A weakly supervised deep learning model based on CT images for predicting WHO/ISUP nuclear grade of clear cell renal cell carcinoma

A weakly supervised deep learning model based on CT images for predicting WHO/ISUP nuclear grade of clear cell renal cell carcinoma | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A weakly supervised deep learning model based on CT images for predicting WHO/ISUP nuclear grade of clear cell renal cell carcinoma Cuiping Li, Shuai Li, Chao Zhu, Baoxin Qian, Jianying Li, Xingwang Wu, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6489710/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Purpose: The study aims to evaluate the performance of weakly supervised deep learning models in distinguishing high-grade and low-grade clear cell renal cell carcinoma (ccRCC) in comparison to strongly supervised deep learning models. Method: Pathologically confirmed ccRCC from two hospitals between January 2017 and April 2022 were included. The strongly and weakly supervised deep learning models were on three-phase images (CMP, NP and UP) based on the 3D ResNet-18 network and 2D ResNet-18 network, respectively using three-phase images (CMP, NP and UP) to form six deep learning models. Accuracy, sensitivity, specificity, and area-under-the-curve (AUC) were used to assess the discriminatory efficacy of the deep learning models. Results: A total of 306 ccRCC patients were collected in this retrospective study. Among them, 165 were low-grade ccRCC, and 141 were high-grade ccRCC. Data were divided into a training set (n=214) and a testing set (n=92) according to the ratio of 7:3. Among the weakly supervised deep learning models based on three-phase images, the model based on CMP images has the highest diagnostic performance, and its accuracy, sensitivity, specificity, and AUC values are 0.859, 0.857, 0.860 and 0.907, respectively. These values had no significant difference from the strongly supervised deep learning model based on CMP images (the accuracy, sensitivity, specificity, and AUC values were 0.848, 0.854, 0.843 and 0.922, respectively). Conclusion: The weakly supervised deep learning model developed in this study using CMP images has the same high diagnostic performance as the strongly supervised deep learning model in distinguishing high-grade ccRCC from low-grade ccRCC. With richer samples and sufficient developing, the weakly supervised deep learning model may become a routine clinical tool to reduce the physical toll of biopsy on patients. clear cell renal cell carcinoma weakly supervised deep learning WHO/IUSP nuclear grade computer tomography Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Clear cell renal cell carcinoma (ccRCC) is the most common malignancy of the kidney, accounting for approximately 80% of renal cell carcinoma (RCC) [1]. Papillary renal cell carcinoma (pRCC) and chromophobe renal cell carcinoma (chRCC) are relatively rare renal malignancies with a better prognosis than ccRCC [2, 3]. According to the grading criteria established by the World Health Organization/International Society of Urological Pathology (WHO/ISUP) in 2016, RCCs were classified into four grades (I, II, III, IV). The higher the grade, the worse the patient’s prognosis [4]. Percutaneous puncture biopsy is the gold standard for preoperative diagnosis of kidney tumors. Still, it is an invasive test and may lead to complications such as bleeding and infection [5]. In addition, due to the heterogeneous nature of the tumor, the nuclear grade obtained by puncture biopsy is lower than the true one, which in turn delays the patient’s treatment [6, 7]. Cross-sectional imaging, such as CT, MRI and ultrasound, enables preoperative non-invasive diagnosis of renal tumors. Studies have shown that imaging can be used not only to determine the benignity or malignancy of renal tumors, but also preoperatively assess the histological grade of ccRCC [8-11]. Some traditional radiological characteristics have been shown to correlate closely with tumor gradings, such as tumor size and enhancement pattern [12-14]. However, conventional imaging characteristics provide limited information and are susceptible to the subjective judgment of radiologists. Thus, preoperative prediction of pathologic grading of ccRCC remains challenging. In the past few years, deep learn-based algorithms, such as convolutional neural networks (CNN), have been developed and evaluated for benign and malignant renal tumors discrimination, ccRCC nuclear grading [15, 16]. Several studies have utilized weakly supervised CNN for lesion segmentation [17-21]. Most studies have focused on deep learning based on single-phase CT images to distinguish nuclear grades of ccRCC and have mainly used strongly supervised deep learning models. To our knowledge, few studies have explored the use of weakly supervised deep learning models to distinguish high-grade ccRCC from low-grade ccRCC. In this study, we aimed to build weakly supervised deep learning-based model to classify high-grade ccRCC versus low-grade ccRCC. In addition, this paper explores the advantages and disadvantages between weakly supervised deep learning and strongly supervised deep learning, as well as the impact of CT images from different imaging phases on deep learning results. The established deep learning model may provide a convenient, non-invasive and accurate method for WHO/ISUP nuclear grading of ccRCC, which not only reduces patient suffering but also helps clinicians to make decisions. Materials and Methods This multi-institutional retrospective study was approved by the local ethics committee and the requirement for informed consent was waived (ID: PJ2022-06-48). Patients for the study were primarily sourced from two hospitals, and patients with kidney tumors who underwent abdominal CT scans and were pathologically confirmed at both hospitals from January 2017 to April 2022 were collected. The inclusion criteria were as follows: (1) Patients with pathologically confirmed ccRCC and clear WHO/ISUP nuclear grade; (2) Patients all had preoperative CT plain and triple-phase enhancement scans. The exclusion criteria were as follows: (1) Significant artifacts on CT images; (2) Patient had received previous tumor treatment. In this study, WHO/ISUP nuclear grading system was used as the reference standard. Grades I and II were considered low-grade ccRCC, while grades III and IV were considered high-grade ccRCC. The flow chart for study population recruitment is shown in Figure 1 . Finally, a total of 306 patients were included in our study. CT imaging acquisition All CT images obtained in this study were based on the three different CT scanners ( Table A1 ). The scanning parameters are as follows: tube voltage of 120 -140 kV, tube current of 220-300 mA, detector collimation of 64 × 0.625 mm, matrix of 512 × 512, rack rotation time of 0.5 s, and slicer thickness of 5 mm. A total of 60-100 mL of non-ionic contrast agent was injected into the anterior elbow vein at 2.5-3.0 mL/s using a high-pressure syringe. A total of four phases of CT images were obtained, starting with the unenhanced phase (UP, no contrast injection), followed by the corticomedullary phase (CMP, 30 s after contrast injection) and the nephrographic phase (NP, 90 s after contrast injection). Finally, the excretory phase (EP, 6-8 minutes after contrast injection). Strongly supervised deep learning Two radiologists with more than 10 years of experience in diagnostic abdominal imaging outlined the tumor lesion to obtain a three-dimensional region of interest (3D-ROI), mainly using ITK-SNAP software for layer-by-layer annotation on three-phase images (UP, CMP, NP). The main lesion was defined as the lesion with maximum tumor diameter when multiple lesions existed. The border of the tumor was outlined along the tumor in all three-phase images. Two radiologists who were blinded to the clinical data reviewed the images and resolved the discrepancies with consensus. Then, to eliminate the effect of pixel spacing, we resampled the tumor volume to a certain isotropic resolution of 1.0×1.0×1.0 mm 3 and the ROI was resized 32×32×32 pixels using cubic spline interpolation. The patients were randomly divided into the training and testing sets according to the ratio of 7:3. We initially obtained 214 training images and 92 testing images to train and validate the network in each sequence. For each sequence of a patient, the label remained the same. Due to the limited amount of data, we performed data augmentation on each training images such as random rotation to increase the data in the training set. After the augmentation, each training image was rotated 30°, 60°, 90° and 180°, we created 856 (214×4) images in each sequence. We used three-phase images (UP, CMP, NP) to construct strongly supervised deep learning (DL) model based on three dimensions-residential (3D ResNet-18) network [22]. Traditional deep learning algorithms simply stack images on top of each other to build a model. ResNet extracts the residual features and uses skipped connections to subtract the features learned from the input layer. The architecture of the strongly supervised DL network is shown in Figure 2 . Detailed description of the computing environment and hyper-parameter is provided in Supplementary Appendix 1 . The ROI (32×32×32 pixel) of each patient were fed into the ResNet-18 network as inputs. ResNet-18 consists of 17 convolutional layers and one fully connected layer. Through processing by a sequence of convolution and pooling layers, the WHO/ISUP nuclear grade of ccRCC could be obtained from the last full-connection layer of the DL network. The abovementioned DL model was performed on UP, CMP and NP phases, respectively, thus acquiring corresponding three single phase-based predict results. Weakly supervised deep learning Based on the annotated 3D-ROI information, the slice where the lesion was located was selected. Thus, each patient had multiple 2D slices. For each slice of the same patient, the label remained the same. All slices were standardized such that the intensity was distributed to have mean = 0 and standard deviation = 1 and the matrices of all slices were resized to 256×256 pixels. All images were divided into training and testing sets according to the groups mentioned above. To evaluate the effect of different sequences on the training efficacy, we evaluated three-phase images (UP, CMP, NP) as input sources for training. The three-phase images were individually imported into the 2D ResNet-18 network to build three models. The architecture of 2D ResNet-18 is a two dimensions version of the 3D ResNet-18, where 3D convolution is replaced by 2D convolution and pooling operation is replaced by 2D pooling, while the BatchNorm and activation function (ReLu) remain unchanged. Training details are the same as strongly supervised DL. The output results in one probability vector for a single layer of images. Since each image in single phase contains multiple slices, after predicting the ccRCC grade of a single slice using the 2D ResNet-18 classification network, the next step will be to evaluate the classification performance of the model from the perspective of one ccRCC cases. The architecture also involves a clustering technique [23]: patients will be classified as a given grade ccRCC patients under that grade if the CT slices of ccRCC lesions at the joint with Intersection Over Union (IOU) >0.3 of adjacent ccRCC lesion slices and if there are more than three consecutive slices with the same lesion grade. We statistically calculated the total probability vector for each patient and finally obtained the graded judgment for that patient. To further understand the deep learning model applied in our prediction task, we used a deep learning visualization technique called class activation map for generating a heat map of class activations on the input