Multimodal artificial intelligence model based on CT for differentiating primary renal sarcomas from renal cell carcinomas: a dual-center retrospective study

Introduction

Adult primary renal sarcomas are rare renal malignancies derived from the mesenchymal tissue of the kidney. Given the rarity of renal sarcoma, the current epidemiologic data are lacking. In previous studies, the prevalence of renal sarcoma was approximately 1% of primary renal malignancies (1-3). As the second most common genitourinary sarcomas, the renal sarcomas accounted for 2% of all soft tissue sarcomas (4).

Renal cancer is one of the top ten most common cancers worldwide, accounting for 2.2% of all malignant tumors. As the most common solid tumor of the kidney, renal cell carcinoma (RCC) accounts for about 80–90% of primary malignant tumors of the kidney (5,6). For early-stage patients such as localized and locally advanced RCCs, nephron sparing surgery is the preferred treatment to preserve kidney function better (7-9).

The prognosis of patients with renal sarcomas is worse than that of RCCs due to the rapid growth and high malignancy (2). The treatment strategy for primary renal sarcoma typically involves radical nephrectomy instead of nephron sparing surgery in order to reduce positive surgical margins (10-12). Therefore, accurately diagnosing primary renal sarcoma can assist in preoperative clinical decision making and improve patient prognosis, which is an important issue to be solved in clinical practice. But the preoperative differential diagnosis between primary renal sarcoma and RCC before surgery is difficult, especially for the lager diameter tumors.

In recent years, abdominal imaging has been widely used in health examinations, and the detective rate and diagnostic accuracy (ACC) of renal tumors have significantly improved (13). Computed tomography (CT) is the cornerstone of imaging method for screening and diagnosis of renal mass. Based on the imaging features of sarcoma summarized in previous researches, Uhlig et al. found that the imaging features associated with renal sarcoma were: tumor laterality (right side), larger maximum diameter, irregular shape, ill-defined margins, vascular invasion, necrosis, and organ invasion (14-17). However, the studies above did not propose standard imaging evaluation method and neglected the differentiation between RCC and renal sarcoma. Moreover, the high-throughput information of images needs further exploration.

The application of artificial intelligence (AI) and radiomics in medical image analysis has been developed rapidly. Radiomics can be used for disease screening, diagnosis, prognosis analysis and decision making (18,19). Many studies have demonstrated that deep learning (DL) can be useful for image interpretation tasks and the models showed expert-level performance (20-22). Compared with traditional feature engineering-based method, DL radiomics can learn features automatically without precise tumor annotation and distinguish the features adaptive to specific task though self-learning strategy.

Therefore, we aimed to establish a DL model to distinguish adult primary renal sarcoma and RCC preoperatively, which can assist in formulating individualized treatment strategies for adult primary renal sarcoma patients. We present this article in accordance with the TRIPOD reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2025-539/rc).

Methods

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The First Affiliated Hospital, Sun Yat-sen University (FAH-SYSU) (No. [2022] 155) and the Ethics Committee of Sun Yat-sen University Cancer Center (No. B2021-314-01). Consent to participate was waived by the Ethics Committees.

Patient cohort

Patients pathologically diagnosed with renal sarcoma at FAH-SYSU and Sun Yat-sen University Cancer Center (SYSUCC) from 2009 to 2021 were retrospectively included in this study. The pathological subtypes of renal sarcoma were based on World Health Organization classification of renal tumors (23). Participants were excluded if they had previous history of sarcoma, presented with renal metastasis of non-renal originated sarcomas, with unsatisfied images quality, and of age less than 18 or more than 80. The control group randomly included from consecutive patients who were initially diagnosed from 2018 to 2020, with a pathological diagnosis of RCC (in order to match the sample sizes between two groups). Cystic renal mass was excluded. The flowchart of patients’ enrollment is shown in Figure 1.

Figure 1 Flowchart of the patient enrollment. CT, computed tomography; RCC, renal cell carcinoma.

