non-ccRCCs (comprising 10 papillary and 8 chromophobe RCCs), and 87 were ccRCCs. Among the 87 ccRCCs, 61 were classified as low grade (Fuhrman grades I–II), 18 as high grade (Fuhrman grades III–IV), and 8 cases were ungraded in the pathology reports and thus excluded from the Fuhrman grade correlation analysis. [Fig. 1] summarizes the inclusion and exclusion criteria and the final number of analyzed CT scans in different phases.

Fig. 1 Flowchart showing the study design. ccRCC, clear cell renal cell carcinoma; DECT, dual-energy computed tomography.

Image Acquisition

DECT was performed as per the routine institutional practice on a dual-source dual-energy 256-slice CT machine after intravenous administration of 1.8 to 2 mL/kg of nonionic iodinated contrast at a rate of 3.8 mL/s using a pressure injector. The DECT scan acquisition parameters are detailed in [Table 1]. Three postcontrast phases were acquired: corticomedullary phase at 35 seconds, nephrographic phase at 90 seconds, and a delayed phase at 15 minutes. Virtual noncontrast (VNC) images were generated from the dual-energy images, and noncontrast images were not acquired separately. The CT scans retrieved for the study were anonymized by removing their DICOM metadata and assigning each scan a new, unique study identification number.

Table 1 Dual-energy CT (DECT) acquisition parameters

Method of DECT	Dual source
Detector collimation (mm)	0.6
Tube voltage (kV)  Tube A  Tube B	140 100
Tube current time product (mA)  Tube A  Tube B	230 178
Gantry rotation time (s)	0.5
Pitch	0.6
Acquisition mode	Helical
Reconstructed section thickness (mm)	1.5
Matrix size	256 × 256
Reconstruction kernel	D30f
Reconstruction algorithm	Filtered back projection

Image Analysis

The DECT images were processed using a dedicated dual-energy software package (syngo.via VB10A, Siemens Healthineers). The software generated color-coded iodine maps by employing the iodine subtraction algorithm (Liver VNC, Siemens Healthineers). Two independent readers (with 10 and 7 years of experience in body imaging), who were blinded to the final histopathology results, performed the analysis. The readers drew circular regions of interest (ROIs) on the largest axial section of the tumor. In homogeneous lesions, the ROIs encompassed as much tumor area as possible. For heterogeneous lesions, the ROIs were carefully placed to include only the most avidly enhancing areas while excluding necrotic regions. Representative images showing the ROI placement are shown in [Fig. 2]. The iodine concentration (IC) within the tumor, measured in milligrams per milliliter, and the iodine ratio (IR), calculated as the IC in the tumor divided by the IC in the aorta at the level of the renal artery supplying the kidney with the tumor, were recorded for each ROI.

Fig. 2 Representative dual-energy CT image with iodine overlay map showing placement of the circular region of interest (ROI) over the most enhancing part of the tumor in (a) clear cell renal cell carcinoma (RCC) with (b) corresponding corticomedullary and (c) nephrographic phase images showing the hyperenhancing nature of the lesion. Similar dual-energy CT image with (d) iodine overlay map, (e) corticomedullary, and (f) nephrographic phase images in a case of non–clear cell RCC.

Machine Learning Models

Inter-reader agreement for the calculated parameters, IC and IR, was calculated using intraclass coefficient (ICC). Six ML models, including Logistic Regression (LR), Support Vector Machine (SVM), Random Forest Classifier (RFC), AdaBoost, Naive Bayes (NB), and Artificial Neural Network (ANN), were trained to perform binary classification between ccRCC and non-ccRCC, and to predict the grade of ccRCC as low or high, using the two DECT parameters, IC and IR. We further attempted to ensemble the outputs of each of these models using another set of ML architectures to see if the accuracy was improved when multiple ML techniques were applied together.

Each of the ML architectures used in the study is briefly described below.

• SVM: It is a method that finds the best way to separate different groups of data points by drawing a line (or in more complex scenarios, a plane, or a hyperplane) between them. This line is drawn to maximize the distance from the nearest points of any group, ensuring the clearest distinction. This model is robust to outliers and quickly converges to its final model.

• RFC: It uses many decision trees to make predictions. Each tree votes on the outcome, and the majority vote is taken as the final prediction. This approach reduces errors and improves accuracy by combining the strengths of multiple trees.

