Development and validation of a deep learning radiomics model for predicting capsular invasion in small renal masses: a multicenter retrospective study

Introduction

Renal cell carcinoma (RCC) is one of the most common malignant tumors affecting the urinary system, with an increasing incidence worldwide, and clear cell renal cell carcinoma (ccRCC) is its most common pathological subtype (1,2). Small renal masses (SRMs) are defined as contrast-enhancing renal cortical neoplasms ≤4 cm in greatest diameter, and 70–80% of SRM are malignant, usually stage T1a RCC (3-5). As the current detection rate of SRM has increased, the treatment modalities for SRM have changed, with primary intervention (partial nephrectomy, thermal ablation) and active surveillance (AS) becoming the dominant therapeutic approaches (5,6).

The renal capsule is a fibrous membrane that functions as a barrier to prevent outward invasion of renal malignancy into perirenal adipose tissue. Once the renal capsule is broken through, it would result in a leap from T1 to T3 in the tumor-node-metastasis (TNM) staging system of the 8th edition of the American Joint Committee on Cancer (AJCC) guideline, influencing the clinician to take a different surgical approach (7). It is well known that as the size of a renal malignancy grows, the likelihood of the risk of capsule invasion increases (8). Many studies have shown that capsule invasion in urologic tumors not only makes surgery more difficult, but is also a biomarker of poor prognosis (9-12). Therefore, early and accurate evaluation of renal capsule status in SRM is essential. Although there is no conclusive evidence that SRM capsule invasion is a decisive factor influencing the treatment modality, it could be a contributing factor for SRM patients to adopt more aggressive interventions than AS, despite the fact that SRM has comparable outcomes for AS and primary intervention at 5 years (13).

It is difficult for radiologists to confirm renal capsule invasion on computed tomography (CT) images prior to SRM treatment. Radiologists are limited to assessing renal capsule status exclusively in SRM that have invaded the perirenal adipose tissue but cannot evaluate its status in masses adjacent to or protruding from the renal silhouette (14,15). Precise determination of capsular invasion pretreatment could enable timely interventions to mitigate pathological progression.

Radiomics and deep learning techniques have been increasingly applied to the field of medical imaging with the development of computer science. Radiomics and deep learning are analytical methods that extract a large number of high-throughput features from medical images, which reflect the micro and macro characteristics of tissues and lesions (16,17). At present, the research of radiomics and deep learning in renal cancer mainly focuses on the identification of benign and malignant tumors in kidney (18-21), the classification of histological subtypes of renal cancer (22,23), and the prediction of renal cancer pathological grading (24,25) and postoperative prognosis (26,27). To the best of our knowledge, there are few studies reporting on renal capsule invasion using radiomics and deep learning methods, especially SRM.

Our study aims to explore the predictive performance of a deep learning radiomics (DLR) model based on CT images for renal capsule invasion in SRM and to evaluate its potential clinical utility. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2539/rc).

Methods

Patients

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the ethics committees of all the involved hospitals: the First Affiliated Hospital of Shandong First Medical University (No. 2024S943) (Center 1), the Third Affiliated Hospital of Shenzhen University (No. 2023-LHQRMYY-KYLL-55) (Center 2), and the Union Hospital, Tongji Medical College, Huazhong University of Science and Technology (No. 2023S187) (Center 3), and informed consent was waived due to the retrospective study. A total of 413 SRM were retrospectively recruited between July 2017 and September 2024 from the First Affiliated Hospital of Shandong First Medical University (Center 1), the Third Affiliated Hospital of Shenzhen University (Center 2), and the Union Hospital, Tongji Medical College, Huazhong University of Science and Technology (Center 3). The inclusion criteria were as follows: (I) pathologically confirmed T1a RCC with postoperative pathological assessment of renal capsule status; (II) preoperative kidney CT enhancement. Exclusion criteria: (I) preoperative CT corticomedullary phase (CMP) was absent and image quality was poor; (II) clinicopathologic data were incomplete. The SRM from Center 1 and Center 2 were used as the training set, while the SRM from Center 3 served as the independent external testing set. The whole process of enrollment is shown in Figure 1.

Figure 1 The whole process of enrollment. Center 1: the First Affiliated Hospital of Shandong First Medical University; Center 2: the Third Affiliated Hospital of Shenzhen University; Center 3: the Union Hospital, Tongji Medical College, Huazhong University of Science and Technology. CMP, corticomedullary phase; SRM, small renal mass.