Data collection and analysis

Demographic data and pathological findings were collected from the medical records of all patients. The estimated glomerular filtration rate (mL/min/1.73 m²), were calculated by the Cockcroft-Gault formula to assess the renal function (24).

CT images within 1 month before intervention were collected. The CT scanners and corresponding scanning parameters used at each institution are detailed in Table S1. The CT examinations were performed in two centers with the same protocol: an unenhanced phase (UP) and three standard enhanced phases CT was performed, including corticomedullary phase (CMP), and nephrographic phase (NP) and excretory phase (EP). The three standard enhanced phase images were acquired at 30, 70 and 180 s, respectively, after the injection of contrast agent (1.5 mL/kg, 3.0–4.0 mL/s, Ultravist, Bayer Schering Pharma, Berlin, Germany) via a pump injector. Due to the focus of this study was on the tumor area in CT images, we took UP, CMP and NP into analysis. Image analysis was performed by two physicians with more than 3 years of experience in abdominal imaging independently. The readers were blind to patient’s inclusion process, clinical and histopathological information. Renal tumor radiological features were evaluated according to the 13 features, including the intratumoral arteries, defined as visible arteries of the mass, which were summarized according to previous studies (14-17).

Image preprocessing

The preprocessing steps performed before DL model construction including format conversion, pixel resizing, tumor segmentation, normalization and augmentation. MicroDicom (version 3.8.1) software was used to convert image files into consecutive joint photographic experts group (JPEG) images with original pixel values extracted and mapped to JPEG color space to avoid information loss. We extracted images at the corresponding tumor levels from the 3-phase CT images of the same patient, while discarding any additional images at the upper or lower region of the tumor. This ensured that we maintained the balance of the 3-phase CT image datasets. Pixel resizing, Z-score standardization, and image normalization (pixel values were transformed to a distribution with a mean of 0 and standard deviation of 1) were applied to eliminate baseline intensity shifts across different imaging devices. Tumor segmentation with the region of interest outline applied to renal mass regions of the images, removing meaningless backgrounds to improve the model performance (delineated by H.L. under the supervision of H.W., a senior radiologist and further reviewed and fine-tuned by another high-seniority radiologist Y.G.). The image data normalization and augmentation such as image flipping were performed to improve data quality and enrich the data sets.

Model development and validation

In this study, we used residual neural network 34 (ResNet-34) for model development (25). The DL models was trained with the image data sets obtained after image preprocessing (the parameters were listed in the Table S2). To evaluate the performance of the model across different CT modalities, we input CT images from different dataset either individually or in combination. Hence, 7 models were established: (I) 3 single-mode models trained with image dataset separately; (II) 3 bimodal models trained with two randomly combined data sets; (III) a trimodal model trained with all 3 CT data sets. The study design and basic network structure is showed in Figure 2.

Figure 2 Study design and basic structure of the deep learning models. Cov, convolutional layer; fc, full connecting layer.

We trained and validated the DL model according to the 5-fold cross-validation (CV) strategy. Since the neural network in our models was an end-to-end images classifier and each patient has multiple consecutive CT slices from whole tumors, all CT slices from a single patient were assigned to either the training or validation set as a whole, then we took the mean of all the image-level results from each patient as a patient-level prediction. The diagnostic performance of DL models was evaluated by receiver operating characteristic (ROC) curve, area under the curve (AUC).

Statistical analysis

All analyses were performed using SPSS (version 25.0) and R studio (version 2022.02.0+443). The Shapiro-Wilk test was used to assessing the normality of the continuous variables. Continuous variables were tested using two independent samples t-test and Mann Whitney U test. While categorical variables were compared using Chi-squared test or Fisher’s exact test. Univariate and multivariate logistic regression (forward steps selection) was used to filter independent indicators of sarcoma and RCC, and a clinical model was constructed. The maximized Youden index was used to determine the optimal threshold and calculate sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) from the ROC curve. The difference was considered statistically significant at a two-side P<0.05.