• AdaBoost: It is a boosting technique that combines several weak models (in our case, Haar cascade classifiers) to create a strong predictive model. It works by giving more weight to the mistakes of the previous models, focusing on the harder-to-predict instances to improve overall performance.

• NB: It is a probabilistic model based on Bayes' theorem. It assumes that features are independent of each other, which simplifies the calculation of the probability of different outcomes. Despite this simple assumption, it often performs surprisingly well for many types of problems.

• LR: It is a statistical method used to predict the probability of a binary outcome. It models the relationship between the input features and the probability of the outcome using a logistic function, making it suitable for classification tasks.

• ANN: It is a model inspired by the human brain. It consists of layers of interconnected nodes (neurons) that process data by passing it through multiple layers, learning complex patterns and relationships in the data to make accurate predictions.

Each of the features (IC and IR) were normalized between 0 and 1 using min–max normalization before passing them through the ML models. This step is crucial as it puts all parameters on an even scale, allowing each one to contribute equally to the model, regardless of their original scale or units. We did not add any priors to any of the networks. The training process was conducted using a standard ML library called sklearn, which provides tools for building and deploying ML models. The outputs of the ML models were thresholded at a given threshold (0.5) to assign a predicted class to each of the samples. The performance of the ML models was assessed in terms of accuracy, precision, recall, and F1 score. Area under the receiver operator characteristic (ROC) curve and Pearson's correlation coefficient were also calculated.

Results

Patient demographics and tumor characteristics: A total of 112 patients with RCCs were included in the study (mean age: 65 years; male:female: 61:51). Of these, 87 were pathologically proven ccRCCs, while 25 were non-ccRCCs. Non-ccRCCs included papillary RCC, chromophobe RCC, and collecting duct RCC ([Table 2]).

Table 2 Patient characteristics

Total number of patients	112
Mean age (y)	65
Gender (male:female)	61:51
Histologic type of RCC ( n  = 112) Clear cell (ccRCC) Non–clear cell RCC  Papillary cell  Chromophobe  Other non–clear cell subtypes	87 25 10 8 7
Histologic grade of ccRCC ( n  = 79)  Low  High	61 18

Grading was available in 79 cases of ccRCC, of which 61 were low-grade tumors and 18 were high-grade tumors.

RCC Subtyping (ccRCC vs. non-ccRCC)

Performance of Individual DECT Parameters

The mean IC and IR values were 2.49 mg/mL and 58.7 for reader 1 and 2.44 mg/mL and 57.8 for reader 2, respectively. The correlation coefficient between the two readers was 0.89 each for IC and IR ([Table 3]). This is represented graphically in [Fig. 3] where the IC ([Fig. 3a]) and IR ([Fig. 3b]) as measured by reader 1 and reader 2 form the x and y-axis, respectively; all the data points are seen clustered around a straight line, suggesting high agreement between the two readers.

Fig. 3 Scatter plots showing inter-reader agreement for (a) iodine concentration and (b) iodine ratio. Values from reader 1 are plotted along the x-axis (IC_1 and IR_1), while values determined by reader 2 are plotted along the y-axis (IC_2 and IR_2).

Table 3 Dual-energy CT parameters

	Iodine concentration (mg/mL)		Iodine ratio
	Clear cell RCC (n = 87)	Non–clear cell RCC (n = 25)	Clear cell RCC (n = 87)	Non–clear cell RCC (n = 25)
Mean	2.76	1.44	65.70	33.57
Median	2.60	1.4	60.58	32.15
Interquartile range	1.22	1.35	31.0	32.55
Standard deviation	1.17	0.82	25.05	21.94
Inter-reader agreement	0.894 (p < 0.0001)		0.889 (p < 0.0001)

Abbreviations: CT, computed tomography; RCC, renal cell carcinoma.

Note: The values of iodine concentration and iodine ratio were higher for ccRCC than non-ccRCC. Inter-reader agreement was very high for both parameters and the results were statistically significant.

For distinguishing ccRCC and non-ccRCCs, IC alone had an accuracy of 77.7% at a threshold of 6.8 mg/mL, while IR alone had an accuracy of 77.5% at a threshold of 155.65, with higher values indicating ccRCC in both parameters.