CT imaging parameters

All patients underwent a preoperative triphasic renal CT scan with copious amounts of water before the scans. After the unenhanced phase (UP) scans, non-ionic contrast agent was injected into the elbow vein at a rate of 3 mL/s using a high-pressure syringe at a dosage of 1.5–2 mL/kg, and CMP and nephrographic phase (NP) images were acquired from the top of the diaphragm to the iliac wing after 30 and 70 s, respectively. The specific parameters of the CT scanners in all hospitals are shown in Table S1.

Image segmentation and preprocessing

The Image Biomarker Standardization Initiative (IBIS) was followed to preprocess the renal CMP phase images. Before segmentation of the region of interest (ROI), gray-scale standardization of the images was first performed by setting the window width of all CT images to 300 Hounsfield Unit (HU) and the window level to 100 HU. Then, linear interpolation was used to isotropically resample all voxels to 1×1×1 mm³ to ensure the consistency of spatial resolution under different imaging parameters. A radiologist (5 years of experience in diagnosis) manually outlined the three-dimensional SRM contours layer by layer on the CMP images using ITK-SNAP version 3.8.0 (www.itksnap.org), and all segmented masks were confirmed by another radiologist with 10 years of working experience, and when disputes arose, they were resolved by consensus. The radiologists were blinded to the clinicopathologic data except for renal cancer. After one month, 30 SRM were randomly selected for SRM segmentation by the same radiologist to assess intra-observer reproducibility. High intra-observer correlation coefficient (ICC) in tumor segmentation ensures the reliability of extracted imaging features, forming a fundamental basis for building robust and generalizable predictive models.

Radiomics features extraction and selection

The open-source tool PyRadiomics (28) (https://pyradiomics.readthedocs.io) was used to extract SRM radiomics features based on SRM three-dimensional ROIs. A bin with a fixed width of 25 HU was used for discretization of the image grayscale. Radiomics features include: (I) first order statistics feature; (II) shape-based feature; and (III) texture feature. And texture features were extracted using several different methods, including the Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Neighboring Gray Tone Difference Matrix (NGTDM) and Gray Level Dependence Matrix (GLDM).

Firstly, the z-score normalization was performed for radiomics features and deep learning features respectively. ICC >0.75 was considered to have good features reproducibility. The Mann-Whitney U test was then used for further feature selection, keeping only radiomics features and deep learning features with P<0.05. The Spearman rank correlation analysis was used to calculate the correlation among features and only one feature with a correlation coefficient of more than 0.9 between any two features was retained. Subsequently, the most significant features associated with SRM capsule invasion were identified and selected by maximum relevance and minimum redundancy (mRMR) analysis and least absolute shrinkage and selection operator (LASSO) regression (29). The 30 features with the highest mRMR rankings were selected for 10-fold cross validated LASSO regression to develop the radiomics (Rad) model.

Deep learning features extraction and model development

Deep learning features were extracted using the ResNet50 model of transfer learning. The axial images with the largest cross-sectional area in the three-dimensional ROI of the SRM were cropped out, and the linear difference was used to resample these images to 224×224, and the mean and standard deviation of the pixel intensities were normalized to 0 and 1. We then used the pretrained ResNet50 model as the basic architecture to train our task by freezing most of its convolutional layers and adding two new fully-connected layers (with Sigmoid activation function) in order to preserve the generic image features extracted by the pretrained model. After the initial training was completed, the top convolutional layer was gradually unfrozen and a low learning rate fine-tuning strategy was applied to enable the model to adaptively learn high-level specific features of SRM images. Eventually, the established deep transfer learning (DTL) model was set to 50 training epochs with a learning rate of 0.0001 and a batch size of 32. At the end of the model training, the final fully-connected layers were removed and the features of the average pooling layer were extracted as the deep learning features.

Rad and DLR model development

The DLR model for predicting SRM renal capsule invasion was developed by fusing comprehensive radiomics features with deep learning features processed through principal component analysis (PCA). DLR features selection followed the same methodology as for radiomics features. Eight machine learning (ML) algorithms were employed to construct Rad and DLR models for identifying SRM renal capsule invasion. These models were trained and tuned on the training set, and the optimal parameters of the models were determined using Grid Search and the 5‐fold cross‐validation method. Class weight balancing strategies were employed during model training to address the impact of dataset imbalance on performance metrics. The whole study flow is shown in Figure 2.