image. Then, the heat map was overlaid on the original CT image to show the focus of the CNN algorithm, the red and yellow pixel regions correlate more strongly with tumor grading. The architecture of the weakly supervised DL network is shown in Figure 3. Performance evaluation in the testing set After building the models, we examined their performance in predicting the tumors in the independent testing set. We used the area under the receiver operating characteristic curve (AUC) to show the diagnostic ability of models in grading ccRCC patients. Accuracy, sensitivity and specificity were calculated and used to evaluate the ability of two different deep learning methods. Statistical analysis Statistical analysis was performed using SPSS software (version 25.0, IBM) and R package (version 3.5.1). Continuous variables were described as mean ± SD or median (inter-quartile range, IQR) and compared using independent samples t-test or Mann-Whitney U-test . Categorical variables were summarized as frequencies and percentages using the chi-square test or Fisher’s exact test . The predictive performance of strongly supervised deep learning models (UP, CMP and NP) and weakly supervised deep learning models (UP, CMP and NP) was evaluated using AUC. The comparison of AUC with strongly supervised deep learning models in each phase was done using the Delong test. Comparison of sensitivity and specificity between the strongly and weakly supervised deep learning models in the same phase was performed using McNemar’s test. Two sides p < 0.05 was considered significantly different. Results Study population A total of 306 patients with pathologically confirmed ccRCC from two hospitals were included in this study, including 165 patients with low-grade ccRCC and 141 patients with high-grade ccRCC. The patients in this study were randomly divided into training (n = 214) and testing (n = 92) groups according to the ratio of 7:3. The baseline information of the patients in this study is shown in Table 1 and Table A2 . Strongly supervised deep learning analysis The AUC values of the 3D ResNet-18 deep learning models based on UP, CMP and NP images in the training set were 0.894, 0.944 and 0.920, respectively. In the testing set, the AUC values for the three models were 0.875, 0.922 and 0.900, respectively ( Table 2 ). The ROC curves in the training and testing sets are shown in Figure 4 . The confusion matrix results of the three strongly supervised deep learning models in the testing set are shown in Figure A1 . Weakly supervised deep learning analysis From the perspective of a single image of ccRCC. The AUC values of the 2D ResNet-18 deep learning models based on UP, CMP and NP image in the training set were 0.876, 0.912 and 0.893, respectively. In the testing set, the AUC values for the three models were 0.866, 0.907 and 0.884, respectively ( Table 3 ). The ROC curves in the training and testing set are shown in Figure 5 . The confusion matrix results of the three weakly supervised deep learning models in the testing set are shown in Figure A2 . From the perspective of a whole ccRCC case, the sensitivity and specificity values based on UP, CMP and NP in the testing set were 0.841/0.687,0.800/0.831 and 0.863/0.861, respectively. We visualize the heatmap of the tumor to illustrate how much attention the algorithm pays to the segmentation results of the analysis. Examples of the heatmap for the UP, CMP and NP models are shown in Figure 6 . Comparison of strongly and weakly supervised deep learning model In testing set, the results show that the among the weakly supervised deep learning models, the prediction performance based on CMP images is the best, followed by NP images and finally UP images (Sensitivity: 0.857 > 0.833 > 0786, Specificity: 0.860 > 0.840 > 0.800). The McNemar’s test showed that there was no statistical difference between the specificity and sensitivity of the strongly supervised deep learning model and the weakly supervised deep learning model based on CMP images (CMP_sensitivity: strongly vs. weakly = 0.854 vs. 0.857, McNemar test P > 0.05; CMP_specificity: strongly vs. weakly = 0.843 vs. 0.860, McNemar’s test P > 0.05). Discussion We developed a weakly supervised deep learning model based on three-phase CT images to distinguish high-grade ccRCC from low-grade ccRCC and compared the results with the strongly supervised deep learning models. The results showed that the CMP-based images had the highest discrimination performance in both strongly and weakly supervised deep learning models, with AUC values of 0.922 and 0.907, respectively. The AUC, sensitivity, and specificity of the strongly supervised deep learning models based on UP, CMP and NP images were better than those of the corresponding weakly supervised deep learning models, but the differences between them were not statistically significant, indicating that the discriminative efficacy of the weakly supervised deep learning models was not weaker than that of the strongly supervised deep learning. And the weakly supervised deep learning could save more time because it does not require labeling of tumors. The gold standard for differentiating the WHO/ISUP grade of ccRCC prior to developing a clinical treatment strategy is by kidney puncture histological biopsy. However, this is after all an invasive test that may not only lead to bleeding, infection, or damage to surrounding organs, but also lead to metastasis of malignant tumors. Due to the heterogeneous nature of the tumor, the biopsy results obtained by puncture are often lower than the true WHO/ISUP grade, delaying patient treatment [5-7]. Artificial intelligence based on CT or MRI images is increasingly used in the study of kidney tumors [15, 16, 24, 25]. It mainly includes benign and malignant identification, grading of tumors and predicting the prognosis of tumors. Gao et al.[26] extracted radiomics features form multiphase CT images to distinguish smll (< 4cm) ccRCC, achieving AUC of 0.940 and 0.902 in the training and testing sets, respectively. In contrast to their focus solely on tumor diameters below 4 cm, our study extends beyond this limitation, enhancing its applicability in clinical settings. Shu et al. [27] extracted radiomics features from CMP and NP images of 260 ccRCC patients and developed CMP model, NP model and combined model for differentiating Fuhrman grade. The results showed that the NP model had the highest diagnostic efficacy with an AUC value of 0.818 and the CMP model had an AUC value of 0.766, the difference was not statistically significant ( P =0.0844). This is not quite consistent with our findings, where our deep learning model based on CMP images had significantly better AUC values than the deep learning model based on NP images, and the difference was statistically significant. The reason for this discrepancy may be because radiomics and deep learning belong to two different research methods, and there are also significant differences between the Fuhrman grading system and the WHO/ISUP grading system. In previous studies, most of them were based on the Fuhrman grading system [27-29]. The Fuhrman grading system, proposed by Fuhrman in 1982, is a widely used pathological grading system for renal cell carcinoma [28]. This grading system is only based on the analysis results of 103 cases of renal cell carcinoma, of which only 85 cases were followed up, and does not consider the tissue type of renal cell carcinoma. In practice, the classification system has some problems, such as difficulty in interpretation and poor repeatability [29, 30]. Therefore, in the 2016 edition of the WHO New Classification of Renal Tumors, this system was replaced by a new grading standard called the WHO/ISUP grading system [31]. Because the latter grading system is superior to the former and has better interpretability [30-32]. Deep learning models have been widely used in previous studies to evaluate kidney tumors [15, 16, 33-37]. Toda et al. [34] developed a deep learning-based algorithm for fully automated detection of small (≤4cm) RCC in contrast-enhanced CT images and evaluated its performance. The AUC values of this deep learning algorithm were 0.930 and 0.933 in dataset A and dataset B, respectively. Han et al. [35] constructed a deep learning framework to distinguish the three major subtypes of RCC (ccRCC, pRCC and chRCC) based on three-phase CT images, and the network had an accuracy of 0.85, sensitivity of 0.64-0.98, specificity of 0.83-0.93 and AUC value of 0.90. Xi et al. [38] collected MRI images (T1-enhanced and T2WI) from a total of 1162 patients with renal tumors and combined clinical characteristics to build a deep learning model based on ResNet to identify benign and malignant renal tumors. The accuracy, sensitivity and specificity of the deep learning model were 0.70, 0.92 and 0.41, respectively. Lin et al. [16] used multiple methods (image cropping, setting the attention level, selecting model complexity, and applying transfer learning) to build deep learning models used to distinguish high-grade ccRCC from low-grade ccRCC with AUC values of 0.82±0.11 and 0.81±0.04 in the internal test set and external test set, respectively. Xu et al. [15] divided 706 ccRCC patients into training and validation cohorts and built a deep learning framework initialized by self-supervised pre-training to identify high-grade ccRCC with AUC values of 0.864-0.882. In our study, a total of six deep learning models were built based on UP, CMP and NP images, three of which were strongly supervised and three were weakly supervised. Previous studies have mainly built deep models based on strong supervision, and few studies have built deep learning models based on weak supervision to distinguish high-grade ccRCC from low-grade ccRCC. The use of weakly supervised deep learning model would help physicians to save a lot of time, and the models has good performance in distinguishing WHO/IUSP grade of ccRCC than previous studies. ResNet networks are increasingly being used in deep learning studies of medical images. Xu et al. [37] collected MRI images (T2WI and DWI) from a total 217 renal tumor patients and built a deep learning model based on the ResNet-18 network to assess the benignity and malignancy of tumors. And compared it with the radiomics model, which resulted in a significantly better deep learning model than the radiomics model with an AUC value of 0.925. Lin et al. [16] built several deep learning models based on ResNet-18, ResNet-34, and ResNet-50 networks to evaluate the WHO/ISUP grade of ccRCC. The results showed that the deep learning models built based on ResNet-18 network provided the best results in differentiation. The above study showed that the RseNet-18 network was widely used in deep learning research. The six deep learning models established in our study were all built based on the ResNet-18 network which is different from the neural networks used by some other studies. For example, Nikpanah et al. [36] built a model based on the AlexNet network to distinguish between ccRCC and oncocytoma. Zuo et al. [39] used six neural networks (MobileNetV2, Efficient Net, ShuffleNet, ResNet-34, ResNet-50 and ResNet-101) to build a deep learning model to distinguish between pRCC and chRCC, with the model built on the MobileNetV2 network having better diagnostic results. There are still some limitations in this study. First, although the samples in this study were obtained from two large hospitals, we did not have an advantageous sample size compared with previous deep learning studies and hope to prospectively collect more cases for the study in the future. Second, this study only used the ResNet-18 network to build a deep learning model and did not use neural networks such as VGGNet and AlexNet to build a deep learning model to evaluate the WHO/ISUP grade of ccRCC. Third, this study built a deep learning model to distinguish high-grade ccRCC from low-grade ccRCC but did not collect clinical features to build a clinical model to assess its grade. Clinical models and radiomics models can be added in future studies to assess the WHO/ISUP grade of ccRCC. Conclusion In our study, we developed a weakly supervised deep learning model based on the 2D ResNet-18 network to identify focal locations for ccRCC and further evaluate its WHO/ISUP grading. These findings confirms that our weakly supervised deep learning model performs either favorably or comparably in terms of grade assessment of ccRCC with biopsies whilst enjoying the non-invasive and labor-saving property, which provides the original basis for the use of weakly supervised deep learning models in big data in the future. However, a larger sample size and prospective studies are needed to validate that the model can be used as a routine clinical tool. Declarations Data Availability Statement Data will be available upon request. Because the data used in this study comes from real patient data, which is related to the privacy of patients, if necessary, it is necessary to apply to the hospital management system for use, and can be uploaded to the public database after the hospital agrees. Acknowledgements The authors thank all staff members involved in the acquisition of data. They are grateful to the technical assistance provided by Huiying Medical Technology (Beijing). Ethics Statement In accordance with the Declaration of Helsinki, the studies involving human participants were reviewed and approved by the first affiliated hospital of Anhui medical university’s Ethics Committee (ID: PJ2022-06-48). Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements. Author Contributions Cuiping Li, Xingwang Wu, Yankun Gao and Chao Zhu: conception and design. Cuiping Li, Yankun Gao and Shuai Li: collection and arrangement of data. Cuiping Li and Yankun Gao: data analysis and manuscript writing. Baoxin Qian: Data analysis, validation, and visualization. Jianying Li, Xingwang Wu and Yankun Gao: review & editing. All authors contributed to the article and approved the submitted version. Funding The authors received no funding for this work. Conflict of interest statement The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Reference Motzer R J, Jonasch E, Michaelson M D, et al. NCCN Guidelines Insights: Kidney Cancer, Version 2.2020. J Natl Compr Canc Netw 2019; 17:1278-1285. Capitanio U, Cloutier V, Zini L, et al. A critical assessment of the prognostic value of clear cell, papillary and chromophobe histological subtypes in renal cell carcinoma: a population-based study. BJU International 2009; 103:1496-1500. Keegan K A, Schupp C W, Chamie K, et al. Histopathology of surgically treated renal cell carcinoma: survival differences by subtype and stage. J Urol 2012; 188:391-397. Moch H, Cubilla A L, Humphrey P A, 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. Caoili E M and Davenport M S Role of percutaneous needle biopsy for renal masses. Semin Intervent Radiol 2014; 31:20-26. Volpe A, Mattar K, Finelli A, et al. Contemporary results of percutaneous biopsy of 100 small renal masses: a single center experience. J Urol 2008; 180:2333-2337. Harris C R, Whitson J M, and Meng M V Under-grading of <4 cm renal masses on renal biopsy. BJU Int 2012; 110:794-797. Shin T, Duddalwar V A, Ukimura O, et al. Does Computed Tomography Still Have Limitations to Distinguish Benign from Malignant Renal Tumors for Radiologists? Urol Int 2017; 99:229-236. Kim J H, Sun H Y, Hwang J, et al. Diagnostic accuracy of contrast-enhanced computed tomography and contrast-enhanced magnetic resonance imaging of small renal masses in real practice: sensitivity and specificity according to subjective radiologic interpretation. World J Surg Oncol 2016; 14:260. Chen S H, Wu Y P, Li X D, et al. R.E.N.A.L. Nephrometry Score: A Preoperative Risk Factor Predicting the Fuhrman Grade of Clear-Cell Renal Carcinoma. J Cancer 2017; 8:3725-3732. Oh S, Sung D J, Yang K S, et al. Correlation of CT imaging features and tumor size with Fuhrman grade of clear cell renal cell carcinoma. Acta Radiol 2017; 58:376-384. Zhu Y H, Wang X, Zhang J, et al. Low enhancement on multiphase contrast-enhanced CT images: an independent predictor of the presence of high tumor grade of clear cell renal cell carcinoma. AJR Am J Roentgenol 2014; 203:W295-300. Zhao J, Zhang P, Chen X, et al. Lesion Size and Iodine Quantification to Distinguish Low-Grade From High-Grade Clear Cell Renal Cell Carcinoma Using Dual-Energy Spectral Computed Tomography. J Comput Assist Tomogr 2016; 40:673-677. Chen C, Kang Q, Xu B, et al. Differentiation of low- and high-grade clear cell renal cell carcinoma: Tumor size versus CT perfusion parameters. Clin Imaging 2017; 46:14-19. Xu L, Yang C, Zhang F, et al. Deep Learning Using CT Images to Grade Clear Cell Renal Cell Carcinoma: Development and Validation of a Prediction Model. Cancers (Basel) 2022; 14: Lin F, Ma C, Xu J, et al. A CT-based deep learning model for predicting the nuclear grade of clear cell renal cell carcinoma. Eur J Radiol 2020; 129:109079. Yang G, Wang C, Yang J, et al. Weakly-supervised convolutional neural networks of renal tumor segmentation in abdominal CTA images. BMC Med Imaging 2020; 20:37. Wang, C. W., Y. C. Lee, C. C. Chang, et al., A Weakly Supervised Deep Learning Method for Guiding Ovarian Cancer Treatment and Identifying an Effective Biomarker. Cancers (Basel), 2022. 14(7). Liu, M. Z., C. Swintelski, S. Sun, et al., Weakly Supervised Deep Learning Approach to Breast MRI Assessment. Acad Radiol, 2022. 29 Suppl 1: p. S166-S172. Zhou, C., Y. Jin, Y. Chen, et al., Histopathology classification and localization of colorectal cancer using global labels by weakly supervised deep learning. Comput Med Imaging Graph, 2021. 88: p. 101861. Zhou, J., L. Y. Luo, Q. Dou, et al., Weakly supervised 3D deep learning for breast cancer classification and localization of the lesions in MR images. J Magn Reson Imaging, 2019. 50(4): p. 1144-1151. Ebrahimi A, Luo S, and Chiong R. Introducing transfer learning to 3D ResNet-18 for Alzheimer’s disease detection on MRI images . in 2020 35th international conference on image and vision computing New Zealand (IVCNZ) . 2020. IEEE. Ma L, Wang Y, Guo L, et al. Developing and verifying automatic detection of active pulmonary tuberculosis from multi-slice spiral CT images based on deep learning. J Xray Sci Technol 2020; 28:939-951. Zabihollahy F, Schieda N, Krishna S, et al. Automated classification of solid renal masses on contrast-enhanced computed tomography images using convolutional neural network with decision fusion. Eur Radiol 2020; 30:5183-5190. Pedersen M, Andersen M B, Christiansen H, et al. Classification of renal tumour using convolutional neural networks to detect oncocytoma. Eur J Radiol 2020; 133:109343. Gao Y, Wang X, Zhao X, et al. 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. Shu J, Tang Y, Cui J, et al. Clear cell renal cell carcinoma: CT-based radiomics features for the prediction of Fuhrman grade. Eur J Radiol 2018; 109:8-12. Ding J, Xing Z, Jiang Z, et al. CT-based radiomic model predicts high grade of clear cell renal cell carcinoma. Eur J Radiol 2018; 103:51-56. Hussain M A, Hamarneh G, and Garbi R Learnable image histograms-based deep radiomics for renal cell carcinoma grading and staging. Comput Med Imaging Graph 2021; 90:101924. Shu J, Wen D, Xi Y, et al. Clear cell renal cell carcinoma: Machine learning-based computed tomography radiomics analysis for the prediction of WHO/ISUP grade. Eur J Radiol 2019; 121:108738. Wang R, Hu Z, Shen X, et al. Computed Tomography-Based Radiomics Model for Predicting the WHO/ISUP Grade of Clear Cell Renal Cell Carcinoma Preoperatively: A Multicenter Study. Front Oncol 2021; 11:543854. Zhao Y, Chang M, Wang R, et al. Deep Learning Based on MRI for Differentiation of Low- and High-Grade in Low-Stage Renal Cell Carcinoma. J Magn Reson Imaging 2020; 52:1542-1549. Wen-Zhi G, Tai T, Zhixin F, et al. Prediction of pathological staging and grading of renal clear cell carcinoma based on deep learning algorithms. J Int Med Res 2022; 50:3000605221135163. Toda N, Hashimoto M, Arita Y, et al. Deep Learning Algorithm for Fully Automated Detection of Small (</=4 cm) Renal Cell Carcinoma in Contrast-Enhanced Computed Tomography Using a Multicenter Database. Invest Radiol 2022; 57:327-333. Han S, Hwang S I, and Lee H J The Classification of Renal Cancer in 3-Phase CT Images Using a Deep Learning Method. Journal of Digital Imaging 2019; 32:638-643. Nikpanah M, Xu Z, Jin D, et al. A deep-learning based artificial intelligence (AI) approach for differentiation of clear cell renal cell carcinoma from oncocytoma on multi-phasic MRI. Clin Imaging 2021; 77:291-298. Xu Q, Zhu Q, Liu H, et al. 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 2022; 55:1251-1259. Xi I L, Zhao Y, Wang R, et al. Deep Learning to Distinguish Benign from Malignant Renal Lesions Based on Routine MR Imaging. Clin Cancer Res 2020; 26:1944-1952. Zuo T, Zheng Y, He L, et al. Automated Classification of Papillary Renal Cell Carcinoma and Chromophobe Renal Cell Carcinoma Based on a Small Computed Tomography Imaging Dataset Using Deep Learning. Front Oncol 2021; 11:746750. Tables Table 1. Patient demographics and characteristics of ccRCC in training and testing set. Variable Training set (n=214) Testing set (n=92) t/Z value P -value Age, year 58.2±11.6 55.5±12.6 1.833 0.068 Sex, % 1.485 0.223 Male 65% (139/214) 57.6% (53/92) Female 35% (75/214) 42.4% (39/92) Tumor diameter, cm 5.2±2.5 5.3±2.4 -0.313 0.754 ≤4cm 42.1% (90/214) 35.9% (33/92) 1.024 0.311 ＞4cm 57.9% (124/214) 64.1% (59/92) WHO/ISUP grade 0.121 0.728 High-grade 46.7% (100/214) 44.6% (41/92) Low-grade 53.3% (114/214) 55.4% (51/92) Side, % 1.494 0.222 Right 46.7% (100/214) 45.7% (42/92) Left 53.3% (114/214) 54.3% (50/92) Note: ccRCC, clear cell renal cell carcinoma; WHO/ISUP, the World Health Organization/International Society of Urological Pathology. Table 2. Diagnostic performance measures of the 3D ResNet-18 deep learning models in the training and testing sets. Models AUC (95% CI) training / testing Accuracy training / testing Sensitivity training /testing Specificity training /testing Delong Test P value_test CMP 0.944 (0.917-0.971) / 0.922 (0.867-0.977) 0.846 / 0.848 0.800 / 0.854 0.886 / 0.843 1.00 NP 0.920 (0.886-0.954) / 0.900 (0.840-0.960) 0.832 / 0.826 0.823 / 0.805 0.839 / 0.843 0.23 UP 0.894 (0.854-0.934) / 0.875 (0.805-0.945) 0.799 / 0.793 0.810 / 0.822 0.789 / 0.766 <0.05 Note: AUC, area under the curve; CI, confidence interval; CMP, corticomedullary phase; NP, nephrographic phase; UP, unenhanced phase. Table 3. Diagnostic performance measures of the 2D ResNet-18 deep learning models in the training and testing sets. Model Split Single slice >3 slices AUC (95%CI) Accuracy Sensitivity Specificity Accuracy Sensitivity Specificity CMP training 0.912(0.897-0.926) 0.926 0.927 0.924 0.864 0.865 0.864 testing 0.907(0.888-0.927) 0.862 0.863 0.861 0.859 0.857 0.860 NP training 0.893(0.877-0.909) 0.914 0.911 0.918 0.846 0.835 0.855 testing 0.884(0.863-0.906) 0.813 0.800 0.831 0.837 0.833 0.840 UP training 0.876(0.859-0.893) 0.897 0.912 0.875 0.822 0.804 0.838 testing 0.866(0.843-0.890) 0.772 0.841 0.687 0.793 0.786 0.800 Note: AUC, area under the curve; CI, confidence interval; CMP, corticomedullary phase; NP, nephrographic phase; UP, unenhanced phase. Additional Declarations No competing interests reported. Supplementary Files Supplementary.docx Supplementary material Supplementary Appendix 1: Detailed description of the computing environment and hyper-parameter is provided in Supplementary Appendix 1. Figure A1: a, b and c are confusion matrices of the 3D ResNet-18 deep learning models based on CMP, NP and UP images in the testing set. A total of 92 patients are included, with 41 patients of high-grade ccRCC (labeled 1) and the remaining 51 patients of low-grade ccRCC (labeled 0). Figure A2: a, b, and c are confusion matrices of the 2D ResNet-18 deep learning models based on CMP, NP and UP images in the testing set. A total of 92 patients are included, with 41 patients of high-grade ccRCC (labeled 1) and the remaining 51 patients of low-grade ccRCC (labeled 0). Table A1: The Details on three CT scanners. Table A2:Patient demographics and characteristics of ccRCC in training and testing set. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 18 Jun, 2025 Reviewers invited by journal 30 May, 2025 Editor invited by journal 08 May, 2025 Editor assigned by journal 29 Apr, 2025 Submission checks completed at journal 29 Apr, 2025 First submitted to journal 20 Apr, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6489710","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":465070733,"identity":"a66685a4-e14f-4378-bd43-8b1dc4459bd0","order_by":0,"name":"Cuiping Li","email":"","orcid":"","institution":"First Affiliated Hospital of Anhui Medical University","correspondingAuthor":false,"prefix":"","firstName":"Cuiping","middleName":"","lastName":"Li","suffix":""},{"id":465070734,"identity":"aac0beea-4d50-44a7-9874-a5ee8307bc13","order_by":1,"name":"Shuai Li","email":"","orcid":"","institution":"First Affiliated Hospital of Anhui Medical University","correspondingAuthor":false,"prefix":"","firstName":"Shuai","middleName":"","lastName":"Li","suffix":""},{"id":465070735,"identity":"d4ee0803-d60c-41ac-9b42-1bd1eb063516","order_by":2,"name":"Chao Zhu","email":"","orcid":"","institution":"First Affiliated Hospital of Anhui Medical University","correspondingAuthor":false,"prefix":"","firstName":"Chao","middleName":"","lastName":"Zhu","suffix":""},{"id":465070736,"identity":"c34eb24d-fad0-4edc-af5d-352732c61778","order_by":3,"name":"Baoxin Qian","email":"","orcid":"","institution":"Huiying Medical Technology Co., Ltd, Room A206, B2, Dongsheng Science and Technology Park","correspondingAuthor":false,"prefix":"","firstName":"Baoxin","middleName":"","lastName":"Qian","suffix":""},{"id":465070737,"identity":"8675e127-5d53-4642-be1e-50cd227cd8d5","order_by":4,"name":"Jianying Li","email":"","orcid":"","institution":"CT Research Center, GE Healthcare China","correspondingAuthor":false,"prefix":"","firstName":"Jianying","middleName":"","lastName":"Li","suffix":""},{"id":465070738,"identity":"cedda11f-1931-4c6b-ada3-96e11228e710","order_by":5,"name":"Xingwang Wu","email":"","orcid":"","institution":"First Affiliated Hospital of Anhui Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xingwang","middleName":"","lastName":"Wu","suffix":""},{"id":465070739,"identity":"c2998991-d2b8-4bb1-87fe-2309b76dc570","order_by":6,"name":"Yankun Gao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzklEQVRIiWNgGAWjYBACfvbGxsc/Kv7ZsbE3EKlFsudwszHDmQPJ/DwHiNRicCO9TZix7QDjzBkJxGo5kNjGXMB2h9ng5uONNxhqbKIJO+zAwbbHM3ie8RncTiu2YDiWlttASAvfwcZ2Ax4JZmaD2zlmEowNhwlrYTjM2CbBY8DMuOHmGSK1CBxjbJPmSTgM9D4PkVokexibDWccSAMGMtAvCcT4hV/++cMHH//ZAKPy8MYbH2psiPALEjCQSCBFOUQLqTpGwSgYBaNgZAAASe1Ew543jckAAAAASUVORK5CYII=","orcid":"","institution":"First Affiliated Hospital of Anhui Medical University","correspondingAuthor":true,"prefix":"","firstName":"Yankun","middleName":"","lastName":"Gao","suffix":""}],"badges":[],"createdAt":"2025-04-20 14:08:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6489710/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6489710/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":83893962,"identity":"e1a53d8d-9a68-4eea-93f4-77490ac34eec","added_by":"auto","created_at":"2025-06-04 08:34:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":111234,"visible":true,"origin":"","legend":"\u003cp\u003ePatient selection flowchart. ccRcc, clear cell renal cell carcinoma; CT, computerized tomography.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6489710/v1/216be78318fb266ba3f49b6c.png"},{"id":83893966,"identity":"26a7ad07-8956-4775-b093-c738211642b8","added_by":"auto","created_at":"2025-06-04 08:34:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":138974,"visible":true,"origin":"","legend":"\u003cp\u003eThe 3D ResNet-18 deep learning analysis process. The illustration of the convolutional module in the 3D ResNet’s Building Block. ccRcc, clear cell renal cell carcinoma; ROI, region of interest; UP, unenhanced phase; CMP, corticomedullary phase; NP, nephrographic phase.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6489710/v1/6bd12f3fc923df8177aa1bed.png"},{"id":83893964,"identity":"e04a9072-7520-4f89-8bc8-acf3336e1728","added_by":"auto","created_at":"2025-06-04 08:34:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":131312,"visible":true,"origin":"","legend":"\u003cp\u003eThe architecture of the weakly supervised deep learning network. ccRcc, clear cell renal cell carcinoma; CT, computerized tomography; UP, unenhanced phase; CMP, corticomedullary phase; NP, nephrographic phase; IOU, intersection over union.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6489710/v1/4981f36206e4a911aebdcbe2.png"},{"id":83893963,"identity":"68d062d2-6caa-467f-8da7-7bd0b922b056","added_by":"auto","created_at":"2025-06-04 08:34:49","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":69047,"visible":true,"origin":"","legend":"\u003cp\u003eThe receiver operating characteristic (ROC) curves of the 3D ResNet-18 deep learning models based on CMP (model_1), NP (model_2) and UP (model_3) in the training (\u003cstrong\u003ea\u003c/strong\u003e) and testing (\u003cstrong\u003eb\u003c/strong\u003e) sets. UP, unenhanced phase; CMP, corticomedullary phase; NP, nephrographic phase; AUC, area under the curve.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6489710/v1/2fd144f91afc6c3dac8d5979.png"},{"id":83895419,"identity":"6e2ff867-a6fd-4a2d-9aa7-5fb0b12c5743","added_by":"auto","created_at":"2025-06-04 08:42:49","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":72983,"visible":true,"origin":"","legend":"\u003cp\u003eThe receiver operating characteristic (ROC) curves of the 2D ResNet-18 deep learning models based on CMP (model_1), NP (model_2) and UP (model_3) in the training (\u003cstrong\u003ea\u003c/strong\u003e) and testing (\u003cstrong\u003eb\u003c/strong\u003e) sets. UP, unenhanced phase; CMP, corticomedullary phase; NP, nephrographic phase; AUC, area under the curve.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6489710/v1/354fd258d2ae0dbf25994cd5.png"},{"id":83893969,"identity":"3baf0bd4-73ee-435c-a89f-a660e21398b6","added_by":"auto","created_at":"2025-06-04 08:34:50","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":280532,"visible":true,"origin":"","legend":"\u003cp\u003eFeature heatmaps of a representative patient generated from the 2D ResNet-18 models. Red and yellow pixel areas correlated more strongly with WHO/ISUP grade. \u003cstrong\u003ea/b/c.\u003c/strong\u003e From left to right are the original image of UP/CMP/NP, the corresponding feature heatmap and the specific location of the tumor. WHO/ISUP, the World Health Organization/International Society of Urological Pathology; UP, unenhanced phase; CMP, corticomedullary phase; NP, nephrographic phase.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6489710/v1/8f95d73da1e66e12b7c041a3.png"},{"id":83895790,"identity":"9e88211c-24bb-44a9-a594-d72b3300a9d7","added_by":"auto","created_at":"2025-06-04 08:50:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1401334,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6489710/v1/757adca0-0e3b-47ae-9742-c5289f248c37.pdf"},{"id":83893968,"identity":"7ce6cffe-c833-4691-a312-ca7a55bda812","added_by":"auto","created_at":"2025-06-04 08:34:49","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":204728,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary material\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupplementary Appendix 1:\u003c/strong\u003e Detailed description of the computing environment and hyper-parameter is provided in Supplementary Appendix 1.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure A1:\u003c/strong\u003e \u003cstrong\u003ea\u003c/strong\u003e, \u003cstrong\u003eb\u003c/strong\u003e and \u003cstrong\u003ec\u003c/strong\u003e are confusion matrices of the 3D ResNet-18 deep learning models based on CMP, NP and UP images in the testing set. A total of 92 patients are included, with 41 patients of high-grade ccRCC (labeled 1) and the remaining 51 patients of low-grade ccRCC (labeled 0).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure A2:\u003c/strong\u003e \u003cstrong\u003ea\u003c/strong\u003e, \u003cstrong\u003eb\u003c/strong\u003e, and \u003cstrong\u003ec\u003c/strong\u003e are confusion matrices of the 2D ResNet-18 deep learning models based on CMP, NP and UP images in the testing set. A total of 92 patients are included, with 41 patients of high-grade ccRCC (labeled 1) and the remaining 51 patients of low-grade ccRCC (labeled 0).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable A1: \u003c/strong\u003eThe Details on three CT scanners.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable A2:\u003c/strong\u003ePatient demographics and characteristics of ccRCC in training and testing set.\u003c/p\u003e","description":"","filename":"Supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-6489710/v1/3a5f418ae73a550af47b1bde.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A weakly supervised deep learning model based on CT images for predicting WHO/ISUP nuclear grade of clear cell renal cell carcinoma","fulltext":[{"header":"Introduction ","content":"\u003cp\u003eClear cell renal cell carcinoma (ccRCC) is the most common malignancy of the kidney, accounting for approximately 80% of renal cell carcinoma (RCC) [1]. Papillary renal cell carcinoma (pRCC) and chromophobe renal cell carcinoma (chRCC) are relatively rare renal malignancies with a better prognosis than ccRCC [2, 3]. According to the grading criteria established by the World Health Organization/International Society of Urological Pathology (WHO/ISUP) in 2016, RCCs were classified into four grades (I, II, III, IV). The higher the grade, the worse the patient’s prognosis [4]. Percutaneous puncture biopsy is the gold standard for preoperative diagnosis of kidney tumors. Still, it is an invasive test and may lead to complications such as bleeding and infection [5]. In addition, due to the heterogeneous nature of the tumor, the nuclear grade obtained by puncture biopsy is lower than the true one, which in turn delays the patient’s treatment [6, 7].\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp;Cross-sectional imaging, such as CT, MRI and ultrasound, enables preoperative non-invasive diagnosis of renal tumors. Studies have shown that imaging can be used not only to determine the benignity or malignancy of renal tumors, but also preoperatively assess the histological grade of ccRCC [8-11]. Some traditional radiological characteristics have been shown to correlate closely with tumor gradings, such as tumor size and enhancement pattern [12-14]. However, conventional imaging characteristics provide limited information and are susceptible to the subjective judgment of radiologists. Thus, preoperative prediction of pathologic grading of ccRCC remains challenging.\u003c/p\u003e\n\u003cp\u003eIn the past few years, deep learn-based algorithms, such as convolutional neural networks (CNN), have been developed and evaluated for benign and malignant renal tumors discrimination, ccRCC nuclear grading [15, 16]. Several studies have utilized weakly supervised CNN for lesion segmentation [17-21]. Most studies have focused on deep learning based on single-phase CT images to distinguish nuclear grades of ccRCC and have mainly used strongly supervised deep learning models. To our knowledge, few studies have explored the use of weakly supervised deep learning models to distinguish high-grade ccRCC from low-grade ccRCC.