Performance of ML Models

Among the individual ML models, RFC and AdaBoost had a similar performance with an accuracy of 89.2% each, F1 score of 0.886 and 0.881, precision of 0.896 and 0.889, recall of 0.892 each, and area under the curve (AUC) of 0.888 and 0.970. The performance of the other ML models was good with an accuracy of 85.6, 82.8, 89.2, 83.8, and 79.3% for LR, SVM, NB, and ANN, respectively ([Table 4]). Combining the ML models improved the performance of all ML models. AdaBoost performed the best with an accuracy of 100%. The performance of other ML models also improved with an accuracy of 87.4, 89.2, 91.0, 87.4, and 86.5% for LR, SVM, RFC, NB, and ANN, respectively ([Table 5]). The results for histologic subtyping of RCC are depicted graphically using box and whisker plots in [Fig. 4], showing the accuracy of IC ([Fig. 4a]) and IR ([Fig. 4b]) individually without the use of ML, as well as the highest performing individual algorithm—the RFC ([Fig. 4c])—and the best performing combined ML algorithm—AdaBoost ([Fig. 4d]). Wide separation of the boxes of ccRCC and non-ccRCC, as seen in [Fig. 4d], indicates higher accuracy.

Fig. 4 Box and whisker plots for renal cell carcinoma (RCC) subtyping using (a) iodine concentration (IC), (b) iodine ratio (IR), (c) Random Forest classifier alone (RF), and (d) AdaBoost boosted by other machine learning models.

Table 4 Accuracy of DECT biomarkers and individual machine learning models for differentiating clear cell and non–clear cell RCC

	Iodine concentration	Iodine ratio	Logistic regression	Support Vector Machine	Random Forest Classifier	AdaBoost	Naive Bayes	Artificial Neural Network
Accuracy	0.777	0.775	0.856	0.828	0.892	0.892	0.838	0.793
F1 score	0.679	0.676	0.842	0.814	0.881	0.886	0.838	0.717
Precision	0.603	0.600	0.850	0.815	0.896	0.889	0.838	0.836
Recall	0.777	0.775	0.856	0.829	0.892	0.892	0.838	0.793
Threshold	6.800	155.65	0.5	0.5	0.5	0.5	0.5	0.5
Area under ROC	0.825	0.832	0.846	0.823	0.888	0.970	0.848	0.820

Abbreviations: DECT, dual-energy computed tomography; RCC, renal cell carcinoma; ROC, receiver operating characteristic curve.

Note: The accuracy of individual DECT parameters for prediction of the RCC subtype was moderate (IC and IR in the first two columns). All the machine learning models performed better than the individual DECT parameters. The Random Forest classifier and AdaBoost models had the highest accuracy and area under the curve.

Table 5 Accuracy of combined machine learning models for differentiating clear cell and non–clear cell RCC

	Logistic regression	Support Vector Machine	Random Forest classifier	AdaBoost	Naive Bayes	Artificial Neural Network
Accuracy	0.874	0.892	0.910	1.0	0.874	0.865
F1 score	0.858	0.881	0.910	1.0	0.872	0.853
Precision	0.880	0.896	0.910	1.0	0.871	0.860
Recall	0.874	0.892	0.910	1.0	0.874	0.865
Threshold	0.5	0.5	0.5	0.5	0.5	0.5
Area under ROC	0.866	0.885	0.972	1.0	0.911	0.870

Abbreviations: DECT, dual-energy computed tomography; RCC, renal cell carcinoma; ROC, receiver operating characteristic curve.

Note: Combining multiple machine learning algorithms to create another level of architecture improved the performance of all the models, with AdaBoost performing the best.

Fuhrman Grade Prediction for ccRCCs (Low vs. High Grade)

Performance of Individual DECT Parameters

To differentiate between high- and low-grade ccRCCs, IC alone had an accuracy of 77.9%, while IR alone had an accuracy of 77.6%. The discriminatory threshold values for IC and IR were 6.80 mg/mL and 128.55, respectively ([Table 6]).

Table 6 Accuracy of DECT biomarkers and individual machine learning models for differentiating high- and low-grade clear cell RCC

	Iodine concentration	Iodine ratio	Logistic regression	Support Vector Machine	Random Forest	AdaBoost	Naïve Bayes	Artificial Neural Network
Accuracy	0.779	0.776	0.776	0.776	0.776	0.789	0.776	0.776
F1 score	0.682	0.678	0.678	0.678	0.678	0.708	0.678	0.678
Precision	0.607	0.603	0.603	0.603	0.603	0.834	0.603	0.603
Recall	0.779	0.776	0.776	0.776	0.776	0.790	0.776	0.776
Threshold	6.80	128.55	0.5	0.5	0.5	0.5	0.5	0.5
Area under ROC	0.524	0.463	0.544	0.994	0.922	0.898	0.515	0.538

Abbreviations: DECT, dual-energy computed tomography; RCC, renal cell carcinoma; ROC, receiver operating characteristic curve.