Results

Patient characteristics

In total, 85 patients were retrospectively included in this study, including 27 renal sarcomas and 58 RCCs from the FAH-SYSU and the SYSUCC. The baseline characteristics were shown in Table 1, the most common histological types of renal sarcoma were liposarcoma (n=9, 33.3%), followed by leiomyosarcoma (n=8, 29.6%). The most common pathological type of RCCs was clear cell renal carcinoma (ccRCC; n=37, 63.8%), followed by chromophobe RCC (n=11, 18.9%). As detailed in Table 2, there was no statistical difference in the demographic distribution of renal sarcoma and RCC patients.

Clinical characteristics and radiological features

The distribution of preoperative clinical data and radiological features among different groups were shown in Table 2. Between renal sarcoma and RCC cohorts, white blood cell count (P=0.03), tumor size (P=0.01), intratumoral arteries (P=0.03), Gerota’s fascia invasion (P=0.001), perinephric invasion (P=0.001) showed significant difference.

Development and validation of models

According to the multivariate analysis in Figure 3A, intratumoral arteries [odds ratio (OR) =0.27, 95% confidence interval (CI): 0.08–0.93, P=0.04] and Gerota’s fascia invasion (OR =4.47, 95% CI: 1.37–14.57, P=0.01) were independent associated with renal sarcoma. When intratumoral arteries used, the diagnostic model yielded an AUC =0.63 (95% CI: 0.50–0.75), and another model developed by Gerota’s fascia invasion yielded an AUC =0.71 (95% CI: 0.59–0.83). While the indicators above were combined to develop a clinical model (the full parameters of equation were shown in Table S3), the AUC was 0.77 (95% CI: 0.66–0.87). The ROC curve analysis was showed in Figure 3B, the calibration curve and decision curve analysis were illustrated in Figure S1. To test the robustness of the model, we used 5-fold CV and Bootstrap resampling (1,000 iterations), yielding AUC of 0.79±0.11 and 0.78 (95% CI: 0.70–0.85) respectively (5-fold CV AUC values were 0.72, 0.77, 0.80, 0.81, and 0.85, respectively), the details of confusion metrics were in Table S4. The cutoff value of the clinical model was 0.28, the sensitivity, specificity, PPV and NPV of the model was 0.74, 0.67, 0.51 and 0.85 respectively. The metrics of the model adjusted by tumor size are detailed in the Table S5. The tumor size across folds was showed on Table S6.

Figure 3 Indicators and diagnostic performance of the clinical model. (A) Multivariate Logistic regression for screening independent indicators of renal sarcoma. (B) ROC curve, AUC and 95% CI. AUC, area under the curve; CI, confidence interval; OR, odds ratio; ROC, receiver operating characteristic; WBC, white blood cell count.

Totally, 7,482 multimodal CT images were obtained and used to develop and validate the AI models. The performances of the 7 different DL models were validated by 5-fold CV, and the results were presented in Table 3. The scatter plots were utilized to visualize the ACC and AUC distribution in each fold (Figure 4). Overall, the AUC results of 7 models were in the range of 0.90 to 0.95. The overall ACC of the 7 models for differentiating renal sarcoma and RCC were in the range of 0.85 to 0.95. We compared the diagnostic performance between the UP model and radiologists assisted by the clinical model, with detailed data available in the Table S7.

Figure 4 Performance of deep learning models trained on multimodal CT images based on 5-fold cross-validation. Scatter plot showed the ACC (A) and AUC (B) results in each fold. ACC, accuracy; AUC, area under the curve; CMP, corticomedullary phase; CT, computed tomography; DL, deep learning; NP, nephrographic phase; UP, unenhanced phase.

