Nephroblastoma, also known as Wilms tumor, is the most common renal tumor in pediatric patients (1), accounting for more than 90% of all malignant kidney tumors in children (2), it is also the most frequent pediatric abdominal malignant tumor and the fourth most frequent pediatric malignant tumor overall (3). Under the comprehensive treatment of surgery, chemotherapy, and radiotherapy, the five-year survival rate of nephroblastoma patients has exceeded 90% (4). The current focus of treatment is on how to better protect kidney function and improve long-term survival quality based on ensuring survival rates.

Nephron sparing surgery (NSS) helps to protect the kidneys in the long term and prevent the development of chronic kidney disease, which is especially important in children with a long survival period and facing progressive renal insufficiency (5, 6). Since the successful application in bilateral nephroblastoma, NSS has also been increasingly used to remove selected cases of unilateral nephroblastoma (7–9). Studies have found that in the group of patients with localized disease, the event-free survival and overall survival after NSS appeared to be as good as after total nephrectomy (TN), and local recurrence rates were the same as after TN (8). The application of NSS in unilateral nephroblastoma is limited due to the possible risk of recurrence caused by positive margins. If nephroblastoma patients without the involvement of renal capsule, sinus vessels, and lymph nodes (named localized stage here) can be accurately distinguished before surgery, it will be of great significance to the management of nephroblastoma and will promote the precision treatment of nephroblastoma.

Radiomics is a novel technology that can extract a large number of quantitative information describing the pathophysiological status and phenotypic characteristics of lesions from CT, PET, MRI and other medical images (10–12). Models can be built through in-depth analysis of the extracted mineable radiomic features, combined with clinical examinations and testing information, and using suitable machine learning algorithms to provide clinical decision support. It has shown great advantages in metastasis prediction (13), tumor diagnosis (14), treatment response evaluation (15), and prognosis analysis (16).

The purpose of this study is to establish a radiomic approach based on CT images to develop a non-invasive tool for preoperative prediction of localized nephroblastoma of pediatric patients.

A total of 105 patients (50 males/55 females) were enrolled in the study, including 59 cases with non-localized stage and 46 cases with localized stage. They were divided into training cohort (n= 73, 37 males/36 females) and validation cohort (n= 32, 13 males/19 females) according to the order of CT scanning time. Age of the total study sample ranged from 0.1 to 10.3 years, with a median age of 2.1 years. Age of the training cohort ranged from 0.1 to 7.9 years with a median age of 1.9 years, whereas age of the validation cohort ranged from 0.1 to 10.3 years with a median age of 2.9 years.

A total of 1652 radiomic features have been extracted on each VOI. For original interpolated VOIs, 105 radiomic features were obtained including 14 shape features, 18 first order statistics features and 73 texture features. The remaining 1547 radiomic features were extracted from derived interpolated VOIs. They consisted of 306 first order statistics features and 1241 texture features. ICC analysis showed that the radiomic features from the VOIs delineated by the two senior radiologists had good inter- and intra- observer repeatability (ICC ranged from 0.73 to 1.0).

In order to determine the optimal parameter combination, we have adjusted different parameters (including the voxel spacing of isotropic data, the Pearson threshold and the number of retained features) and then established the corresponding LR models in the training cohort. The LR model with the optimal parameter combination was chosen as the final LR model. We found that the optimal parameter combination was the voxel spacing of 1×1×1 mm³, the Pearson threshold of 0.90 and the feature number of 10. When one parameter remains unchanged and other parameters change, the AUC values of the LR models are shown in Figure 3 . The AUC value decreases with the increase of the voxel spacing of isotropic data, illustrating that clearer image data is more conducive to radiomics analysis. A total of 11 patients (including 4 cases of non-localized stage and 7 cases of localized stage) were able to correctly predict whether they were localized or not, even if the voxel spacing was changed. As for the Pearson threshold and the number of retained features, no clear pattern was observed in this single-center study. In the Pearson correlation analysis, a total of 354 radiomic features (including 3 shape features, 73 first order statistics features and 278 texture features) were obtained with threshold 0.90. About 78.57% of radiomic features have been eliminated. After the SFFS feature-ranking algorithm, top 10 radiomic features (including 3 first order statistics features and 7 texture features) were selected as our radiomic signature ( Table 2 ).