Figure 2 Workflow of the study. AUC, area under the curve; DCA, decision curve analysis; DLR, deep learning radiomics; DTL, deep transfer learning; LASSO, least absolute shrinkage and selection operator; mRMR, maximum relevance and minimum redundancy; Rad, radiomics; ROC, receiver operating characteristic; SHAP, SHapley Additive exPlanations.

Statistical analysis

Baseline data were processed using SPSS (22.0) and a statistical package (Python version). Keras (https://github.com/fchollet/keras) and TensorFlow (https://www.tensorflow.org) in Python (version 3.7.12) were used to build ResNet50 structure with an NVIDIA GeForce RTX 4060 GPU graphics processor. Continuous variables were presented as mean ± standard deviation and categorical variables were presented as frequency and percentage. Normal or non-normal distribution of continuous variables was determined by the Kolmogorov-Smirnov test. The Mann-Whitney U test or Student’s t-test was used for continuous variables in the SRM renal capsule invasion and renal capsule non-invasion groups, and the Chi-squared test or Fisher exact test was used for categorical variables, with two-sided P<0.05 indicating statistical significance. Metrics were used to evaluate the performance of the Rad, DTL, and DLR models, including accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV), receiver operating characteristic curve (ROC), and area under the curve (AUC). The integrated discrimination index (IDI) and DeLong test were used to compare the performance of different models. The feature contributions to the DLR model were visualized using the SHapley Additive exPlanations (SHAP). Decision curve analysis (DCA) and calibration curve were used to evaluate the clinical application value and verify the degree of calibration of the models, respectively.

Results

Clinical baseline characteristics

A total of 413 individuals (mean age ± standard deviation, 57.35±11.40 years; mean size ± standard deviation, 2.97±0.78 cm; 285 men) from three centers were eventually enrolled in the study from July 2017 to September 2024, and 330 individuals (57.27±11.58 years; 3.00±0.76 cm; 226 men) from Center 1 (n=176) and Center 2 (n=154) were grouped into training set, and 83 individuals (57.67±10.76 years; 2.85±0.83 cm; 59 men) of Center 3 as an external testing set. We analyzed the baseline characteristics and the relevant preoperative CT image qualitative features of renal capsule invasion in the training set and the external testing set for 413 patients, as illustrated in Table 1.

Features extraction, selection, and DLR model building

A total of 1,834 radiomics features were extracted. 23 radiomics features were ultimately retained through ICC, Mann-Whitney U test, Spearman rank correlation analysis, and mRMR/LASSO methods to construct the Rad model for SRM capsular invasion. A pretrained ResNet50 was fine-tuned for our task, yielding 2,048 deep learning features extracted from average pooling layer. 8 deep learning features were kept after features compression via PCA. A pretrained ResNet50 was fine-tuned for our task, yielding 2,048 deep learning features extracted from average pooling layers. After feature compression via PCA, 8 deep learning features were retained. Feature fusion was done between the extracted radiomics features and the compressed deep learning features. Following the same radiomics feature selection process, only 11 features were ultimately preserved to construct the DLR model for SRM capsular invasion. Rad and DLR feature selection results are shown in Figure S1.

On the training set, the ACC and AUC for the Rad, DTL, and DLR models were 0.782, 0.846 (95% CI: 0.803–0.888); 0.812, 0.890 (95% CI: 0.854–0.927); and 0.764, 0.855 (95% CI: 0.815–0.895), respectively. On the external testing set, the ACC and AUC for Rad, DTL, and DLR models were 0.699, 0.746 (95% CI: 0.639–0.852); 0.687, 0.715 (95% CI: 0.601–0.829); and 0.675, 0.734 (95% CI: 0.623–0.843). Model evaluation metrics are detailed in Table 2 and Figure 3. On the external testing set, IDI and DeLong tests indicated no statistically significant difference in performance between the DLR model and the Rad or DTL models. Performance comparisons of the three models are shown in Figure S2.

Figure 3 ROC curves of the Rad, DTL, and DLR models in training and external testing sets. AUC, area under the curve; DLR, deep learning radiomics; DTL, deep transfer learning; Rad, radiomics; ROC, receiver operating characteristic.

DLR model visualization displays SHAP value importance for all features (Figure 4). Deep learning features ranked as the top three most important features for SRM renal capsule invasion in the DLR model, with DL_1 and DL_3 exhibiting negative correlations with SRM renal capsule invasion. The importance of radiomics features was relatively weaker. DCA and calibration curves were used to evaluate the clinical utility and calibration of the Rad, DTL, and DLR models on the training and external testing sets, respectively (Figure S3). Although the DLR model demonstrated a superior net clinical benefit range on the external testing set compared to the other models, the advantage was not highly significant.