\u003c/p\u003e\n\u003cp\u003eIn this study, we aimed to build weakly supervised deep learning-based model to classify high-grade ccRCC versus low-grade ccRCC. In addition, this paper explores the advantages and disadvantages between weakly supervised deep learning and strongly supervised deep learning, as well as the impact of CT images from different imaging phases on deep learning results. The established deep learning model may provide a convenient, non-invasive and accurate method for WHO/ISUP nuclear grading of ccRCC, which not only reduces patient suffering but also helps clinicians to make decisions.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eThis multi-institutional retrospective study was approved by the local ethics committee and the requirement for informed consent was waived (ID: PJ2022-06-48). Patients for the study were primarily sourced from two hospitals, and patients with kidney tumors who underwent abdominal CT scans and were pathologically confirmed at both hospitals from January 2017 to April 2022 were collected. The inclusion criteria were as follows: (1) Patients with pathologically confirmed ccRCC and clear WHO/ISUP nuclear grade; (2) Patients all had preoperative CT plain and triple-phase enhancement scans. The exclusion criteria were as follows: (1) Significant artifacts on CT images; (2) Patient had received previous tumor treatment. In this study, WHO/ISUP nuclear grading system was used as the reference standard. Grades I and II were considered low-grade ccRCC, while grades III and IV were considered high-grade ccRCC. The flow chart for study population recruitment is shown in \u003cstrong\u003eFigure 1\u003c/strong\u003e. Finally, a total of 306 patients were included in our study.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eCT imaging acquisition\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eAll CT images obtained in this study were based on the three different CT scanners (\u003cstrong\u003eTable A1\u003c/strong\u003e). The scanning parameters are as follows: tube voltage of 120 -140 kV, tube current of 220-300 mA, detector collimation of 64\u0026nbsp;×\u0026nbsp;0.625 mm, matrix of 512\u0026nbsp;×\u0026nbsp;512, rack rotation time of 0.5 s, and slicer thickness of 5 mm. A total of 60-100 mL of non-ionic contrast agent was injected into the anterior elbow vein at 2.5-3.0 mL/s using a high-pressure syringe. A total of four phases of CT images were obtained, starting with the unenhanced phase (UP, no contrast injection), followed by the corticomedullary phase (CMP, 30 s after contrast injection) and the nephrographic phase (NP, 90 s after contrast injection). Finally, the excretory phase (EP, 6-8 minutes after contrast injection).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eStrongly supervised deep learning\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eTwo radiologists with more than 10 years of experience in diagnostic abdominal imaging outlined the tumor lesion to obtain a three-dimensional region of interest (3D-ROI), mainly using ITK-SNAP software for layer-by-layer annotation on three-phase images (UP, CMP, NP). The main lesion was defined as the lesion with maximum tumor diameter when multiple lesions existed. The border of the tumor was outlined along the tumor in all three-phase images. Two radiologists who were blinded to the clinical data reviewed the images and resolved the discrepancies with consensus. Then, to eliminate the effect of pixel spacing, we resampled the tumor volume to a certain isotropic resolution of 1.0×1.0×1.0 mm\u003csup\u003e3\u0026nbsp;\u003c/sup\u003eand the ROI was resized 32×32×32 pixels using cubic spline interpolation.\u003c/p\u003e\n\u003cp\u003eThe patients were randomly divided into the training and testing sets according to the ratio of 7:3. We initially obtained 214 training images and 92 testing images to train and validate the network in each sequence. For each sequence of a patient, the label remained the same. Due to the limited amount of data, we performed data augmentation on each training images such as random rotation to increase the data in the training set. After the augmentation, each training image was rotated 30°, 60°, 90°\u0026nbsp;and 180°, we created 856 (214×4) images in each sequence.\u003c/p\u003e\n\u003cp\u003eWe used three-phase images (UP, CMP, NP) to construct strongly supervised deep learning (DL) model based on three dimensions-residential (3D ResNet-18) network [22]. Traditional deep learning algorithms simply stack images on top of each other to build a model. ResNet extracts the residual features and uses skipped connections to subtract the features learned from the input layer. The architecture of the strongly supervised DL network is shown in \u003cstrong\u003eFigure 2\u003c/strong\u003e. Detailed description of the computing environment and hyper-parameter is provided in \u003cstrong\u003eSupplementary Appendix 1\u003c/strong\u003e. The ROI (32×32×32 pixel) of each patient were fed into the ResNet-18 network as inputs. ResNet-18 consists of 17 convolutional layers and one fully connected layer. Through processing by a sequence of convolution and pooling layers, the WHO/ISUP nuclear grade of ccRCC could be obtained from the last full-connection layer of the DL network. The abovementioned DL model was performed on UP, CMP and NP phases, respectively, thus acquiring corresponding three single phase-based predict results.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eWeakly supervised deep learning\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eBased on the annotated 3D-ROI information, the slice where the lesion was located was selected. Thus, each patient had multiple 2D slices. For each slice of the same patient, the label remained the same. All slices were standardized such that the intensity was distributed to have mean = 0 and standard deviation = 1 and the matrices of all slices were resized to 256×256 pixels. All images were divided into training and testing sets according to the groups mentioned above. To evaluate the effect of different sequences on the training efficacy, we evaluated three-phase images (UP, CMP, NP) as input sources for training. The three-phase images were individually imported into the 2D ResNet-18 network to build three models. The architecture of 2D ResNet-18 is a two dimensions version of the 3D ResNet-18, where 3D convolution is replaced by 2D convolution and pooling operation is replaced by 2D pooling, while the BatchNorm and activation function (ReLu) remain unchanged. Training details are the same as strongly supervised DL. The output results in one probability vector for a single layer of images. Since each image in single phase contains multiple slices, after predicting the ccRCC grade of a single slice using the 2D ResNet-18 classification network, the next step will be to evaluate the classification performance of the model from the perspective of one ccRCC cases. The architecture also involves a clustering technique [23]: patients will be classified as a given grade ccRCC patients under that grade if the CT slices of ccRCC lesions at the joint with Intersection Over Union (IOU) \u0026gt;0.3 of adjacent ccRCC lesion slices and if there are more than three consecutive slices with the same lesion grade. We statistically calculated the total probability vector for each patient and finally obtained the graded judgment for that patient. To further understand the deep learning model applied in our prediction task, we used a deep learning visualization technique called class activation map for generating a heat map of class activations on the input image. Then, the heat map was overlaid on the original CT image to show the focus of the CNN algorithm, the red and yellow pixel regions correlate more strongly with tumor grading. The architecture of the weakly supervised DL network is shown in \u003cstrong\u003eFigure 3.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003ePerformance evaluation in the testing set\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eAfter building the models, we examined their performance in predicting the tumors in the independent testing set. We used the area under the receiver operating characteristic curve (AUC) to show the diagnostic ability of models in grading ccRCC patients. Accuracy, sensitivity and specificity were calculated and used to evaluate the ability of two different deep learning methods.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eStatistical analysis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eStatistical analysis was performed using SPSS software (version 25.0, IBM) and R package (version 3.5.1). Continuous variables were described as mean\u0026nbsp;±\u0026nbsp;SD or median (inter-quartile range, IQR) and compared using \u003cem\u003eindependent samples\u003c/em\u003e \u003cem\u003et-test\u003c/em\u003e or \u003cem\u003eMann-Whitney U-test\u003c/em\u003e. Categorical variables were summarized as frequencies and percentages using the \u003cem\u003echi-square test\u003c/em\u003e or \u003cem\u003eFisher’s exact test\u003c/em\u003e. The predictive performance of strongly supervised deep learning models (UP, CMP and NP) and weakly supervised deep learning models (UP, CMP and NP) was evaluated using AUC. The comparison of AUC with strongly supervised deep learning models in each phase was done using the Delong test. Comparison of sensitivity and specificity between the strongly and weakly supervised deep learning models in the same phase was performed using McNemar’s test. Two sides \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05 was considered significantly different.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cem\u003eStudy population\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eA total of 306 patients with pathologically confirmed ccRCC from two hospitals were included in this study, including 165 patients with low-grade ccRCC and 141 patients with high-grade ccRCC. The patients in this study were randomly divided into training (n = 214) and testing (n = 92) groups according to the ratio of 7:3. The baseline information of the patients in this study is shown in \u003cstrong\u003eTable 1\u0026nbsp;\u003c/strong\u003eand\u003cstrong\u003e\u0026nbsp;Table A2\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eStrongly supervised deep learning analysis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe AUC values of the 3D ResNet-18 deep learning models based on UP, CMP and NP images in the training set were 0.894, 0.944 and 0.920, respectively. In the testing set, the AUC values for the three models were 0.875, 0.922 and 0.900, respectively (\u003cstrong\u003eTable 2\u003c/strong\u003e). The ROC curves in the training and testing sets are shown in \u003cstrong\u003eFigure 4\u003c/strong\u003e. The confusion matrix results of the three strongly supervised deep learning models in the testing set are shown in \u003cstrong\u003eFigure A1\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eWeakly supervised deep learning analysis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eFrom the perspective of a single image of ccRCC. The AUC values of the 2D ResNet-18 deep learning models based on UP, CMP and NP image in the training set were 0.876, 0.912 and 0.893, respectively. In the testing set, the AUC values for the three models were 0.866, 0.907 and 0.884, respectively (\u003cstrong\u003eTable 3\u003c/strong\u003e). The ROC curves in the training and testing set are shown in \u003cstrong\u003eFigure 5\u003c/strong\u003e. The confusion matrix results of the three weakly supervised deep learning models in the testing set are shown in \u003cstrong\u003eFigure A2\u003c/strong\u003e. From the perspective of a whole ccRCC case, the sensitivity and specificity values based on UP, CMP and NP in the testing set were 0.841/0.687,0.800/0.831 and 0.863/0.861, respectively. We visualize the heatmap of the tumor to illustrate how much attention the algorithm pays to the segmentation results of the analysis. Examples of the heatmap for the UP, CMP and NP models are shown in \u003cstrong\u003eFigure 6\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eComparison of strongly and weakly supervised deep learning model\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eIn testing set, the results show that the among the weakly supervised deep learning models, the prediction performance based on CMP images is the best, followed by NP images and finally UP images (Sensitivity: 0.857 \u0026gt; 0.833 \u0026gt; 0786, Specificity: 0.860 \u0026gt; 0.840 \u0026gt; 0.800).