In addition to the final LR model, SVM-RFE, RF, AdaBoost, GBDT, and XGBoost models were constructed using the optimal parameter combination and the corresponding machine learning classifiers. Table 3 shows the AUC values of these models for predicting the localized status in patients with nephroblastoma. Although the same radiomic signature was used, only AUC values of the final LR model were higher than 0.70 in the training and validation cohorts (training cohort, 0.796; validation cohort, 0.710). Other prediction models performed poorly in distinguishing localized group from non-localized group. It designates the potential of LR-based radiomics in classifying localized and non-localized nephroblastoma in pediatric patients. For the final LR model, the radiomic signature calculation formula is presented as follows: Rad_ signature = 0.22 - 0.52 × Feature_1 - 0.56 × Feature_2 - 0.21 × Feature_3 - 0.35 × Feature_4 - 0.84 × Feature_5 - 0.02 × Feature_6 + 0.58 × Feature_7 + 0.27 × Feature_8 + 0.33 × Feature_9 - 0.31 × Feature_10.

We further explored the performances of our radiomic signature calculation formula based on the final LR model using more performance metrics (such as accuracy, sensitivity and specificity). All performance metrics of the final LR model for predicting the localized nephroblastoma are presented in Figure 4 . The AUC, accuracy, sensitivity and specificity are 0.796, 0.795, 0.732 and 0.875 for the training cohort respectively, and 0.710, 0.719, 0.611 and 0.857 for the validation cohort respectively. Results manifest the effectiveness in distinguishing patients with localized and non-localized nephroblastoma ( Figure 5 ).

The overall survival outcomes of most nephroblastoma patients are good, and further studies should focus on how to use the optimal treatment under more precise risk-stratified strategies (24) to improve the long-term prognosis of patients with nephroblastoma. Our result can be regarded as a kind of non-invasive risk-stratified strategies prior to surgery for nephroblastoma patients using radiomics technique, which goes beyond the ability of the human eye to identify the completely localized status in CT images.

Unlike previous surveys that concerned of all stages of nephroblastoma, our current study highlighted on the detection of nephroblastoma patients at a very early stage without involvement of renal capsule, sinus vessels, and lymph nodes. These patients can not only undergo complete tumor resection, but also have the potential to perform nephron sparing nephrectomy. Therefore, we developed a new method to divide patients into localized group and non-localized group non-invasively before operation, so as to guide the personalized treatment of nephroblastoma. To the best of our knowledge, the relationship between radiomic features and the involvement of renal capsule, sinus vessels, and lymph nodes of nephroblastoma has not been evaluated previously.

At present, lots of studies have shown that radiomic technology has the power to predict metastasis of different diseases. For example, one of our previous studies showed that radiomic features could effectively predict preoperative lymph node metastasis in patients with gastric cancer (13). In that study, tumor radiomics and LN radiomics were integrated to build an effective non-invasive tool to guide treatment decision-making in gastric cancer, which was notably useful in patients with T2-stage, diffuse subtype, and moderately/well differentiated gastric cancer. Zhao et al. found that radiomic features from preoperative CT images could predict future distant metastasis of local renal cell carcinoma after surgical resection, and these predictive radiomic features were related to certain important biological pathways (extracellular matrix-receptor interaction, focal adhesion and Phosphoinositide 3-kinases/Akt pathways) (25). Their study suggested that these findings provided support for the biological interpretation of radiomics model and might enhance the molecular basis of radiomics. Both of the above studies showed that the radiomic technology has offered a possible solution for studying the correlation between medical imaging and tumor metastasis in patients with nephroblastoma.

