A MRI-based radiomics nomogram for differentiation of fat-poor renal angiomyolipoma from nodular clear cell renal cell carcinoma

doi:10.21203/rs.3.rs-4121605/v1

Objectives

This study aims to assess the effectiveness of a radiomics nomogram based on multiparametric MRI in distinguishing between fat-poor renal angiomyolipoma (fp-AML) and renal clear cell carcinoma (ccRCC) in small renal tumors (≤ 4 cm).

Methods

A retrospective analysis was conducted from July 2017 to July 2023, including 65 ccRCC and 23 fp-AML patients with confirmed pathology. T2WI, DWI and ADC sequences were used to extract radiomics features. A t-test and LASSO were used to select features. A radiomic signature was created after calculating a radiomic score (Rad-score). To create a clinical model, clinical characteristics and subject MRI features were evaluated. Then, the Rad scores and clinical characteristics were integrated to develop a radiomics nomogram. The nomogram's performance was assessed regarding calibration, discrimination, and clinical utility.

Results

In the study, 88 renal masses from 86 patients were included. Each patient had their 3948 radiomics features retrieved, and 13 of those features were used to create the radiomics signature. The radiomics nomogram (AUC, 0.833; 95% CI, 0.675–0.992) showed superior discriminating capacity in the validation set compared to the clinical model (AUC, 0.774; 95% CI, 0.554–0.995) and radiomics signature (AUC, 0.778; 95% CI, 0.571–0.985) alone. Decision curve analysis indicated that the nomogram outperformed the clinical model in terms of clinical utility.

Conclusions

The multiparametric MRI-based radiomics nomograms effectively use Rad scores and clinical factors as non-invasive preoperative predictors, successfully differentiating fp-AML from ccRCC.

Fat-poor renal angiomyolipoma

Renal clear cell carcinoma

Magnetic resonance imaging

Radiomics

The detection of small renal masses (SRMs) is continuously rising due to the extensive utilization of imaging equipment [Feng et al. 2018]. Renal angiomyolipoma(AML)is usually characterized by macroscopic fat. However, in up to 5% of AML cases, imaging may not display visible fat. This difficulty in distinguishing between fp-AML and clear cell carcinoma (ccRCC) can result in misdiagnosis and unnecessary surgical treatment [Woo et al. 2017; Thiravit et al.2018; Sasaguri et al. 2018]. Studies showed that 2–6% of surgically removed kidney masses are AML [Hindman et al. 2012]. Percutaneous renal biopsy is invasive, subject to sampling error and is associated with complications, providing an incorrect diagnosis or failure to make a diagnosis in up to 20% of biopsy cases [Leon et al. 2019; Bhindi et al. 2018; Lim et al. 2019].

To differentiate fp-AML from ccRCC, radiological features including high attenuation at unenhanced CT, absence of pseudo-capsule, uniform enhancement pattern, low signal on T2-weighted imaging, and uniform high signal on diffusion-weighted imaging (DWI) images have been suggested [Li et al. 2020; Park et al. 2019]. However, the radiological features are subjective visual assessments, have poor inter-observer agreement, and limited reproducibility. Other researchers have attempted to develop quantitative methods to assess fp-AML better, but these findings are controversial. For instance, although some studies [Silva et al. 2021; Li et al. 2019] found a low ADC value can differentiate fp-AML from ccRCC, others [Ding et al. 2016] have indicated that there is no statistically significant difference in the ADC values of the two types. Similarly, quantitative assessment of signal intensity (SI) reduction on opposed phase T1-weighted image/chemical shift image has been considered an effective method for determining fp-AML [Sasiwimonphan et al. 2016]. However, Ferre et al. [Ferré et al. 2015] reported no difference in SI reduction between fp-AML and ccRCC. Since conventional CT and MRI often fail to provide accurate differentiation, there is a need for a more reliable method to reduce biopsies and unnecessary resections of benign renal masses.

Radiomics is a computer-aided technique designed to transform medical images into quantitative, mineable information, including histograms, textures, and shape features. Radiomics can extract large-scale imaging parameters that the human eye cannot observe, potentially improving diagnostic accuracy and prognosis. Ma Y et al. [Ma et al. 2020] demonstrates the superior performance of radiomics CT analysis compared to conventional CT analysis. With higher image contrast between soft tissues, MRI, in combination with radiomics, has the potential to deliver a more exhaustive evaluation of renal masses than traditional CT. As a result, our investigation endeavors to develop a nomogram that combines a radiomics model and clinical features, thus facilitating the preoperative distinction between fp-AML and ccRCC.