Note: For grade prediction, individual DECT parameters and machine learning models had similar performance with moderate accuracy.

Performance of ML Models

The best performing ML model was AdaBoost with an accuracy of 78.9% and F1 score of 0.708. The ML models including LR, SVM, RFC, NB, and ANN had a similar accuracy of 77.6%. Combining ML techniques improved the accuracy of all the models except LR and ANN. The best performing model was again AdaBoost with an accuracy of 100%. The accuracies of SVM, RFC, and NB were 97.4, 98.7, and 94.7%, respectively ([Table 7]). The results for grading of ccRCC are depicted graphically using box and whisker plots in [Fig. 5], showing the accuracy of IC ([Fig. 5a]) and IR ([Fig. 5b]) individually without the use of ML, as well as the highest performing individual algorithm—the RFC ([Fig. 5c])—and the best performing combined ML algorithm—AdaBoost ([Fig. 5d]). Wide separation of the boxes of ccRCC and non-ccRCC, as seen in [Fig. 5d], indicates higher accuracy.

Fig. 5 Box and whisker plots for grading of clear cell renal cell carcinoma (ccRCC) using (a) iodine concentration (IC), (b) iodine ratio (IR), and (c) Random Forest classifier alone (RF) and (d) AdaBoost in combination with other machine learning algorithms.

Table 7 Accuracy of DECT biomarkers and combined machine learning models for differentiating high-grade and low-grade clear cell RCC

	Logistic regression	Support vector machine	Random Forest classifier	AdaBoost	Naive Bayes	Artificial Neural Network
Accuracy	0.776	0.974	0.987	1.0	0.947	0.776
F1 score	0.678	0.974	0.987	1.0	0.948	0.678
Precision	0.603	0.974	0.987	1.0	0.951	0.603
Recall	0.776	0.974	0.987	1.0	0.947	0.776
Threshold	0.5	0.5	0.5	0.5	0.5	0.5
Area under ROC	0.975	0.986	1.0	1.0	0.988	0.973

Abbreviations: DECT, dual-energy computed tomography; RCC, renal cell carcinoma; ROC, receiver operating characteristic curve.

Note: Combining the ML techniques improved the accuracy of the models for grade prediction of clear cell RCC (ccRCC), with AdaBoost performing the best.

Discussion

In this study, we have assessed the performance of quantitative DECT biomarkers and various ML models to distinguish the type (ccRCC vs. non-ccRCC) and grade (low vs. high) of ccRCCs. IC and IR were used as objective DECT parameters to represent the enhancement characteristics of RCC, which are generally evaluated subjectively in order to identify ccRCCs and non-ccRCCs. The inter-reader agreement for both DECT parameters was high, suggesting inter-rater reproducibility of results. The accuracy of IC alone and IR alone for distinguishing ccRCC and non-ccRCCs was moderate. Among the ML models, RF and AdaBoost had a very high accuracy and combining ML techniques improved their performance. AdaBoost had an accuracy of 100%, which meant that it was able to predict the correct type of RCC and the right grade of ccRCC in all cases.

Previous studies on DECT in RCC have found that IC and IR can be used to predict the RCC subtype with a high accuracy, which can further be improved with the combination of perfusion CT and radiomics.[11] A recent study also developed a nomogram for the preoperative classification of high-grade versus low-grade tumors among ccRCC using clinical parameters such age, systemic immune-inflammation index, and slope of spectral CT curve in the cortical phase, using data from 73 ccRCC patients.[19] Most of the published literature on ML for prediction of RCC subtype or grade is based on the use of radiomics and texture parameters. Several studies have trained ML models using textural data with a high degree of accuracy.[18] [20] [21]

However, few studies have used DECT data to train ML models.