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe McNemar’s test showed that there was no statistical difference between the specificity and sensitivity of the strongly supervised deep learning model and the weakly supervised deep learning model based on CMP images (CMP_sensitivity: strongly vs. weakly = 0.854 vs. 0.857, McNemar test\u003cem\u003e\u0026nbsp;P\u0026nbsp;\u003c/em\u003e\u0026gt; 0.05; CMP_specificity: strongly vs. weakly = 0.843 vs. 0.860, McNemar’s test \u003cem\u003eP\u003c/em\u003e \u0026gt; 0.05).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003e\u0026nbsp; \u0026nbsp; We developed a weakly supervised deep learning model based on three-phase CT images to distinguish high-grade ccRCC from low-grade ccRCC and compared the results with the strongly supervised deep learning models. The results showed that the CMP-based images had the highest discrimination performance in both strongly and weakly supervised deep learning models, with AUC values of 0.922 and 0.907, respectively. The AUC, sensitivity, and specificity of the strongly supervised deep learning models based on UP, CMP and NP images were better than those of the corresponding weakly supervised deep learning models, but the differences between them were not statistically significant, indicating that the discriminative efficacy of the weakly supervised deep learning models was not weaker than that of the strongly supervised deep learning. And the weakly supervised deep learning could save more time because it does not require labeling of tumors.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; The gold standard for differentiating the WHO/ISUP grade of ccRCC prior to developing a clinical treatment strategy is by kidney puncture histological biopsy. However, this is after all an invasive test that may not only lead to bleeding, infection, or damage to surrounding organs, but also lead to metastasis of malignant tumors. Due to the heterogeneous nature of the tumor, the biopsy results obtained by puncture are often lower than the true WHO/ISUP grade, delaying patient treatment [5-7]. Artificial intelligence based on CT or MRI images is increasingly used in the study of kidney tumors [15, 16, 24, 25]. It mainly includes benign and malignant identification, grading of tumors and predicting the prognosis of tumors. Gao et al.[26] extracted radiomics features form multiphase CT images to distinguish smll (\u0026lt; 4cm) ccRCC, achieving AUC of 0.940 and 0.902 in the training and testing sets, respectively. In contrast to their focus solely on tumor diameters below 4 cm, our study extends beyond this limitation, enhancing its applicability in clinical settings. Shu et al. [27] extracted radiomics features from CMP and NP images of 260 ccRCC patients and developed CMP model, NP model and combined model for differentiating Fuhrman grade. The results showed that the NP model had the highest diagnostic efficacy with an AUC value of 0.818 and the CMP model had an AUC value of 0.766, the difference was not statistically significant (\u003cem\u003eP\u003c/em\u003e=0.0844). This is not quite consistent with our findings, where our deep learning model based on CMP images had significantly better AUC values than the deep learning model based on NP images, and the difference was statistically significant. The reason for this discrepancy may be because radiomics and deep learning belong to two different research methods, and there are also significant differences between the Fuhrman grading system and the WHO/ISUP grading system. In previous studies, most of them were based on the Fuhrman grading system [27-29]. The Fuhrman grading system, proposed by Fuhrman in 1982, is a widely used pathological grading system for renal cell carcinoma [28]. This grading system is only based on the analysis results of 103 cases of renal cell carcinoma, of which only 85 cases were followed up, and does not consider the tissue type of renal cell carcinoma. In practice, the classification system has some problems, such as difficulty in interpretation and poor repeatability [29, 30]. Therefore, in the 2016 edition of the WHO New Classification of Renal Tumors, this system was replaced by a new grading standard called the WHO/ISUP grading system [31]. Because the latter grading system is superior to the former and has better interpretability [30-32]. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; Deep learning models have been widely used in previous studies to evaluate kidney tumors [15, 16, 33-37]. Toda et al. [34] developed a deep learning-based algorithm for fully automated detection of small (≤4cm) RCC in contrast-enhanced CT images and evaluated its performance. The AUC values of this deep learning algorithm were 0.930 and 0.933 in dataset A and dataset B, respectively. Han et al.\u0026nbsp;[35]\u0026nbsp;constructed a deep learning framework to distinguish the three major subtypes of RCC (ccRCC, pRCC and chRCC) based on three-phase CT images, and the network had an accuracy of 0.85, sensitivity of 0.64-0.98, specificity of 0.83-0.93 and AUC value of 0.90. Xi et al.\u0026nbsp;[38]\u0026nbsp;collected MRI images (T1-enhanced and T2WI) from a total of 1162 patients with renal tumors and combined clinical characteristics to build a deep learning model based on ResNet to identify benign and malignant renal tumors. The accuracy, sensitivity and specificity of the deep learning model were 0.70, 0.92 and 0.41, respectively. Lin et al.\u0026nbsp;[16]\u0026nbsp;used multiple methods (image cropping, setting the attention level, selecting model complexity, and applying transfer learning) to build deep learning models used to distinguish high-grade ccRCC from low-grade ccRCC with AUC values of 0.82±0.11 and 0.81±0.04 in the internal test set and external test set, respectively. Xu et al.\u0026nbsp;[15]\u0026nbsp;divided 706 ccRCC patients into training and validation cohorts and built a deep learning framework initialized by self-supervised pre-training to identify high-grade ccRCC with AUC values of 0.864-0.882. In our study, a total of six deep learning models were built based on UP, CMP and NP images, three of which were strongly supervised and three were weakly supervised. Previous studies have mainly built deep models based on strong supervision, and few studies have built deep learning models based on weak supervision to distinguish high-grade ccRCC from low-grade ccRCC. The use of weakly supervised deep learning model would help physicians to save a lot of time, and the models has good performance in distinguishing WHO/IUSP grade of ccRCC than previous studies.\u003c/p\u003e\n\u003cp\u003eResNet networks are increasingly being used in deep learning studies of medical images. Xu et al. [37] collected MRI images (T2WI and DWI) from a total 217 renal tumor patients and built a deep learning model based on the ResNet-18 network to assess the benignity and malignancy of tumors. And compared it with the radiomics model, which resulted in a significantly better deep learning model than the radiomics model with an AUC value of 0.925. Lin et al. [16] built several deep learning models based on ResNet-18, ResNet-34, and ResNet-50 networks to evaluate the WHO/ISUP grade of ccRCC. The results showed that the deep learning models built based on ResNet-18 network provided the best results in differentiation. The above study showed that the RseNet-18 network was widely used in deep learning research. The six deep learning models established in our study were all built based on the ResNet-18 network which is different from the neural networks used by some other studies. For example, Nikpanah et al. [36] built a model based on the AlexNet network to distinguish between ccRCC and oncocytoma. Zuo et al. [39] used six neural networks (MobileNetV2, Efficient Net, ShuffleNet, ResNet-34, ResNet-50 and ResNet-101) to build a deep learning model to distinguish between pRCC and chRCC, with the model built on the MobileNetV2 network having better diagnostic results.\u003c/p\u003e\n\u003cp\u003eThere are still some limitations in this study. First, although the samples in this study were obtained from two large hospitals, we did not have an advantageous sample size compared with previous deep learning studies and hope to prospectively collect more cases for the study in the future. Second, this study only used the ResNet-18 network to build a deep learning model and did not use neural networks such as VGGNet and AlexNet to build a deep learning model to evaluate the WHO/ISUP grade of ccRCC. Third, this study built a deep learning model to distinguish high-grade ccRCC from low-grade ccRCC but did not collect clinical features to build a clinical model to assess its grade. Clinical models and radiomics models can be added in future studies to assess the WHO/ISUP grade of ccRCC.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn our study, we developed a weakly supervised deep learning model based on the 2D ResNet-18 network to identify focal locations for ccRCC and further evaluate its WHO/ISUP grading. These findings confirms that our weakly supervised deep learning model performs either favorably or comparably in terms of grade assessment of ccRCC with biopsies whilst enjoying the non-invasive and labor-saving property, which provides the original basis for the use of weakly supervised deep learning models in big data in the future. However, a larger sample size and prospective studies are needed to validate that the model can be used as a routine clinical tool.\u0026nbsp;\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData Availability Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData will be available upon request. Because the data used in this study comes from real patient data, which is related to the privacy of patients, if necessary, it is necessary to apply to the hospital management system for use, and can be uploaded to the public database after the hospital agrees.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors thank all staff members involved in the acquisition of data. They are grateful to the technical assistance provided by Huiying Medical Technology (Beijing).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Statement\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn accordance with the Declaration of Helsinki, the studies involving human participants were reviewed and approved by the first affiliated hospital of Anhui medical university’s Ethics Committee (ID: PJ2022-06-48). Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCuiping Li, Xingwang Wu, Yankun Gao and Chao Zhu: conception and design. Cuiping Li, Yankun Gao and Shuai Li: collection and arrangement of data. Cuiping Li and Yankun Gao: data analysis and manuscript writing. Baoxin Qian: Data analysis, validation, and visualization. Jianying Li, Xingwang Wu and Yankun Gao: review \u0026amp; editing. All authors contributed to the article and approved the submitted version.