Patients

This retrospective study was approved by the ethics review board of our Hospital, with a waiver of informed patient consent. Data from 86 patients (88 renal masses) were collected for analysis, including 65 ccRCCs and 23 fp-AMLs. The study was conducted at our Hospital from July 2017 to July 2023. The inclusion and exclusion criteria, as shown in Fig. 1, were used for participant selection. The following were the specific inclusion criteria: (1) Patients confirmed to have AML or ccRCC based on a definitive pathological diagnosis.; (2) patients who received a routine MRI scan 15 days before surgery; (3) Patients with a full range of clinical and pathological data. The following were the exclusion standards: (1) MRI examination was on a 1.5 T MRI instrument; (2) patients with a mass larger than 4 cm or less than 1cm in diameter; (3) patients with poor image quality that was difficult to analyze; (4) patients with AML with intra-focal fat visible to the naked eye on MRI images.

MRI examination

The patients’ images were obtained using a Siemens 3.0T or Philips 3.0 T MRI scanner for examination. A 12-channel phased-array body coil was used to perform a plain scan of the kidney MRI after a 4- to 6-hour fast before scanning, covering both sides of the kidney. Among the T2WI scan parameters: TR 2000 to 6200 ms, TE 82 to 104 ms, FOV 40 cm × 40 cm2; the b value chosen for DWI in this study was 800 s/mm².

MRI feature evaluation

Two radiologists with 20 and 7 years of abdominal imaging experience independently reviewed the images of the study cohort, remaining blinded to histopathological details. These imaging characteristics were evaluated by the radiologists: tumor origination (left/ right renal), tumor location (upper, middle, or lower pole), maximal size on T2WI images, shape (regular/irregular), margin (smooth/nonsmooth), pseudocapsule (present/absent), extrarenal extension (absent/≤50% extension />50% extension), the T2WI, DWI signal (homogeneously hypointense/homogeneously hyperintense/ heterogeneous) and ADC values. All images were first reviewed by two radiologists individually, after which disagreements between the readers were discussed until consistent results were produced. Clinical data, including age and gender, were recorded simultaneously.

Construction of the clinical factor model

The clinical factors (including clinical data and MRI features) were compared between the two groups using univariate analysis. Subsequently, a multiple logistic regression analysis was employed to construct the clinical factors model, incorporating the significant variables identified from the univariate analysis as inputs.

Segmentation of tumors

Patients' images were exported in DICOM format and the resulting image data were imported into 3D slicer software for outlining the region of interest (ROI). Two experienced abdominal radiologists manually drew the ROI on a slice-by-slice sequence of T2WI, DWI and ADC maps. They had over five years of experience each and were blind to the clinical data and pathological findings.

Radiomics feature extraction and selection

The PyRadiomics package, an open-source Python package, was used to extract features from T2WI, DWI and ADC maps. Each patient's ROI was grouped by tumor type, and the radiomics features were extracted. There were seven image types (original, exponential, gradient, logarithm, square, squareroot, wavelet) and seven feature classes (shape features, first order features, gray level co-occurrence matrix (glcm) features, gray level dependence matrix(gldm) features, gray level run length matrix(glrlm) features, gray level size zone matrix(glszm) features, neigbouring gray tone difference matrix (ngtdm)) features adopted for each sequence. The details of the radiomics features are described at

(https://pyradiomics.readthedocs.io/en/2.1.2/features.html).

Construction and evaluation of the radiomics model

Before modeling, three feature selection steps were carried out to prevent overfitting. Initially, stable features with an intraclass correlation coefficient (ICC) exceeding 0.8 were chosen from the two sets of ROIs. Following this, a t-test was conducted for each stable feature. Finally, the features that demonstrated statistical significance (p < 0.05) in the t-test were evaluated using the least absolute shrinkage and selection operator (LASSO) regression algorithm.