In the present study, we have trained several types of ML models including SVM, RFC, AdaBoost, NB, LR, and ANN. We have further ensembled the outputs of each of these models with another set of ML architectures to further improve their accuracy. The algorithms were trained on imaging-based quantitative markers, including IC and IR, after normalizing their values using min–max normalization. The performance was then assessed at a given threshold of 0.5, and we found that there was high accuracy of prediction of ccRCC versus non-ccRCC, which was further improved when an additional set of ML architectures was added to the ensemble output. It is well known that ccRCCs are hyperenhancing as compared to non-ccRCCs, and this was probably reflected in the quantitative DECT parameters used to train the ML models. We did not perform external validation due to the small sample size and retrospective design of the study; however, future prospective studies using larger datasets could be used to validate these results in a more “real-world” setting.

As seen in our study, the combination of DECT parameters with ML has the potential to improve the accuracy of predicting the histological subtype and grade of RCCs noninvasively. A recently published study by Bing et al trained ML models using DECT-based radiomics to predict the number of interstitial collagen fibers and pseudocapsule thickness in ccRCCs. They found that the combined model, which used data from iodine-based material decomposition images and mixed energy images, had the best performance with a specificity of 0.87 and a sensitivity of 0.75.[22] In our study, we used DECT data, combining IC and IR values to train various ML models with a high accuracy of predicting RCC subtypes and ccRCC grades. We also ensembled multiple ML algorithms together to a higher level of architecture to further boost the performance of the ML models.

Limitations

This study was limited by its retrospective nature and small sample size. Therefore, prediction of various subtypes of non-ccRCC was not possible. The dataset included only RCCs, and the algorithm was trained only to differentiate ccRCC from non-ccRCC. We did not include any renal lesions other than RCCs, such as angiomyolipomas or other non-RCC malignancies, and thus the algorithm would not be able to differentiate such lesions from RCCs.

While surgery would still be required irrespective of the histologic subtype of RCC, the differentiation between ccRCC and non-ccRCC is clinically relevant as it affects the therapeutic choice, response to treatment, and prognosis. While studies have found that there is no significant difference in the prognosis of papillary and chromophobe RCCs, both are significantly better than ccRCC.[2] [10] Biopsy is often required prior to starting immunotherapy in metastatic disease, and being able to distinguish ccRCC from non-ccRCC would be useful in such cases.

DECT was performed only on a dual-source machine. This limits its generalizability as different DECT techniques may yield different values of IC.

There may be the problem of overfitting in the present study resulting in such high accuracies, as we did not perform external validation. Validation of these results using a larger, external dataset could further check and correct for this issue inherent to ML models.

Thus, a larger prospective cohort study including non-RCC renal masses can be conducted to develop more robust and widely applicable ML models that can predict the different subtypes of RCCs including the less common non-ccRCCs and differentiate RCCs from other renal masses.

Conclusion

DECT-based quantitative biomarkers have a moderate diagnostic accuracy to differentiate ccRCC from non-ccRCC and to predict the grade of the tumor. The use of ML models can significantly improve the accuracy of the DECT parameters. Thus, DECT parameters combined with ML can provide noninvasive, accurate, and reliable biomarkers for RCC subtype prediction, which is crucial for prognostication and choice of targeted therapy.