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors received no funding for this work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of interest statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e"},{"header":"Reference","content":"\u003col\u003e\n \u003cli\u003eMotzer R J, Jonasch E, Michaelson M D, et al. NCCN Guidelines Insights: Kidney Cancer, Version 2.2020. J Natl Compr Canc Netw 2019; 17:1278-1285.\u003c/li\u003e\n \u003cli\u003eCapitanio U, Cloutier V, Zini L, et al. A critical assessment of the prognostic value of clear cell, papillary and chromophobe histological subtypes in renal cell carcinoma: a population-based study. BJU International 2009; 103:1496-1500.\u003c/li\u003e\n \u003cli\u003eKeegan K A, Schupp C W, Chamie K, et al. Histopathology of surgically treated renal cell carcinoma: survival differences by subtype and stage. J Urol 2012; 188:391-397.\u003c/li\u003e\n \u003cli\u003eMoch H, Cubilla A L, Humphrey P A, 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.\u003c/li\u003e\n \u003cli\u003eCaoili E M and Davenport M S Role of percutaneous needle biopsy for renal masses. Semin Intervent Radiol 2014; 31:20-26.\u003c/li\u003e\n \u003cli\u003eVolpe A, Mattar K, Finelli A, et al. Contemporary results of percutaneous biopsy of 100 small renal masses: a single center experience. J Urol 2008; 180:2333-2337.\u003c/li\u003e\n \u003cli\u003eHarris C R, Whitson J M, and Meng M V Under-grading of \u0026lt;4 cm renal masses on renal biopsy. BJU Int 2012; 110:794-797.\u003c/li\u003e\n \u003cli\u003eShin T, Duddalwar V A, Ukimura O, et al. Does Computed Tomography Still Have Limitations to Distinguish Benign from Malignant Renal Tumors for Radiologists? Urol Int 2017; 99:229-236.\u003c/li\u003e\n \u003cli\u003eKim J H, Sun H Y, Hwang J, et al. Diagnostic accuracy of contrast-enhanced computed tomography and contrast-enhanced magnetic resonance imaging of small renal masses in real practice: sensitivity and specificity according to subjective radiologic interpretation. World J Surg Oncol 2016; 14:260.\u003c/li\u003e\n \u003cli\u003eChen S H, Wu Y P, Li X D, et al. R.E.N.A.L. Nephrometry Score: A Preoperative Risk Factor Predicting the Fuhrman Grade of Clear-Cell Renal Carcinoma. J Cancer 2017; 8:3725-3732.\u003c/li\u003e\n \u003cli\u003eOh S, Sung D J, Yang K S, et al. Correlation of CT imaging features and tumor size with Fuhrman grade of clear cell renal cell carcinoma. Acta Radiol 2017; 58:376-384.\u003c/li\u003e\n \u003cli\u003eZhu Y H, Wang X, Zhang J, et al. Low enhancement on multiphase contrast-enhanced CT images: an independent predictor of the presence of high tumor grade of clear cell renal cell carcinoma. AJR Am J Roentgenol 2014; 203:W295-300.\u003c/li\u003e\n \u003cli\u003eZhao J, Zhang P, Chen X, et al. Lesion Size and Iodine Quantification to Distinguish Low-Grade From High-Grade Clear Cell Renal Cell Carcinoma Using Dual-Energy Spectral Computed Tomography. J Comput Assist Tomogr 2016; 40:673-677.\u003c/li\u003e\n \u003cli\u003eChen C, Kang Q, Xu B, et al. Differentiation of low- and high-grade clear cell renal cell carcinoma: Tumor size versus CT perfusion parameters. Clin Imaging 2017; 46:14-19.\u003c/li\u003e\n \u003cli\u003eXu L, Yang C, Zhang F, et al. Deep Learning Using CT Images to Grade Clear Cell Renal Cell Carcinoma: Development and Validation of a Prediction Model. Cancers (Basel) 2022; 14:\u003c/li\u003e\n \u003cli\u003eLin F, Ma C, Xu J, et al. A CT-based deep learning model for predicting the nuclear grade of clear cell renal cell carcinoma. Eur J Radiol 2020; 129:109079.\u003c/li\u003e\n \u003cli\u003eYang G, Wang C, Yang J, et al. Weakly-supervised convolutional neural networks of renal tumor segmentation in abdominal CTA images. BMC Med Imaging 2020; 20:37.\u003c/li\u003e\n \u003cli\u003eWang, C. W., Y. C. Lee, C. C. Chang, et al., A Weakly Supervised Deep Learning Method for Guiding Ovarian Cancer Treatment and Identifying an Effective Biomarker. Cancers (Basel), 2022. 14(7).\u003c/li\u003e\n \u003cli\u003eLiu, M. Z., C. Swintelski, S. Sun, et al., Weakly Supervised Deep Learning Approach to Breast MRI Assessment. Acad Radiol, 2022. 29 Suppl 1: p. S166-S172.\u003c/li\u003e\n \u003cli\u003eZhou, C., Y. Jin, Y. Chen, et al., Histopathology classification and localization of colorectal cancer using global labels by weakly supervised deep learning. Comput Med Imaging Graph, 2021. 88: p. 101861.\u003c/li\u003e\n \u003cli\u003eZhou, J., L. Y. Luo, Q. Dou, et al., Weakly supervised 3D deep learning for breast cancer classification and localization of the lesions in MR images. J Magn Reson Imaging, 2019. 50(4): p. 1144-1151.\u003c/li\u003e\n \u003cli\u003eEbrahimi A, Luo S, and Chiong R. \u003cem\u003eIntroducing transfer learning to 3D ResNet-18 for Alzheimer\u0026rsquo;s disease detection on MRI images\u003c/em\u003e. in \u003cem\u003e2020 35th international conference on image and vision computing New Zealand (IVCNZ)\u003c/em\u003e. 2020. IEEE.\u003c/li\u003e\n \u003cli\u003eMa L, Wang Y, Guo L, et al. Developing and verifying automatic detection of active pulmonary tuberculosis from multi-slice spiral CT images based on deep learning. J Xray Sci Technol 2020; 28:939-951.\u003c/li\u003e\n \u003cli\u003eZabihollahy F, Schieda N, Krishna S, et al. Automated classification of solid renal masses on contrast-enhanced computed tomography images using convolutional neural network with decision fusion. Eur Radiol 2020; 30:5183-5190.\u003c/li\u003e\n \u003cli\u003ePedersen M, Andersen M B, Christiansen H, et al. Classification of renal tumour using convolutional neural networks to detect oncocytoma. Eur J Radiol 2020; 133:109343.\u003c/li\u003e\n \u003cli\u003eGao Y, Wang X, Zhao X, et al. Multiphase CT radiomics nomogram for preoperatively predicting the WHO/ISUP nuclear grade of small (\u0026lt; 4 cm) clear cell renal cell carcinoma. BMC Cancer 2023; 23:953.\u003c/li\u003e\n \u003cli\u003eShu J, Tang Y, Cui J, et al. Clear cell renal cell carcinoma: CT-based radiomics features for the prediction of Fuhrman grade. Eur J Radiol 2018; 109:8-12.\u003c/li\u003e\n \u003cli\u003eDing J, Xing Z, Jiang Z, et al. CT-based radiomic model predicts high grade of clear cell renal cell carcinoma. Eur J Radiol 2018; 103:51-56.\u003c/li\u003e\n \u003cli\u003eHussain M A, Hamarneh G, and Garbi R Learnable image histograms-based deep radiomics for renal cell carcinoma grading and staging. Comput Med Imaging Graph 2021; 90:101924.\u003c/li\u003e\n \u003cli\u003eShu J, Wen D, Xi Y, et al. Clear cell renal cell carcinoma: Machine learning-based computed tomography radiomics analysis for the prediction of WHO/ISUP grade. Eur J Radiol 2019; 121:108738.\u003c/li\u003e\n \u003cli\u003eWang R, Hu Z, Shen X, et al. Computed Tomography-Based Radiomics Model for Predicting the WHO/ISUP Grade of Clear Cell Renal Cell Carcinoma Preoperatively: A Multicenter Study. Front Oncol 2021; 11:543854.\u003c/li\u003e\n \u003cli\u003eZhao Y, Chang M, Wang R, et al. Deep Learning Based on MRI for Differentiation of Low- and High-Grade in Low-Stage Renal Cell Carcinoma. J Magn Reson Imaging 2020; 52:1542-1549.\u003c/li\u003e\n \u003cli\u003eWen-Zhi G, Tai T, Zhixin F, et al. Prediction of pathological staging and grading of renal clear cell carcinoma based on deep learning algorithms. J Int Med Res 2022; 50:3000605221135163.\u003c/li\u003e\n \u003cli\u003eToda N, Hashimoto M, Arita Y, et al. Deep Learning Algorithm for Fully Automated Detection of Small (\u0026lt;/=4 cm) Renal Cell Carcinoma in Contrast-Enhanced Computed Tomography Using a Multicenter Database. Invest Radiol 2022; 57:327-333.\u003c/li\u003e\n \u003cli\u003eHan S, Hwang S I, and Lee H J The Classification of Renal Cancer in 3-Phase CT Images Using a Deep Learning Method. Journal of Digital Imaging 2019; 32:638-643.\u003c/li\u003e\n \u003cli\u003eNikpanah M, Xu Z, Jin D, et al. A deep-learning based artificial intelligence (AI) approach for differentiation of clear cell renal cell carcinoma from oncocytoma on multi-phasic MRI. Clin Imaging 2021; 77:291-298.\u003c/li\u003e\n \u003cli\u003eXu Q, Zhu Q, Liu H, et al. 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 2022; 55:1251-1259.\u003c/li\u003e\n \u003cli\u003eXi I L, Zhao Y, Wang R, et al. Deep Learning to Distinguish Benign from Malignant Renal Lesions Based on Routine MR Imaging. Clin Cancer Res 2020; 26:1944-1952.\u003c/li\u003e\n \u003cli\u003eZuo T, Zheng Y, He L, et al. Automated Classification of Papillary Renal Cell Carcinoma and Chromophobe Renal Cell Carcinoma Based on a Small Computed Tomography Imaging Dataset Using Deep Learning. Front Oncol 2021; 11:746750.\u003cstrong\u003e\u003c/strong\u003e\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1.\u0026nbsp;\u003c/strong\u003ePatient demographics and characteristics of ccRCC in training and testing set.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003eVariable\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003eTraining set (n=214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003eTesting set (n=92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003et/Z value\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u003cem\u003eP\u003c/em\u003e-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003eAge, year\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e58.2\u0026plusmn;11.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e55.5\u0026plusmn;12.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e1.833\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e0.068\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003eSex, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e1.485\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e0.223\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003eMale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e65% (139/214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e57.6% (53/92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e35% (75/214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e42.4% (39/92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003eTumor diameter, cm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e5.2\u0026plusmn;2.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e5.3\u0026plusmn;2.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e-0.313\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e0.754\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026le;4cm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e42.1% (90/214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e35.9% (33/92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e1.024\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e0.311\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003e\u0026nbsp; ＞4cm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e57.9% (124/214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e64.1% (59/92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003eWHO/ISUP grade\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e0.121\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e0.728\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003e\u0026nbsp; High-grade\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e46.7% (100/214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e44.6% (41/92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003e\u0026nbsp; Low-grade\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e53.3% (114/214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e55.4% (51/92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003eSide, %\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e1.494\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e0.222\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003e\u0026nbsp; Right\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e46.7% (100/214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e45.7% (42/92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 102px;\"\u003e\n \u003cp\u003e\u0026nbsp; Left\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 143px;\"\u003e\n \u003cp\u003e53.3% (114/214)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 120px;\"\u003e\n \u003cp\u003e54.3% (50/92)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 90px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eNote: ccRCC, clear cell renal cell carcinoma; WHO/ISUP, the World Health Organization/International Society of Urological Pathology.