The study population was divided into a training cohort and a test cohort in a stratified manner. The effectiveness of selected features was tested on a validation cohort using the logistic regression algorithm, with classification accuracy measured by the area under the ROC curve. A radiomics score was generated for each patient using these features and combined with major clinical characteristics to create a radiomics nomogram. The calibration and goodness of fit of the nomogram were evaluated using calibration curves and the Hosmer-Lemeshow test. The diagnostic performance of the clinical model, radiomics signature, and radiomics nomogram in distinguishing between fp-AMLs and ccRCCs was assessed based on AUC in both training and validation sets. To evaluate the clinical applicability of the radiomics nomogram in terms of net benefit at different threshold probabilities, a decision curve analysis was conducted. Figure 2 depicts a flow chart of the whole radiomics analysis.

Statistical analysis

R statistical software and SPSS 26 were used to conduct the statistical analysis. For continuous variables, the t-test or Mann-Whitney U-test was employed, and for categorical variables, the chi-square test or Fisher's exact test. Statistics were deemed to be different if p < 0.05.

Clinical variables and clinical model design

In our study group, there were 23 patients with fp-AMLs (26.1%) and 65 patients with ccRCCs (73.9%). Table 1 lists several clinical elements for patient training and validation. Age, gender, and T2WI signal were independent predictors according to multiple logistic regression. Elderly male patients or tumors showing heterogeneous signal on T2WI may have ccRCC.

The train set AUC and test set AUC for the clinical model, respectively, were 0.891(95%CI: 0.791 ~ 0.990) and 0.774(95%CI: 0.554–0.995).

Construction and evaluation of radiomics models

In this study, 1316 radiomics features were retrieved from T2, DWI or ADC maps. The three sequences generated 3948 radiomics characteristics, of which 2648 radiomics features were stable during ICC screening. After a t-test, 1035 features were selected, and ultimately, LASSO identified the 13 most important features.

The train set AUC and test set AUC for the radiomics model, respectively, were 0.923 (95%CI: 0.832–1.000) and 0.778 (95%CI: 0.571–0.985).

Radiomics Nomogram Construction and Evaluation

The radiomics nomogram (Fig. 3), was built using the age, gender, T2WI signal, and Rad-score, with an AUC of 0.974 (95%CI: 0.938–1.000) and 0.833(95% CI: 0.675–0.992) for the train set and the test set. Table 2 provides an overview of each model's diagnostic performance. The calibration curve is displayed in Fig. 4. Figures 5 show the AUC images of the clinical factors, radiomics model, and radiomics nomograms in the training and test sets, demonstrating that the radiomics nomogram outperforms the radiomics-only model and the clinical model in differentiating between small fp-AML and ccRCC. Figure 6 shows the Nomo-scores of each patient. Additionally, Fig. 7 illustrates the DCA Decision Curve.

AML and ccRCC are respectively the most common benign and malignant renal tumors. It is essential to distinguish between these two entities due to their distinct treatment approaches and prognoses. For ccRCC, common treatment methods involve surgical intervention and ablation therapy, whereas AML often involves observation and follow-up, with treatment being considered for patients with tumor diameters greater than 4cm or those experiencing tumor bleeding [Dell'Atti et al. 2022; Silson et al. 2021].

In clinical practice, CT and MRI are widely used as effective imaging assessment tools for renal tumors. The features that differentiate fat-poor AML and ccRCC include age, gender, CT density on plain scan, and T2WI signal at all [Tannka et al. 2017; Jomoto et al.2022]. The logistic regression analysis results of this study reveal that ccRCC is more prevalent in elderly male patients. Furthermore, ccRCC exhibits a more heterogeneous T2WI signal compared to fat-poor AML. This increased heterogeneity in ccRCC is attributed to its stronger tumor heterogeneity, making it more susceptible to cystic degeneration and necrosis. Consequently, ccRCC demonstrates a greater mixture of high and low abnormal signals on T2WI, consistent with previous research findings. However, DWI signal and ADC values were not included as an independent predictive factor in the clinical model. It is believed that this may be due to smaller ccRCC tumors often having lower malignancy, resulting in DWI features of these low-grade tumors being somewhat similar to lipid-poor AML. Therefore, the similarity between these features emphasizes the necessity of utilizing more detailed and advanced radiomics and machine learning techniques in tumor differentiation to enhance accuracy.