Regarding single-mode models, the UP model could reach a satisfied diagnostic performance, yielding AUC of 0.95±0.09, with ACC of 0.94±0.07, sensitivity of 0.95±0.05 and specificity of 0.93±0.15. Contrary to our expectations, the models trained by contrast-enhanced CT images did not demonstrate improved classification performance. The CMP and NP model produced AUC of 0.94±0.07 and 0.91±0.13 respectively. There were no significant differences between the mean of three models (P=0.80).

Regarding bimodal and trimodal models, the UP + NP model yielded nearly AUC of 0.95±0.06, and ACC of 0.94±0.07. Although the mean of the model was not statistically different from UP model (P>0.99), the combined data seemed to enhance the stability and robustness of the model. The other three kind of combination did not show satisfied improved effects, with the best AUC of 0.94±0.10 and ACC of 0.94±0.08 (CMP + NP model). Overall, Utilization of enhanced phase CT images or combinations of different CT modalities did not significantly improve the performance.

To better understand the classification process and enhance the interpretability of DL model, we visualized the features learned by the model with heat map techniques. The heat map revealed the attention of the model. We chose the UP + CMP model and used the method of gradient class activation mapping (Grad-CAM) to generate heat maps for images input. We found that the values within the tumor were higher than the surrounding area for both renal sarcoma and RCC, with renal sarcoma having higher values. Examples of typical CT slides and heat maps of renal sarcoma and RCC are illustrated in Figure 5.

Figure 5 Examples of typical CT images and corresponding heat maps of renal sarcoma and RCC. The red boxes indicate the tumor region (region of interest). Histopathology images: hematoxylin and eosin staining, 5× magnification. CMP, corticomedullary phase; CT, computed tomography; RCC, renal cell carcinoma; RS, renal sarcoma; UP, unenhanced phase.

Discussion

Renal sarcoma is a rare malignant tumor, with an incidence rate of approximately 1% of all renal tumors (2,26,27). Compared to RCC, renal sarcomas are characterized by high malignancy and poor prognosis, requiring more aggressive surgical intervenes. Therefore, it is vital to differentiate renal sarcoma from RCC, especially the larger diameter RCC, preoperatively. However, there are limited systemic studies on the imaging characteristics and diagnostic methods for renal sarcoma. Hence, a novel method for differential diagnosis of renal sarcoma is needed.

In this study, we collected clinical and CT imaging data from 27 renal sarcomas and 58 RCCs, obtaining a total of 7,482 kidney tumor CT images. We analyzed the clinical and imaging data to investigate the differences in clinical factors and radiological features between renal sarcoma and RCC, and construct and validate a diagnostic model for renal sarcoma. The clinical model showed a good performance in differentiating renal sarcoma from RCC, with an AUC of 0.79. The multimodal DL models further improved the diagnostic performance, with AUC results in range of 0.90–0.95 by 5-fold CV. The results indicated the good discrimination of the DL models.

We found that the intratumoral arteries and Gerota’s fascia invasion were independent indicators and highly suggestive of renal sarcoma. Previous studies have found that CT/magnetic resonance imaging features, including tumor laterality (right side), larger maximum diameter, irregular shape, ill-defined margins, vascular invasion, necrosis, and organ invasion, are associated with renal sarcomas. In this research, absence of intratumoral arteries was associated with renal sarcoma, which may due to the fact that the blood supply of renal sarcomas often comes from extracapsular vessels (26). The rapid growth rate of renal sarcoma causing relatively lack of blood supply, resulting in hypoxia and starve of the tumor cells, may be the reason of neovascularization outside the tumor. The Gerota’s fascia invasion as an indicator showed biological behavior of infiltrating surrounding tissues, which be attributable to highly malignancy of renal sarcoma. Perinephric invasion, which also indicates tumor invasiveness, showed differences but was not included in the clinical model. After excluding the correlation between perinephric invasion and Gerota’s fascia invasion, we determined that the contribution of the former factor is less than that of the latter one. Notably, no statistically significant differences in tumor configuration and margins were observed in our cohort. This may be attributed to population heterogeneity, and could also be due to biases in the manual interpretation of imaging features. Nevertheless, given the role of tumor configuration in the diagnosis of soft-tissue sarcomas and its value in indicating malignancy grade, we will continue to focus on tumor configuration and further refine our research in larger cohorts in subsequent studies (28,29). Given the significant difference in tumor size between two groups, and a multivariate analysis P value of 0.054, we hypothesized it might act as a confounding factor. Adjusting the clinical model for tumor size yielded minimal changes in both AUC (0.80, 95% CI: 0.71–0.91) and the ORs of key features compared to the original model, confirming that tumor size does not a meaningful confounding factor.