Figure 4 Feature importance of the DLR model on the external testing set using SHAP. DLR, deep learning radiomics; SHAP, SHapley Additive exPlanations.

Discussion

To the best of our knowledge, this is the first study to use CT images of renal CMP to build a predictive model for SRM capsule invasion, combining radiomics and deep learning features. The model achieved an AUC of 0.855 and 0.734 on the training and external testing sets, respectively. The external validation performance, while moderate, holds clinical significance given the specific challenge of preoperatively detecting capsular invasion in SRMs—a task where conventional CT assessment is highly limited. For these subtle lesions, the DLR model providing quantitative risk stratification offers a valuable adjunct to subjective radiological interpretation. In ambiguous clinical scenarios, such as tailoring surveillance intensity for patients on AS or refining surgical planning for surgical candidates, it offers an objective metric to inform personalized strategies. And its utility as a decision-support tool is underscored by a net clinical benefit across relevant thresholds in DCA. Although the DLR model did not achieve a statistically superior AUC compared to single-modality models, its robust generalizability and the complementary information captured by its integrated features are substantive advantages. The significant value of the model in our study is that it provides a feasible, non-invasive method for preoperative prediction of SRM capsule invasion, objectively guiding personalized management.

Notably, SHAP interpretability analysis indicated that deep learning features capture more distinct and deeper image information relevant to predicting SRM capsular invasion compared to handcrafted radiomics features. Thus, the research value of the DLR model lies not only in its successful integration of multi-source information but also in revealing the significant predictive potential inherent in deep learning features. It also highlights the superior predictive performance of deep learning, rather than its inherent interpretability. Techniques like feature visualization that map deep features back to images hold promise in bridging the gap between high-performance models and clinical interpretability needs, thereby providing a clear direction for developing more robust and trustworthy models.

Capsule invasion is an important invasiveness and prognostic indicator (11,12,30,31). Studies have shown that the occurrence of capsule invasion is usually accompanied by other invasive features, such as tumor necrosis, lymphovascular invasion, or metastasis (8,31,32). In our cohort, the indicator of tumor necrosis demonstrated a difference between the two data sets of SRM capsule invasion and non-invasion, and the results were consistent with previous studies (8,33). Of interest, tumor size is significantly and positively correlated with the risk of capsule breakthrough. Especially when the maximum diameter of the tumor exceeds 4 cm, the probability of its invasion of the renal capsule tends to be elevated (8), and the 5-year specific survival rate of such patients is reduced (34). Our study was innovatively limited to SRM with a maximum diameter of ≤4 cm. Compared with the previous study designed by Yang’s team (35) that included full-size renal cancers, our severe filtering allowed for risk stratification by evaluating capsule status at an early stage. Precise risk stratification assessment during the SRM provides more targeted evidence for clinical decision-making.

While multimodal imaging integrating CT and MRI provides a comprehensive diagnostic landscape for SRM (36), the present study is strategically focused on the CMP of CT images. This phase was selected as the primary dataset for model construction because it offers an optimal window for visualizing the pathophysiological mechanisms underlying capsular invasion. CMP, characterized by peak enhancement in the cortex and tumor parenchyma, offers a unique non-invasive advantage for assessing this critical interface in ccRCC. This is because ccRCC features significant vascular proliferation, highlighting perfusion abnormalities around the tumor—such as the peritumoral cortex low-enhancement (PCLE) sign—suggesting direct association with invasive pathological processes including vascular disruption, parenchymal infiltration, or capsular invasion (37). For SRMs, CMP features effectively capture critical tumor-capsule spatial relationships, such as irregular or infiltrative margins, which are established prognostic markers (38). Previous evidence indicates CMP features correlate more strongly with capsular invasion than other phases (35), further confirming this biological principle. Therefore, our study focuses on the CMP phase, which maintains high diagnostic specificity for ccRCC invasive potential (39) while aligning with the pathophysiological basis of the tumor. It also simplifies the evaluation process, reducing analytical complexity and inter-observer variability (40).