In recent years, the development and advancement of radiomics have significantly enhanced the ability to predict and classify renal tumors. CT and MRI imaging have been extensively utilized in radiomics and machine learning research to differentiate renal masses [Kim et al. 2022; Sun et al. 2022]. Given that T2, DWI, and ADC maps are widely adopted sequences in MRI examinations in various clinical settings, these three sequences were selected for feature extraction in the current study. A total of 3948 radiomic features were extracted from the T2WI, DWI and ADC sequences, from which 13 most significant features were ultimately identified to construct the radiomics model. In the test set, the model achieved an AUC value of 0.778, representing a substantial improvement compared to previous studies [Matsumoto et al. 2022; Arita et al. 2021] that had focused on extracting only a few dozen features. This expanded feature analysis permits a more thorough exploration of tumor information, thereby enhancing the predictive and classificatory capabilities of radiomics in the context of renal tumors.

Based on the content mentioned above, this study combined conventional MRI features and radiomics features to establish a radiomics nomogram with improved predictive value for distinguishing lipid-poor AML and small renal ccRCC in both the training and validation sets. The AUC values for the nomogram were 0.974 and 0.833, respectively, surpassing those of the standalone clinical model or the imaging radiomics model. This suggests that radiomics features complement conventional clinical and imaging features. The results align with Nie et al.'s [Nie et al. 2020] work, who developed a CT radiomics nomogram. By leveraging both CT and MRI advantages, combining information from both modalities proves to be more effective in distinguishing these entities. Jian et al. [Jian et al. 2022] conducted a study wherein they focused on the clinical value of urinary creatinine by using T2WI sequences and IVIM sequences alongside clinical data and MRI radiomics models to construct a radiomics nomogram for differentiating lipid-poor AML and RCC. Their clinical-radiomics model achieved a commendable AUC of 0.931, demonstrating superior performance compared to the current study. The reason behind this superiority lies in the ability of IVIM sequences to differentiate between the diffusion of water molecules and tissue microcirculation perfusion, thereby providing more accurate descriptions of cellular functional changes and tissue microstructure when compared to DWI sequences. Consequently, IVIM sequences offer a more comprehensive range of information [Malagi et al. 2022; Malagi et al.2023]. However, despite these advantages, IVIM sequences are not extensively utilized in clinical practice, thereby limiting their practical value. Future research could consider integrating other imaging features and clinical factors to further enhance the application effectiveness of radiomics nomograms. These studies further demonstrate that models built using multiple parameters have higher predictive efficacy. By comprehensively utilizing conventional imaging features and radiomics techniques, the ability to differentiate between these two types of small renal masses is improved, providing strong support for more precise treatment plans for patients.

In addition to radiomics models, deep learning models are also being integrated into the study of renal tumors and have achieved good predictive results [Xu et al. 2022; Xi et al. 2020]. The application of deep learning may bring more innovations to the study of kidney tumors, and when combined with traditional methods such as radiomics, it is expected to achieve better outcomes in clinical practice.

There were several limitations to this study. Firstly, the small sample size of this study, particularly for fp-AML, which corresponds to the low overall incidence of this subtype in the population. In addition, for solid renal masses, MRI is not often the preferred imaging modality, and most of the patients in our study had diagnostic difficulties with enhanced CT or ultrasound. Therefore, the potential for selection bias within the patient cohort of this retrospective study highlights the need for further prospective investigations to accurately assess the diagnostic significance in real-world clinical scenarios. Secondly, the data used to train the radiomics models was obtained on two MRI scanning instruments. Previous literature [Kalpathy et al. 2016; Timmeren et al. 2020] has reported that the CT scanner models affect the reproducibility of radiomics features. Thus, different MRI scanner models may also impact data analysis. However, it is essential to highlight that this variation would enable the creation of more generalized models. Future research opportunities may arise from the need to provide additional training or validation of machine learning models on an external dataset to account for this variability and assess its appropriateness. This would require extra effort in accommodating the variability and evaluating the applicability, thereby offering a potential avenue for further investigation.

To sum up, the research conducted in our study focused on the development of a radiomics nomogram based on multiparametric MRI. This nomogram demonstrated excellent predictive effectiveness in distinguishing fp-AML from small ccRCC before surgical intervention. Utilizing radiomics nomograms, which are quantitative and non-invasive, for clinical decision-making has the potential to be highly beneficial. However, it is essential to validate these findings extensively before implementing them widely in clinical practice.