Currently, there are no non-invasive clinical or radiomics methods available that can provide similar diagnostic performance to this study. In previous studies, Uhlig et al. used machine learning to analyze clinical and radiological features to distinguish renal sarcoma from all other renal tumors (14). However, the model is limited due to deficiency of deep-level exploration in CT images and the incomplete inclusion of CT modes. In recent years, numerous studies have indicated that DL methods can achieve expert-level performance in image feature analysis tasks. As an advanced training framework, the ResNet convolutional neural network has yielded satisfactory results in fields such as cranial nerve imaging and musculoskeletal imaging (20-22,30-32). Our study used DL algorithms to explore the roles of different modalities of CT scans and further investigated the integration of multi-modalities of CT images to make the study more comprehensive. The utilization of end-to-end DL algorithm reduce the need for manual feature engineering and allows for more accurate and robust representations of the data.

In multimodal CT models, the enhanced CT models did not achieve higher AUC than unenhanced model, which can be explained by two key factors. First, our DL models, as end-to-end system, autonomously mined intrinsic discriminative features from unenhanced images such as tumor morphology, intratumoral density, and peritumoral structure relationships, which already have sufficient diagnostic value without relying on contrast signals. Second, the unstable quality of the contrast enhanced phase driven by various factors including renal vascular status, renal function, contrast agent concentration speed and inter-institutional difference may also explain this observation. Compared to single-mode models, the combination of multimodal CT images did not significantly improve the diagnostic performance of the models. In previous renal radiomics research, the selection of CT images was often based on a single modality. For example, Gomes et al. found that a single-mode model trained on CMP images had comparable ACC to models built using images from all phases in the classification task of RCC subtypes (33). Yan et al. also proposed that the CMP was the best choice for model construction in the task of classifying ccRCC and papillary RCC (pRCC) (34).

There are several limitations in this study. First, this study was exploratory study due to lacking of research on factors related to renal sarcoma, and the results of this study need further validation. Further, the sample size of this study was small, which may restrict statistical power and raise concerns about potential model overfitting. This constraint is a result of the intrinsic rarity of primary renal sarcomas, making large-cohort accumulation extremely challenging even in dual-center retrospective studies. Therefore, we implemented multiple targeted strategies like using forward steps selection in logistic regression control variables including, employing the shallow variant ResNet-34, L2 regularization and CV to mitigate the risks. Finally, the clinical application of the model in this study remained uncertain, which require further test. In future research, we will continue to collect cases, actively promote the implementation of multicenter studies to expand the cohort and establish independent external and prospective validation set, and simultaneously integrate multi-modal data (e.g., pathological and genetic data) into the model framework—aiming to further verify the model’s applicability across diverse populations and clinical settings while enhancing its robustness through multi-dimensional data support.