Currently, there are very few studies focusing on SRM, especially SRM capsule invasion. Anush et al. (41) developed a fully automated method based on U-Net for the detection of SRM on contrast-enhanced magnetic resonance imaging (CE-MRI), which was evaluated on CE-MRI scans of 118 patients with a recall and precision of 86.2% and 83.3%, respectively. Maddalo et al. (42) and Nassiri et al. (43) constructed CT radiomics models to differentiate between benign and malignant tumors in SRM, respectively, with AUC of 0.79±0.12 and 0.77 (95% CI: 0.69–0.85) in the external dataset. Feng et al. (44) developed and validated a nomogram for preoperative differentiation of benign and malignant SRM based on multiphase CT radiomics and clinical data. The results showed that the AUC of this nomogram was 0.988 and 0.968 on the training and testing sets. More specifically, Yang et al. (45) investigated the accuracy of radiomics model in distinguishing between small (<4 cm) renal angiomyolipoma without visible fat (AMLwvf) and RCC. Their study showed that radiomics features extracted from unenhanced CT were enough to differentiate AMLwvf and RCC. Lin et al. (46) and Gao et al. (47) built radiomics models to predict Ki67 and World Health Organization and International Society of Urological Pathology (WHO/ISUP) nuclear grading of SRM, respectively, and the AUC of the models in the test set were 0.668 and 0.902, which fully illustrated the correlation between imaging features and pathological indicators of SRM. As far as we know, there are only two studies on renal capsule invasion based on imaging features. Zhang et al. (48) used CMP CT images for semi-automatic segmentation of ccRCC and then built a radiomics model for predicting renal capsule invasion in ccRCC, with AUC of 0.83 (95% CI: 0.68–0.98) and 0.74 (95% CI: 0.51–0.97) in the two datasets, respectively. Yang et al. (35) developed a radiomics model based on multiphase CT for preoperative prediction of renal capsule invasion in patients with RCC. The results showed that CMP phase had the highest AUC of 0.81 when using a forward neural network (FNN) classifier, and it was also found that features extracted from the tumor region outperformed the border region at multiple phases. It is worth noting that their study was not on SRM.

At the same time, there are some limitations of our study. First, we only choose the CMP phase to study, although this phase shows the best results for SRM, we should try to use multiple phases to build the model and try to include relevant clinical factors to improve the study. Second, although the adopted two-dimensional model cannot leverage the full geometric and spatial information inherent in the three-dimensional data, it significantly reduces the number of model parameters. This reduction not only facilitates more efficient training and convergence but also aligns with our rationale that capsular invasion may be sufficiently predicted by textural and enhancement features observable within a single representative slice. Third, the generalization ability of the DLR model in this study is insufficient, and the model will be further tested and validated on multiple high-quality datasets in the future.

Conclusions

This study developed a DLR model fusing radiomics and deep learning features to predict SRM renal capsule invasion. Although the performance of the model is not significantly superior to that of radiomics or deep learning models, it’s still a valuable exploratory study.

Acknowledgments

The authors thank Bo Liang from the Department of Radiology, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, and Huancheng Yang from the Third Affiliated Hospital of Shenzhen University for their help in data collection.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2539/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2539/dss