Author Contribution

Conceptualization: Shun Chai and Zhanlong MaData curation: Shun Chai and Yawen Yang and Chuanxian MaFormal Analysis: Shun Chai and Yawen YangFunding acquisition: Zhanlong MaInvestigation: Shun Chai and Zhanlong Ma Methodology: Shun Chai and Zhanlong MaProject administration: Zhanlong MaResources: Zhanlong MaSoftware: Shun ChaiSupervision: Zhanlong MaValidation: Yawen Yang and Chuanxian MaVisualization: Shun ChaiWriting – original draft: Shun Chai

Z. Feng, P. Rong, P. Cao, et al. (2018) Machine learning-based quantitative texture analysis of CT images of small renal masses: Differentiation of angiomyolipoma without visible fat from renal cell carcinoma. Eur Radiol. 10.1007/s00330-017-5118-z
S. Woo, C. H. Suh, J. Y. Cho, S. Y. Kim and S. H. Kim (2017) Diagnostic Performance of CT for Diagnosis of Fat-Poor Angiomyolipoma in Patients With Renal Masses: A Systematic Review and Meta-Analysis. AJR Am J Roentgenol. 10.2214/ajr.17.18184
S. Thiravit, W. Teerasamit and P. Thiravit (2018) The different faces of renal angiomyolipomas on radiologic imaging: a pictorial review. Br J Radiol. 10.1259/bjr.20170533
K. Sasaguri and N. Takahashi (2018) CT and MR imaging for solid renal mass characterization. Eur J Radiol. 10.1016/j.ejrad.2017.12.008
N. Hindman, L. Ngo, E. M. Genega, et al. (2012) Angiomyolipoma with minimal fat: can it be differentiated from clear cell renal cell carcinoma by using standard MR techniques? Radiology. 10.1148/radiol.12112087
A. D. de Leon, P. Kapur and I. Pedrosa (2019) Radiomics in Kidney Cancer: MR Imaging. Magn Reson Imaging Clin N Am. 10.1016/j.mric.2018.08.005
B. Bhindi, R. H. Thompson, C. M. Lohse, et al. (2018) The Probability of Aggressive Versus Indolent Histology Based on Renal Tumor Size: Implications for Surveillance and Treatment. Eur Urol. 10.1016/j.eururo.2018.06.003
C. S. Lim, N. Schieda and S. G. Silverman (2019) Update on Indications for Percutaneous Renal Mass Biopsy in the Era of Advanced CT and MRI. AJR Am J Roentgenol. 10.2214/ajr.19.21093
X. L. Li, L. X. Shi, Q. C. Du, W. Wang, L. W. Shao and Y. W. Wang (2020) Magnetic resonance imaging features of minimal-fat angiomyolipoma and causes of preoperative misdiagnosis. World J Clin Cases. 10.12998/wjcc.v8.i12.2502
B. K. Park (2019) Renal Angiomyolipoma Based on New Classification: How to Differentiate It From Renal Cell Carcinoma. AJR Am J Roentgenol. 10.2214/AJR.18.20408
S. de Silva, K. R. Lockhart, P. Aslan, et al. (2021) The diagnostic utility of diffusion weighted MRI imaging and ADC ratio to distinguish benign from malignant renal masses: sorting the kittens from the tigers. BMC Urol. 10.1186/s12894-021-00832-5
H. Li, A. Li, H. Zhu, et al. (2019) Whole-Tumor Quantitative Apparent Diffusion Coefficient Histogram and Texture Analysis to Differentiation of Minimal Fat Angiomyolipoma from Clear Cell Renal Cell Carcinoma. Acad Radiol. 10.1016/j.acra.2018.06.015
Y. Ding, M. Zeng, S. Rao, C. Chen, C. Fu and J. Zhou (2016) Comparison of Biexponential and Monoexponential Model of Diffusion-Weighted Imaging for Distinguishing between Common Renal Cell Carcinoma and Fat Poor Angiomyolipoma. Korean J Radiol. 10.3348/kjr.2016.17.6.853
K. Sasiwimonphan, N. Takahashi, B. C. Leibovich, R. E. Carter, T. D. Atwell and A. Kawashima (2016) Small (<4 cm) Renal Mass: Differentiation of Angiomyolipoma without Visible Fat from Renal Cell Carcinoma Utilizing MR Imaging. Radiology. 10.1148/radiol.2016164024
R. Ferré, F. Cornelis, V. Verkarre, et al. (2015) Double-echo gradient chemical shift MR imaging fails to differentiate minimal fat renal angiomyolipomas from other homogeneous solid renal tumors. Eur J Radiol. 10.1016/j.ejrad.2014.11.040