In this study, we found that the radiomic features extracted from enhanced CT images of preoperative nephroblastoma patients could be used as an independent predictor of localized nephroblastoma. 10 radiomic features related to early localized stage of nephroblastoma were screened out from a large amount of data through a generalized methodology for radiomic feature selection and modeling developed by our team (23), in which Pearson correlation analysis and SFFS algorithm were used for feature selection. The methodology was verified by our own three data sets (Gastric cancer dataset, Osteosarcoma dataset, Pancreatic neuroendocrine tumors dataset) and performed well in these three different origin and different types of solid tumors, which proves the stability and reliability of this methodology in feature selection. Similarly, the results of our LR-based model in this study demonstrated the effectiveness of radiomic features based on preoperative enhanced CT images for predicting localized nephroblastoma at the training cohort (AUC, 0.796; accuracy, 0.795; sensitivity, 0.732; specificity, 0.875) and the validation cohort (AUC, 0.710; accuracy, 0.719; sensitivity, 0.611; specificity, 0.857).

In addition to a single machine learning classifier, the performance of different classifiers based on same radiomic features was also worth exploring. Parmar et al. found that classifiers had a significant impact on the model performance (accounting for 34.21% of the total model differences) (26), so it was a very important step to determine the best machine learning classifier and apply it to radiomics. Our results also supported this conclusion. Through the comparison of 6 machine learning classifiers (LR, SVM-RFE, RF, AdaBoost, GBDT, and XGBoost), we found that LR was the most suitable classifier for our dataset –only the AUC of LR was higher than 0.75 in training cohort (training cohort: 0.796), and it performed equally well on the validation cohort (validation cohort: 0.710). Dong et al. applied multiple machine learning methods based on radiomics to differentiate intracranial ependymoma from medulloblastoma and found that the performance of RF classifier was the best (AUC 0.91,95% CI 0.787-0.968), followed by Support Vector Machines and K neighbors, and Adboost was the worst (AUC 0.75,95% CI 0.604-0.857) (27). Then Paired test showed significant difference (P<0.05) in ROC curve among the classifiers used. In this study, we also compared the model performance of LR with other machine learning methods, which used the same radiomic features for predicting the absence of involvement of the renal capsule, sinus vessels, and lymph nodes in nephroblastoma. The results illustrated that the radiomic model of LR achieved the best performance both in the training cohort and validation cohort (P<0.05).

By comparing radiomic features based on three different types of interpolated images, it was found that features extracted from images with the smallest voxel spacing (1 × 1 × 1 mm³) resulted in the best model accuracy, which was consistent with the truth that smaller voxel associated with the higher resolution (28), even though this smaller voxel was obtained by a resampling technique.

Despite the aforementioned important findings, our study still had several limitations. First of all, this study was retrospective and had a relatively small size of cohort, although external validation was used. In order to obtain a higher level of clinical application evidence, it needed to be verified with prospective and larger data sets. Secondly, our study did not take into account some other features, such as clinical features, genomics and so on. The combination of multi-dimensional data may further improve the predictive ability of the model and thus providing better clinical decision support.

Overall, the results showed that the LR-based radiomic model had a good ability of non-invasive prediction of nephroblastoma without renal capsule, sinus vessels and lymph node involvement., and had the potential to help make surgical decisions for precise treatment and improve the long-term quality of life of nephroblastoma patients. However, more data is needed to improve the global applicability and robustness of the method.

The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding authors.

The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by the ethics committee of the Children’s Hospital, Zhejiang University School of Medicine (2021-IRB-097) and individual consent for this retrospective analysis was waived.

X-HM and JY designed this project and took charge of the development of machine learning and writing this paper. H-CZ and XJ delineated the area of interest in our medical images. J-WL and Y-SD had equal contributions in preparing figures and clinical information. TN and QS supervised the development of this project and guided the writing of this article. X-HM and JY takes mainly responsibility for the integrity of the content of the paper and each author have participated sufficiently in the submission to take public responsibility for the content. All authors contributed to the article and approved the submitted version.