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u003c/strong\u003e Diagnostic performance measures of the 3D ResNet-18 deep learning models in the training and testing sets.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"586\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 56px;\"\u003e\n \u003cp\u003eModels\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 153px;\"\u003e\n \u003cp\u003eAUC (95% CI)\u003c/p\u003e\n \u003cp\u003etraining / testing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003cp\u003etraining / testing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 89px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003cp\u003etraining /testing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 82px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003cp\u003etraining /testing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003eDelong Test\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value_test\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 56px;\"\u003e\n \u003cp\u003eCMP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 153px;\"\u003e\n \u003cp\u003e0.944 (0.917-0.971) /\u003c/p\u003e\n \u003cp\u003e0.922 (0.867-0.977)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.846 /\u003c/p\u003e\n \u003cp\u003e0.848\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 89px;\"\u003e\n \u003cp\u003e0.800 /\u003c/p\u003e\n \u003cp\u003e0.854\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 82px;\"\u003e\n \u003cp\u003e0.886 /\u003c/p\u003e\n \u003cp\u003e0.843\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 56px;\"\u003e\n \u003cp\u003eNP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 153px;\"\u003e\n \u003cp\u003e0.920 (0.886-0.954) /\u003c/p\u003e\n \u003cp\u003e0.900 (0.840-0.960)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.832 /\u003c/p\u003e\n \u003cp\u003e0.826\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 89px;\"\u003e\n \u003cp\u003e0.823 /\u003c/p\u003e\n \u003cp\u003e0.805\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 82px;\"\u003e\n \u003cp\u003e0.839 /\u003c/p\u003e\n \u003cp\u003e0.843\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.23\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 56px;\"\u003e\n \u003cp\u003eUP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 153px;\"\u003e\n \u003cp\u003e0.894 (0.854-0.934) /\u003c/p\u003e\n \u003cp\u003e0.875 (0.805-0.945)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 98px;\"\u003e\n \u003cp\u003e0.799 /\u003c/p\u003e\n \u003cp\u003e0.793\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 89px;\"\u003e\n \u003cp\u003e0.810 /\u003c/p\u003e\n \u003cp\u003e0.822\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 82px;\"\u003e\n \u003cp\u003e0.789 /\u003c/p\u003e\n \u003cp\u003e0.766\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026lt;0.05\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eNote: AUC, area under the curve; CI, confidence interval; CMP, corticomedullary phase; NP, nephrographic phase; UP, unenhanced phase.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u0026nbsp;\u003c/strong\u003eDiagnostic performance measures of the 2D ResNet-18 deep learning models in the training and testing sets.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"709\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 54px;\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 59px;\"\u003e\n \u003cp\u003eSplit\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" style=\"width: 350px;\"\u003e\n \u003cp\u003eSingle slice\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" style=\"width: 227px;\"\u003e\n \u003cp\u003e\u0026gt;3 slices\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 128px;\"\u003e\n \u003cp\u003eAUC (95%CI)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 54px;\"\u003e\n \u003cp\u003eCMP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003etraining\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 128px;\"\u003e\n \u003cp\u003e0.912(0.897-0.926)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.926\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.927\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.924\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.864\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.865\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.864\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003etesting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 128px;\"\u003e\n \u003cp\u003e0.907(0.888-0.927)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.862\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.863\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.861\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.859\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.857\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.860\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 54px;\"\u003e\n \u003cp\u003eNP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003etraining\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 128px;\"\u003e\n \u003cp\u003e0.893(0.877-0.909)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.914\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.911\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.918\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.846\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.835\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.855\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003etesting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 128px;\"\u003e\n \u003cp\u003e0.884(0.863-0.906)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.813\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.800\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.831\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.837\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.833\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.840\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 54px;\"\u003e\n \u003cp\u003eUP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003etraining\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 128px;\"\u003e\n \u003cp\u003e0.876(0.859-0.893)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.897\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.912\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.875\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.822\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.804\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.838\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 59px;\"\u003e\n \u003cp\u003etesting\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 128px;\"\u003e\n \u003cp\u003e0.866(0.843-0.890)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 71px;\"\u003e\n \u003cp\u003e0.772\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.841\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.687\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.793\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.786\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 76px;\"\u003e\n \u003cp\u003e0.800\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;Note: AUC, area under the curve; CI, confidence interval; CMP, corticomedullary phase; NP, nephrographic phase; UP, unenhanced phase.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-cancer","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bcan","sideBox":"Learn more about [BMC Cancer](http://bmccancer.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bcan/default.aspx","title":"BMC Cancer","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"clear cell renal cell carcinoma, weakly supervised deep learning, WHO/IUSP nuclear grade, computer tomography","lastPublishedDoi":"10.21203/rs.3.rs-6489710/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6489710/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003ePurpose:\u003c/strong\u003e The study aims to evaluate the performance of weakly supervised deep learning models in distinguishing high-grade and low-grade clear cell renal cell carcinoma (ccRCC) in comparison to strongly supervised deep learning models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethod: \u003c/strong\u003ePathologically confirmed ccRCC from two hospitals between January 2017 and April 2022 were included. The strongly and weakly supervised deep learning models were on three-phase images (CMP, NP and UP) based on the 3D ResNet-18 network and 2D ResNet-18 network, respectively using three-phase images (CMP, NP and UP) to form six deep learning models. Accuracy, sensitivity, specificity, and area-under-the-curve (AUC) were used to assess the discriminatory efficacy of the deep learning models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eA total of 306 ccRCC patients were collected in this retrospective study. Among them, 165 were low-grade ccRCC, and 141 were high-grade ccRCC. Data were divided into a training set (n=214) and a testing set (n=92) according to the ratio of 7:3. Among the weakly supervised deep learning models based on three-phase images, the model based on CMP images has the highest diagnostic performance, and its accuracy, sensitivity, specificity, and AUC values are 0.859, 0.857, 0.860 and 0.907, respectively. These values had no significant difference from the strongly supervised deep learning model based on CMP images (the accuracy, sensitivity, specificity, and AUC values were 0.848, 0.854, 0.843 and 0.922, respectively).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion: \u003c/strong\u003eThe weakly supervised deep learning model developed in this study using CMP images has the same high diagnostic performance as the strongly supervised deep learning model in distinguishing high-grade ccRCC from low-grade ccRCC. With richer samples and sufficient developing, the weakly supervised deep learning model may become a routine clinical tool to reduce the physical toll of biopsy on patients.\u003c/p\u003e","manuscriptTitle":"A weakly supervised deep learning model based on CT images for predicting WHO/ISUP nuclear grade of clear cell renal cell carcinoma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-04 08:34:45","doi":"10.21203/rs.3.rs-6489710/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"253640400568811598770920083839782440179","date":"2025-06-18T07:22:06+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-30T13:57:55+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-05-08T15:08:32+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-29T04:26:09+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-29T04:23:56+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Cancer","date":"2025-04-20T14:01:50+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-cancer","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bcan","sideBox":"Learn more about [BMC Cancer](http://bmccancer.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bcan/default.aspx","title":"BMC Cancer","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3663598d-2d3f-4584-9e4d-bb594306278a","owner":[],"postedDate":"June 4th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-06-04T08:34:45+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-04 08:34:45","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6489710","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6489710","identity":"rs-6489710","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}