Funding: This study was supported by Shandong-Chongqing Science and Technology Cooperation Project (No. 2024LYXZ021).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2539/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committees of all the involved hospitals: the First Affiliated Hospital of Shandong First Medical University (No. 2024S943) (Center 1), the Third Affiliated Hospital of Shenzhen University (No. 2023-LHQRMYY-KYLL-55) (Center 2), and the Union Hospital, Tongji Medical College, Huazhong University of Science and Technology (No. 2023S187) (Center 3), and informed consent was waived due to the retrospective study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Znaor A, Lortet-Tieulent J, Laversanne M, Jemal A, Bray F. International variations and trends in renal cell carcinoma incidence and mortality. Eur Urol 2015;67:519-30. [Crossref] [PubMed]
2. Hsieh JJ, Purdue MP, Signoretti S, Swanton C, Albiges L, Schmidinger M, Heng DY, Larkin J, Ficarra V. Renal cell carcinoma. Nat Rev Dis Primers 2017;3:17009. [Crossref] [PubMed]
3. Withington J, Neves JB, Barod R. Surgical and Minimally Invasive Therapies for the Management of the Small Renal Mass. Curr Urol Rep 2017;18:61. [Crossref] [PubMed]
4. Sanchez A, Feldman AS, Hakimi AA. Current Management of Small Renal Masses, Including Patient Selection, Renal Tumor Biopsy, Active Surveillance, and Thermal Ablation. J Clin Oncol 2018;36:3591-600. [Crossref] [PubMed]
5. Wang Y, Butaney M, Wilder S, Ghani K, Rogers CG, Lane BR. The evolving management of small renal masses. Nat Rev Urol 2024;21:406-21. [Crossref] [PubMed]
6. Chen DY, Uzzo RG. Evaluation and management of the renal mass. Med Clin North Am 2011;95:179-89. [Crossref] [PubMed]
7. Gallardo E, Méndez-Vidal MJ, Pérez-Gracia JL, Sepúlveda-Sánchez JM, Campayo M, Chirivella-González I, García-Del-Muro X, González-Del-Alba A, Grande E, Suárez C. SEOM clinical guideline for treatment of kidney cancer (2017). Clin Transl Oncol 2018;20:47-56. [Crossref] [PubMed]
8. Zhang XP, Huang ZX, Yu LP, Zhang XW, Li Q, Liu SJ, Xu T. Clinical and pathological analysis of small renal cell carcinoma. Beijing Da Xue Xue Bao Yi Xue Ban 2019;51:623-7. [Crossref] [PubMed]
9. Minervini A, Rosaria Raspollini M, Tuccio A, Di Cristofano C, Siena G, Salvi M, Vittori G, Sebastianelli A, Lapini A, Serni S, Carini M. Pathological characteristics and prognostic effect of peritumoral capsule penetration in renal cell carcinoma after tumor enucleation. Urol Oncol 2014;32:50.e15-22. [Crossref] [PubMed]
10. Song T, Yin Y, Liao B, Zheng S, Wei Q. Capsular invasion in renal cell carcinoma: a meta-analysis. Urol Oncol 2013;31:1321-6. [Crossref] [PubMed]
11. Klatte T, Chung J, Leppert JT, Lam JS, Pantuck AJ, Figlin RA, Belldegrun AS. Prognostic relevance of capsular involvement and collecting system invasion in stage I and II renal cell carcinoma. BJU Int 2007;99:821-4. [Crossref] [PubMed]
12. May M, Brookman-Amissah S, Roigas J, Gilfrich CP, Pflanz S, Hoschke B, Gunia S. Evaluation of renicapsular involvement in Stages I and II renal cell carcinoma from the morphological and prognostic point of view. Urol Oncol 2010;28:274-9. [Crossref] [PubMed]
13. Pierorazio PM, Johnson MH, Ball MW, Gorin MA, Trock BJ, Chang P, Wagner AA, McKiernan JM, Allaf ME. Five-year analysis of a multi-institutional prospective clinical trial of delayed intervention and surveillance for small renal masses: the DISSRM registry. Eur Urol 2015;68:408-15. [Crossref] [PubMed]
14. Zhang H, Wu Y, Xue W, Zuo P, Oesingmann N, Gan Q, Huang Z, Wu M, Hu F, Kuang M, Song B. Arterial spin labelling MRI for detecting pseudocapsule defects and predicting renal capsule invasion in renal cell carcinoma. Clin Radiol 2017;72:936-43. [Crossref] [PubMed]
15. Zhang Y, Tian H, Zhang S, Zhang Q, Wu X. Multislice spiral computed tomography signs of invasion of the renal capsule by renal cell carcinoma. Medicine (Baltimore) 2018;97:e13075. [Crossref] [PubMed]