Y. Ma, F. Cao, X. Xu and W. Ma (2020) Can whole-tumor radiomics-based CT analysis better differentiate fat-poor angiomyolipoma from clear cell renal cell caricinoma: compared with conventional CT analysis? Abdom Radiol (NY). 10.1007/s00261-020-02414-9
L. Dell'Atti, N. Bianchi and G. Aguiari (2022) New Therapeutic Interventions for Kidney Carcinoma: Looking to the Future. Cancers (Basel). 10.3390/cancers14153616
M. P. Wilson, D. Patel, P. Katlariwala and G. Low (2021) A review of clinical and MR imaging features of renal lipid-poor angiomyolipomas. Abdom Radiol (NY). 10.1007/s00261-020-02835-6
H. Tanaka, Y. Fujii, H. Tanaka, et al. (2017) Stepwise algorithm using computed tomography and magnetic resonance imaging for diagnosis of fat-poor angiomyolipoma in small renal masses: Development and external validation. Int J Urol. 10.1111/iju.13354
W. Jomoto, H. Takaki, S. Yamamoto, et al. (2022) Differentiation of Angiomyolipoma With Minimal Fat from Clear Cell Renal Cell Carcinoma Using Non-contrast Multiparametric Magnetic Resonance Imaging. In Vivo. 10.21873/invivo.13016
T. M. Kim, H. Ahn, H. J. Lee, et al. (2022) Differentiating renal epithelioid angiomyolipoma from clear cell carcinoma: using a radiomics model combined with CT imaging characteristics. Abdom Radiol (NY). 10.1007/s00261-022-03571-9
Z. Sun, Y. Cui, C. Xu, et al. (2022) Preoperative Prediction of Inferior Vena Cava Wall Invasion of Tumor Thrombus in Renal Cell Carcinoma: Radiomics Models Based on Magnetic Resonance Imaging. Front Oncol. 10.3389/fonc.2022.863534
S. Matsumoto, Y. Arita, S. Yoshida, et al. (2022) Utility of radiomics features of diffusion-weighted magnetic resonance imaging for differentiation of fat-poor angiomyolipoma from clear cell renal cell carcinoma: model development and external validation. Abdom Radiol (NY). 10.1007/s00261-022-03486-5
Y. Arita, S. Yoshida, T. C. Kwee, et al. (2021) Diagnostic value of texture analysis of apparent diffusion coefficient maps for differentiating fat-poor angiomyolipoma from non-clear-cell renal cell carcinoma. Eur J Radiol. 10.1016/j.ejrad.2021.109895
P. Nie, G. Yang, Z. Wang, et al. (2020) A CT-based radiomics nomogram for differentiation of renal angiomyolipoma without visible fat from homogeneous clear cell renal cell carcinoma. Eur Radiol. 10.1007/s00330-019-06427-x
L. Jian, Y. Liu, Y. Xie, S. Jiang, M. Ye and H. Lin (2022) MRI-Based Radiomics and Urine Creatinine for the Differentiation of Renal Angiomyolipoma With Minimal Fat From Renal Cell Carcinoma: A Preliminary Study. Front Oncol. 10.3389/fonc.2022.876664
A. V. Malagi, A. Netaji, V. Kumar, et al. (2022) IVIM-DKI for differentiation between prostate cancer and benign prostatic hyperplasia: comparison of 1.5 T vs. 3 T MRI. Magma. 10.1007/s10334-021-00932-1
A. V. Malagi, S. Shivaji, D. Kandasamy, et al. (2023) Pancreatic Mass Characterization Using IVIM-DKI MRI and Machine Learning-Based Multi-Parametric Texture Analysis. Bioengineering (Basel). 10.3390/bioengineering10010083
Q. Xu, Q. Zhu, H. Liu, et al. (2022) Differentiating Benign from Malignant Renal Tumors Using T2- and Diffusion-Weighted Images: A Comparison of Deep Learning and Radiomics Models Versus Assessment from Radiologists. J Magn Reson Imaging. 10.1002/jmri.27900
I. L. Xi, Y. Zhao, R. Wang, et al. (2020) Deep Learning to Distinguish Benign from Malignant Renal Lesions Based on Routine MR Imaging. Clin Cancer Res. 10.1158/1078-0432.CCR-19-0374
J. Kalpathy-Cramer, A. Mamomov, B. Zhao, et al. (2016) Radiomics of Lung Nodules: A Multi-Institutional Study of Robustness and Agreement of Quantitative Imaging Features. Tomography. 10.18383/j.tom.2016.00235
J. E. van Timmeren, D. Cester, S. Tanadini-Lang, H. Alkadhi and B. Baessler (2020) Radiomics in medical imaging-"how-to" guide and critical reflection. Insights Imaging. 10.1186/s13244-020-00887-2