Conclusions

This study achieved effective differential diagnosis between renal sarcoma and RCC. Despite the small sample size in this study, the ResNet-34 network models based on multimodal CT images showed strong performance and generalization ability by 5-fold CV. The DL models established in this study might aid in accurate preoperative diagnosis of renal sarcoma and assist in managing the treatment of renal cancer patients.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tau.amegroups.com/article/view/10.21037/tau-2025-539/rc

Data Sharing Statement: Available at https://tau.amegroups.com/article/view/10.21037/tau-2025-539/dss

Peer Review File: Available at https://tau.amegroups.com/article/view/10.21037/tau-2025-539/prf

Funding: This study was supported in part by the National Natural Science Foundation of China (Nos. 82373433, 81725016, 81872094, 82272862, and 81902576), Guangdong Provincial Department of Finance Project in 2022 (No. KS0120220267), and the Guangzhou Science and Technology Projects (No. 202201010910).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tau.amegroups.com/article/view/10.21037/tau-2025-539/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. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The First Affiliated Hospital, Sun Yat-sen University (FAH-SYSU) (No. [2022] 155) and the Ethics Committee of Sun Yat-sen University Cancer Center (No. B2021-314-01). Consent to participate was waived by the Ethics Committees.

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. Dotan ZA, Tal R, Golijanin D, et al. Adult genitourinary sarcoma: the 25-year Memorial Sloan-Kettering experience. J Urol 2006;176:2033-8; discussion 2038-9. [Crossref] [PubMed]
2. Wang X, Xu R, Yan L, et al. Adult renal sarcoma: clinical features and survival in a series of patients treated at a high-volume institution. Urology 2011;77:836-41. [Crossref] [PubMed]
3. Alevizakos M, Gaitanidis A, Korentzelos D, et al. Renal Sarcoma: A Population-Based Study. Clin Genitourin Cancer 2023;21:155-61. [Crossref] [PubMed]
4. Samaratunga H, Delahunt B. Mesenchymal tumors of adult kidney. Semin Diagn Pathol 2015;32:160-71. [Crossref] [PubMed]
5. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
6. Siegel RL, Miller KD, Wagle NS, et al. Cancer statistics, 2023. CA Cancer J Clin 2023;73:17-48. [Crossref] [PubMed]
7. Huang WC, Elkin EB, Levey AS, et al. Partial nephrectomy versus radical nephrectomy in patients with small renal tumors--is there a difference in mortality and cardiovascular outcomes? J Urol 2009;181:55-61; discussion 61-2. [Crossref] [PubMed]
8. Van Poppel H, Da Pozzo L, Albrecht W, et al. A prospective, randomised EORTC intergroup phase 3 study comparing the oncologic outcome of elective nephron-sparing surgery and radical nephrectomy for low-stage renal cell carcinoma. Eur Urol 2011;59:543-52. [Crossref] [PubMed]
9. Ljungberg B, Albiges L, Abu-Ghanem Y, et al. European Association of Urology Guidelines on Renal Cell Carcinoma: The 2022 Update. Eur Urol 2022;82:399-410. [Crossref] [PubMed]
10. Alvarenga JC, Ball AB, Fisher C, et al. Limitations of surgery in the treatment of retroperitoneal sarcoma. Br J Surg 1991;78:912-6. [Crossref] [PubMed]
11. Bevilacqua RG, Rogatko A, Hajdu SI, et al. Prognostic factors in primary retroperitoneal soft-tissue sarcomas. Arch Surg 1991;126:328-34. [Crossref] [PubMed]
12. Heslin MJ, Lewis JJ, Nadler E, et al. Prognostic factors associated with long-term survival for retroperitoneal sarcoma: implications for management. J Clin Oncol 1997;15:2832-9. [Crossref] [PubMed]
13. Roy C, Buy X, Ghali SE. Imaging in Renal Cell Cancer. Eau Update 2003;1:209-14. [Crossref]