16. Kumar V, Gu Y, Basu S, Berglund A, Eschrich SA, Schabath MB, Forster K, Aerts HJ, Dekker A, Fenstermacher D, Goldgof DB, Hall LO, Lambin P, Balagurunathan Y, Gatenby RA, Gillies RJ. Radiomics: the process and the challenges. Magn Reson Imaging 2012;30:1234-48. [Crossref] [PubMed]
17. Cobo M, Menéndez Fernández-Miranda P, Bastarrika G, Lloret Iglesias L. Enhancing radiomics and Deep Learning systems through the standardization of medical imaging workflows. Sci Data 2023;10:732. [Crossref] [PubMed]
18. Xi IL, Zhao Y, Wang R, Chang M, Purkayastha S, Chang K, Huang RY, Silva AC, Vallières M, Habibollahi P, Fan Y, Zou B, Gade TP, Zhang PJ, Soulen MC, Zhang Z, Bai HX, Stavropoulos SW. Deep Learning to Distinguish Benign from Malignant Renal Lesions Based on Routine MR Imaging. Clin Cancer Res 2020;26:1944-52. [Crossref] [PubMed]
19. Klontzas ME, Kalarakis G, Koltsakis E, Papathomas T, Karantanas AH, Tzortzakakis A. Convolutional neural networks for the differentiation between benign and malignant renal tumors with a multicenter international computed tomography dataset. Insights Imaging 2024;15:26. [Crossref] [PubMed]
20. Erdim C, Yardimci AH, Bektas CT, Kocak B, Koca SB, Demir H, Kilickesmez O. Prediction of Benign and Malignant Solid Renal Masses: Machine Learning-Based CT Texture Analysis. Acad Radiol 2020;27:1422-9. [Crossref] [PubMed]
21. Zhou L, Zhang Z, Chen YC, Zhao ZY, Yin XD, Jiang HB. A Deep Learning-Based Radiomics Model for Differentiating Benign and Malignant Renal Tumors. Transl Oncol 2019;12:292-300. [Crossref] [PubMed]
22. Pedersen M, Andersen MB, Christiansen H, Azawi NH. Classification of renal tumour using convolutional neural networks to detect oncocytoma. Eur J Radiol 2020;133:109343. [Crossref] [PubMed]
23. Grajo JR, Batra NV, Bozorgmehri S, Magnelli LL, O'Malley P, Terry R, Su LM, Crispen PL. Association between nuclear grade of renal cell carcinoma and the aorta-lesion-attenuation-difference. Abdom Radiol (NY) 2021;46:5629-38. [Crossref] [PubMed]
24. Xv Y, Lv F, Guo H, Zhou X, Tan H, Xiao M, Zheng Y. 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;12:170. [Crossref] [PubMed]
25. Cui E, Li Z, Ma C, Li Q, Lei Y, Lan Y, Yu J, Zhou Z, Li R, Long W, Lin F. Predicting the ISUP grade of clear cell renal cell carcinoma with multiparametric MR and multiphase CT radiomics. Eur Radiol 2020;30:2912-21. [Crossref] [PubMed]
26. Nie P, Yang G, Wang Y, Xu Y, Yan L, Zhang M, Zhao L, Wang N, Zhao X, Li X, Cheng N, Wang Y, Chen C, Wang N, Duan S, Wang X, Wang Z. A CT-based deep learning radiomics nomogram outperforms the existing prognostic models for outcome prediction in clear cell renal cell carcinoma: a multicenter study. Eur Radiol 2023;33:8858-68. [Crossref] [PubMed]
27. Wang S, Zhu C, Jin Y, Yu H, Wu L, Zhang A, Wang B, Zhai J. A multi-model based on radiogenomics and deep learning techniques associated with histological grade and survival in clear cell renal cell carcinoma. Insights Imaging 2023;14:207. [Crossref] [PubMed]
28. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts HJWL. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
29. Tibshirani R. The lasso method for variable selection in the Cox model. Stat Med 1997;16:385-95. [Crossref] [PubMed]
30. Cho HJ, Kim SJ, Ha US, Hong SH, Kim JC, Choi YJ, Hwang TK. Prognostic value of capsular invasion for localized clear-cell renal cell carcinoma. Eur Urol 2009;56:1006-12. [Crossref] [PubMed]
31. Ha US, Lee KW, Jung JH, Byun SS, Kwak C, Chung J, Hwang EC, Kim YJ, Kwon TG, Kang SH, Hong SH. Renal capsular invasion is a prognostic biomarker in localized clear cell renal cell carcinoma. Sci Rep 2018;8:202. [Crossref] [PubMed]
32. Akbulut D, Kuz ED, Kursun N, Dizbay Sak S. Capsular Invasion Matters Also in "Papillary Patterned" Tumors: A Study on 121 Cases of Encapsulated Conventional Variant of Papillary Thyroid Carcinoma. Endocr Pathol 2021;32:357-67. [Crossref] [PubMed]