Table 1

Clinical factors of the training and validation sets
Clinical factors		Training set (n = 62)			Validation set (n = 26)
	fp-AML	ccRCC	p	fp-AML	ccRCC	p
Age/year (mean ± SD)	47.07 ± 12.22	54.43 ± 12.57	0.045	47.75 ± 10.57	61.78 ± 13.49	0.011
Gender (male/female)	5/10	39/8	0.001	3/5	12/6	0.165
Dmax/cm (mean ± SD)	2.37 ± 0.86	2.43 ± 0.83	0.811	2.19 ± 0.83	2.64 ± 1.11	0.308
Origination (left / right)	5/10	26/21	0.138	4/4	7/11	0.597
location (upper pole/ middle pole/ lower pole)	4/9/2	17/18/12	0.319	1/4/3	3/8/7	0.950
shape(regular/irregular)	8/7	22/25	0.660	5/3	7/11	0.265
margin(smooth/nonsmooth)	11/4	23/24	0.098	8/0	11/7	0.039
capsule(absent/present)	11/4	30/17	0.498	7/1	13/5	0.393
Extrarenal extension(absent/≤50% extension/>50% extension)	5/4/6	21/15/11	0.451	1/1/6	5/7/6	0.142
T2WI signal (homogeneously hypointense/homogeneously hyperintense / heterogeneous)	10/3/2	6/9/32	0.001	6/0/2	7/4/7	0.171
DWI signal (homogeneously hypointense/homogeneously hyperintense / heterogeneous)	10/2/3	22/6/19	0.331	4/2/2	9/2/7	0.606
ADC value	1.06 ± 0.42	1.56 ± 0.43	0.001	1.08 ± 0.40	1.53 ± 0.42	0.02

Table 2

Diagnostic performance of the clinical factor model, the radiomics signature, and the radiomics nomogram
Model	Training set (n = 62)				Validation set (n = 26)
	AUC(95% CI)	Sensitivit	Specificit	Accuracy	AUC(95% CI)	Sensitivit	Specificit	Accuracy
Clinical model	0.891(0.791 ~ 0.990)	0.851	0.800	0.838	0.774(0.554–0.995)	0.667	0.750	0.692
Radiomics signature	0.923(0.832-1.000)	1.000	0.800	0.952	0.778(0.571–0.985)	0.833	0.500	0.731
Radiomics nomogram	0.974(0.938-1.000)	0.978	0.867	0.952	0.833(0.675–0.992)	0.778	0.625	0.771

No competing interests reported.

A MRI-based radiomics nomogram for differentiation of fat-poor renal angiomyolipoma from nodular clear cell renal cell carcinoma

Status:

Version 1

Abstract

Objectives

Methods

Results

Conclusions

Figures

Introduction

Materials and methods

Patients

MRI examination

MRI feature evaluation

Construction of the clinical factor model

Segmentation of tumors

Radiomics feature extraction and selection

Construction and evaluation of the radiomics model

Statistical analysis

Results

Clinical variables and clinical model design

Construction and evaluation of radiomics models

Radiomics Nomogram Construction and Evaluation

Discussion

Declarations

Author Contribution

References

Tables

Additional Declarations

Status:

Version 1