• References

• 1 Moch H, Cubilla AL, Humphrey PA, Reuter VE, Ulbright TM. The 2016 WHO classification of tumours of the urinary system and male genital organs: part A—renal, penile, and testicular tumours. Eur Urol 2016; 70 (01) 93-105
• 2 Leibovich BC, Lohse CM, Crispen PL. et al. Histological subtype is an independent predictor of outcome for patients with renal cell carcinoma. J Urol 2010; 183 (04) 1309-1315
• 3 Moch H, Gasser T, Amin MB, Torhorst J, Sauter G, Mihatsch MJ. Prognostic utility of the recently recommended histologic classification and revised TNM staging system of renal cell carcinoma: a Swiss experience with 588 tumors. Cancer 2000; 89 (03) 604-614
• 4 Atkins MB, Tannir NM. Current and emerging therapies for first-line treatment of metastatic clear cell renal cell carcinoma. Cancer Treat Rev 2018; 70: 127-137
• 5 De Vries-Brilland M, Gross-Goupil M, Boughalem E. et al. Are immune checkpoint inhibitors (ICI) a valid option for papillary renal cell carcinoma (pRCC)? A multicenter retrospective study. J Clin Oncol 2019; 37 (07) 582-582
• 6 Egbert ND, Caoili EM, Cohan RH. et al. Differentiation of papillary renal cell carcinoma subtypes on CT and MRI. AJR Am J Roentgenol 2013; 201 (02) 347-355
• 7 Karlo CA, Di Paolo PL, Chaim J. et al. Radiogenomics of clear cell renal cell carcinoma: associations between CT imaging features and mutations. Radiology 2014; 270 (02) 464-471
• 8 Abel EJ, Carrasco A, Culp SH. et al. Limitations of preoperative biopsy in patients with metastatic renal cell carcinoma: comparison to surgical pathology in 405 cases. BJU Int 2012; 110 (11) 1742-1746
• 9 Young JR, Margolis D, Sauk S, Pantuck AJ, Sayre J, Raman SS. Clear cell renal cell carcinoma: discrimination from other renal cell carcinoma subtypes and oncocytoma at multiphasic multidetector CT. Radiology 2013; 267 (02) 444-453
• 10 Udare A, Walker D, Krishna S. et al. Characterization of clear cell renal cell carcinoma and other renal tumors: evaluation of dual-energy CT using material-specific iodine and fat imaging. Eur Radiol 2020; 30 (04) 2091-2102
• 11 Manoharan D, Netaji A, Diwan K, Sharma S. Normalized dual-energy iodine ratio best differentiates renal cell carcinoma subtypes among quantitative imaging biomarkers from perfusion CT and dual-energy CT. AJR Am J Roentgenol 2020; 215 (06) 1389-1397
• 12 Mileto A, Marin D, Alfaro-Cordoba M. et al. Iodine quantification to distinguish clear cell from papillary renal cell carcinoma at dual-energy multidetector CT: a multireader diagnostic performance study. Radiology 2014; 273 (03) 813-820
• 13 Zarzour JG, Milner D, Valentin R. et al. Quantitative iodine content threshold for discrimination of renal cell carcinomas using rapid kV-switching dual-energy CT. Abdom Radiol (NY) 2017; 42 (03) 727-734
• 14 Salameh JP, McInnes MDF, McGrath TA, Salameh G, Schieda N. Diagnostic accuracy of dual-energy CT for evaluation of renal masses: systematic review and meta-analysis. AJR Am J Roentgenol 2019; 212 (04) W100-W105
• 15 Chen S, Zhang N, Jiang L. et al. Clinical use of a machine learning histopathological image signature in diagnosis and survival prediction of clear cell renal cell carcinoma. Int J Cancer 2021; 148 (03) 780-790
• 16 Lin F, Cui EM, Lei Y, Luo LP. CT-based machine learning model to predict the Fuhrman nuclear grade of clear cell renal cell carcinoma. Abdom Radiol (NY) 2019; 44 (07) 2528-2534
• 17 Suarez-Ibarrola R, Hein S, Reis G, Gratzke C, Miernik A. Current and future applications of machine and deep learning in urology: a review of the literature on urolithiasis, renal cell carcinoma, and bladder and prostate cancer. World J Urol 2020; 38 (10) 2329-2347
• 18 Xv Y, Lv F, Guo H. et al. Machine learning-based CT radiomics approach for predicting WHO/ISUP nuclear grade of clear cell renal cell carcinoma: an exploratory and comparative study. Insights Imaging 2021; 12 (01) 170
• 19 Zhang H, Li F, Jing M, Xi H, Zheng Y, Liu J. Nomogram combining pre-operative clinical characteristics and spectral CT parameters for predicting the WHO/ISUP pathological grading in clear cell renal cell carcinoma. Abdom Radiol (NY) 2024; 49 (04) 1185-1193
• 20 Gurbani S, Morgan D, Jog V. et al. Evaluation of radiomics and machine learning in identification of aggressive tumor features in renal cell carcinoma (RCC). Abdom Radiol (NY) 2021; 46 (09) 4278-4288
• 21 He X, Wei Y, Zhang H. et al. Grading of clear cell renal cell carcinomas by using machine learning based on artificial neural networks and radiomic signatures extracted from multidetector computed tomography images. Acad Radiol 2020; 27 (02) 157-168
• 22 Bing X, Wang N, Li Y. et al. The value of dual-energy computed tomography-based radiomics in the evaluation of interstitial fibers of clear cell renal carcinoma. Technol Cancer Res Treat 2024; 23: 15 330338241235554