33. Chen J, Li Z, Han Z, Kang D, Ma J, Yi Y, Fu F, Guo W, Zheng L, Xi G, He J, Qiu L, Li L, Zhang Q, Wang C, Chen J. Prognostic value of tumor necrosis based on the evaluation of frequency in invasive breast cancer. BMC Cancer 2023;23:530. [Crossref] [PubMed]
34. Jeong IG, Jeong CW, Hong SK, Kwak C, Lee E, Lee SE. Prognostic implication of capsular invasion without perinephric fat infiltration in localized renal cell carcinoma. Urology 2006;67:709-12. [Crossref] [PubMed]
35. Yang L, Gao L, Arefan D, Tan Y, Dan H, Zhang J. A CT-based radiomics model for predicting renal capsule invasion in renal cell carcinoma. BMC Med Imaging 2022;22:15. [Crossref] [PubMed]
36. Blachura T, Radzikowska J, Matusik PS, Kowal A, Jarczewski JD, Skiba Ł, Popiela TJ, Chrzan R. A comparison of effectiveness of the contrast enhanced computed tomography with magnetic resonance imaging in the differential diagnosis of clear cell renal carcinoma. Radiol Oncol 2025;59:193-202. [Crossref] [PubMed]
37. Qu J, Li X, Zhu P, Zhang W, Wang X, Zheng Z, Song X, Dai C, Ma H, Zhou J. Peritumoral Cortex Low-Enhancement Sign on Corticomedullary Phase CT: A Distinctive Indicator of Small Renal Malignancies. Korean J Radiol 2025;26:938-50. [Crossref] [PubMed]
38. Tanaka H, Fukawa Y, Yamamoto K, Tanimoto K, Takemoto A, Mori T, Hasumi H, Kinoshita M, Kanazawa T, Furukawa A, Kimura K, Sato H, Hirakawa A, Fukuda S, Waseda Y, Yoshida S, Campbell SC, Fujii Y. Prognostic Impact and Genomic Backgrounds of Renal Parenchymal Infiltration or Micronodular Spread in Nonmetastatic Clear Cell Renal Cell Carcinoma. Mod Pathol 2024;37:100590. [Crossref] [PubMed]
39. Halefoglu AM, Ozagari AA. Tumor grade estımatıon of clear cell and papıllary renal cell carcınomas usıng contrast-enhanced MDCT and FSE T2 weıghted MR ımagıng: radıology-pathology correlatıon. Radiol Med 2021;126:1139-48. [Crossref] [PubMed]
40. Uhm KH, Jung SW, Hong SH, Ko SJ. Lesion-aware cross-phase attention network for renal tumor subtype classification on multi-phase CT scans. Comput Biol Med 2024;178:108746. [Crossref] [PubMed]
41. Anush A, Rohini G, Nicola S. WalaaEldin EM, Eranga U. Deep-learning-based ensemble method for fully automated detection of renal masses on magnetic resonance images. J Med Imaging (Bellingham) 2023;10:024501. [Crossref] [PubMed]
42. Maddalo M, Bertolotti L, Mazzilli A, Flore AGM, Perotta R, Pagnini F, Ziglioli F, Maestroni U, Martini C, Caruso D, Ghetti C, De Filippo M. Small Renal Masses: Developing a Robust Radiomic Signature. Cancers (Basel) 2023;15:4565. [Crossref] [PubMed]
43. Nassiri N, Maas M, Cacciamani G, Varghese B, Hwang D, Lei X, Aron M, Desai M, Oberai AA, Cen SY, Gill IS, Duddalwar VA. A Radiomic-based Machine Learning Algorithm to Reliably Differentiate Benign Renal Masses from Renal Cell Carcinoma. Eur Urol Focus 2022;8:988-94. [Crossref] [PubMed]
44. Feng S, Gong M, Zhou D, Yuan R, Kong J, Jiang F, Zhang L, Chen W, Li Y. A CT-based radiomics nomogram for differentiation of benign and malignant small renal masses (≤4 cm). Transl Oncol 2023;29:101627. [Crossref] [PubMed]
45. Yang R, Wu J, Sun L, Lai S, Xu Y, Liu X, Ma Y, Zhen X. Radiomics of small renal masses on multiphasic CT: accuracy of machine learning-based classification models for the differentiation of renal cell carcinoma and angiomyolipoma without visible fat. Eur Radiol 2020;30:1254-63. [Crossref] [PubMed]
46. Lin J, Ou Y, Luo M, Jiang X, Cen S, Zeng G. Combining clinical characteristics with CT radiomics to predict Ki67 expression level of small renal mass based on artificial intelligence algorithms. Front Oncol 2025;15:1541143. [Crossref] [PubMed]
47. Gao Y, Wang X, Zhao X, Zhu C, Li C, Li J, Wu X. Multiphase CT radiomics nomogram for preoperatively predicting the WHO/ISUP nuclear grade of small (< 4 cm) clear cell renal cell carcinoma. BMC Cancer 2023;23:953. [Crossref] [PubMed]
48. Zhang Y, Zhao J, Li Z, Yang M, Ye Z. Preoperative prediction of renal fibrous capsule invasion in clear cell renal cell carcinoma using CT-based radiomics model. Br J Radiol 2024;97:1557-67. [Crossref] [PubMed]