14. Uhlig J, Uhlig A, Bachanek S, et al. Primary renal sarcomas: imaging features and discrimination from non-sarcoma renal tumors. Eur Radiol 2022;32:981-9. [Crossref] [PubMed]
15. Karaosmanoğlu AD, Onur MR, Shirkhoda A, et al. Unusual Malignant Solid Neoplasms of the Kidney: Cross-Sectional Imaging Findings. Korean J Radiol 2015;16:853-9. [Crossref] [PubMed]
16. Lalwani N, Prasad SR, Vikram R, et al. Pediatric and adult primary sarcomas of the kidney: a cross-sectional imaging review. Acta Radiol 2011;52:448-57. [Crossref] [PubMed]
17. Katabathina VS, Vikram R, Nagar AM, et al. Mesenchymal neoplasms of the kidney in adults: imaging spectrum with radiologic-pathologic correlation. Radiographics 2010;30:1525-40. [Crossref] [PubMed]
18. Kumar V, Gu Y, Basu S, et al. Radiomics: the process and the challenges. Magn Reson Imaging 2012;30:1234-48. [Crossref] [PubMed]
19. Lambin P, Rios-Velazquez E, Leijenaar R, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012;48:441-6. [Crossref] [PubMed]
20. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 2017;542:115-8. [Crossref] [PubMed]
21. Ting DSW, Cheung CY, Lim G, et al. Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations With Diabetes. JAMA 2017;318:2211-23. [Crossref] [PubMed]
22. Wang K, Lu X, Zhou H, et al. Deep learning Radiomics of shear wave elastography significantly improved diagnostic performance for assessing liver fibrosis in chronic hepatitis B: a prospective multicentre study. Gut 2019;68:729-41. [Crossref] [PubMed]
23. Moch H, Cubilla AL, Humphrey PA, et al. The 2016 WHO Classification of Tumours of the Urinary System and Male Genital Organs-Part A: Renal, Penile, and Testicular Tumours. Eur Urol 2016;70:93-105. [Crossref] [PubMed]
24. Cockcroft DW, Gault MH. Prediction of creatinine clearance from serum creatinine. Nephron 1976;16:31-41. [Crossref] [PubMed]
25. He K, Zhang X, Ren S, et al., editors. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2016.
26. Srinivas V, Sogani PC, Hajdu SI, et al. Sarcomas of the kidney. J Urol 1984;132:13-6. [Crossref] [PubMed]
27. Vogelzang NJ, Fremgen AM, Guinan PD, et al. Primary renal sarcoma in adults. A natural history and management study by the American Cancer Society, Illinois Division. Cancer 1993;71:804-10. [Crossref] [PubMed]
28. Sedaghat S, Salehi Ravesh M, Sedaghat M, et al. Does the primary soft-tissue sarcoma configuration predict configuration of recurrent tumors on magnetic resonance imaging? Acta Radiol 2022;63:642-51. [Crossref] [PubMed]
29. Sedaghat S, Salehi Ravesh M, Sedaghat M, et al. Configuration of soft-tissue sarcoma on MRI correlates with grade of malignancy. Radiol Oncol 2021;55:158-63. [Crossref] [PubMed]
30. Noebauer-Huhmann IM, Vanhoenacker FM, Vilanova JC, et al. Soft tissue tumor imaging in adults: whole-body staging in sarcoma, non-malignant entities requiring special algorithms, pitfalls and special imaging aspects. Guidelines 2024 from the European Society of Musculoskeletal Radiology (ESSR). Eur Radiol 2025;35:351-9. [Crossref] [PubMed]
31. Crombé A, Kind M, Fadli D, et al. Soft-tissue sarcoma in adults: Imaging appearances, pitfalls and diagnostic algorithms. Diagn Interv Imaging 2023;104:207-20. [Crossref] [PubMed]
32. Stramare R, Beltrame V, Gazzola M, et al. Imaging of soft-tissue tumors. J Magn Reson Imaging 2013;37:791-804. [Crossref] [PubMed]
33. Veloso Gomes F, Matos AP, Palas J, et al. Renal cell carcinoma subtype differentiation using single-phase corticomedullary contrast-enhanced CT. Clin Imaging 2015;39:273-7. [Crossref] [PubMed]
34. Yan L, Liu Z, Wang G, et al. Angiomyolipoma with minimal fat: differentiation from clear cell renal cell carcinoma and papillary renal cell carcinoma by texture analysis on CT images. Acad Radiol 2015;22:1115-21. [Crossref] [PubMed]