In this retrospective study, a machine learning model was developed and validated for noninvasive, and individualized prediction of the risk of ccRCC metastasis. This model incorporated clinical-radiological features as well as ultrasound and CT radiomics signature, which exhibited superior predictive performance (training set: AUC = 0.924, 95%CI = 0.887–0.961; testing set 1: AUC = 0.877, 95%CI = 0.771–0.984; testing set 2: AUC = 0.849, 95%CI = 0.754–0.945) in predicting the risk of ccRCC metastasis, along with reliable and reproducible outcomes.
ccRCC is a common cancer with an annual prevalence of 330 000 cases annually[18, 19]. The prognosis for non-metastatic renal cell carcinoma is favorable following partial or radical nephrectomy[20]. However, one-third of ccRCC patients are associated with metastasis[21]. At present, there are limited options for treating patients with metastatic renal cell carcinoma, including cancer-targeted surgery and immune checkpoint inhibitor drugs. However, these treatments can be costly and may not yield optimal results[22, 23]. Moreover, the primary method of detecting ccRCC metastasis is through the use of enhanced CT scan. Despite its widespread use, CT imaging has limitations and provides minimal prediction of the risk of ccRCC metastasis[24–27]. Thus, it is crucial to develop accurate methods for predicting the risk of metastasis in patients, strengthen the active monitoring by physicians, and promptly intervene in patient treatment when necessary. Radiomics is a technique that converts medical images into high-throughput data, which can be used to provide valuable insights into the diagnosis, prognosis, staging, and treatment responses of cancer patients[28–31]. Radiomics has demonstrated outstanding performance in these areas. Numerous studies have indicated that machine learning models based on CT radiomics can accurately predict the risk of ccRCC metastasis[32–34]. Although machine learning models based on radiomics features have shown promise in predicting ccRCC metastasis, many of these studies have overlooked the significance of incorporating clinical and radiological features into the models.
While radiomics-derived information are a valuable component of computerized clinical decision support systems, they are not a universal solution. This study analyzed subtle differences in the clinical-radiological characteristics, CT, and ultrasound radiomics features of patients with ccRCC. With regard to clinical data, we focused on different indicators that can be significant predictors for predicting the risk of ccRCC metastasis. In terms of radiological data, we placed particular emphasis on features such as necrosis, intratumoral hemorrhage, calcification, capsule, and intravascular tumor thrombus, which are known to be associated with an unfavorable prognosis.
Previous research has demonstrated that CCRCC is a tumor with high levels of angiogenesis and vascularity[35]. Ficarra et al. [36] showed that intra-tumor necrosis was an important prognostic indicator in the clinical management of ccRCC patients according to the Mayo Clinic's Staging, Size, Grade, and Necrosis (SSIGN) scoring system[37]. Our findings are in agreement with previous studies, showing that necrosis, intravascular tumor thrombus, and maximum tumor diameter may be independent factors associated with metastasis. These factors have been widely used to establish a clinical-radiological model. Among the clinical models based on clinical-radiological characteristics, the ExtraTrees model exhibited the most outstanding performance in the test sets, with AUC values of 0.726 and 0.708 in the test set 1 and test set 2, respectively.
Despite having lower sensitivity compared to contrast-enhanced CT, ultrasound remains the most frequently employed examination for screening clinical renal diseases. Texture features can be extracted from ultrasound images. However, there are relatively few studies on the prediction of ccRCC based on ultrasound images. In this study, the AUC values of the training set, testing set 1 and testing set 2 models based on ultrasound images were 0.761, 0.676 and 0.664, respectively.
Over time, the contrast-enhanced CT scan yielded varying information, and it is widely accepted that the ML-based radiomics model utilizes features extracted from full-phase images as the primary indicators for predicting the risk of ccRCC metastasis. In the training set, the AUC values of the model based on cortical phase, unenhanced phase and medullary phase CT images were 0.805, 0.816 and 0.815, respectively; in the test set 1, those of the model based on cortical phase, unenhanced phase and medullary phase CT images were 0.712, 0.783 and 0.806, respectively; and in the test set 2, those of the model based on cortical phase, unenhanced phase and medullary phase CT images were 0.711, 0.686 and 0.749, respectively. Therefore, according to the results of this experiment, the medullary phase images have the best performance in constructing a model for predicting LNM in ccRCC patients based on the images obtained from a single period of enhanced CT scanning.
In addition, this study further integrates three different phases (plain, cortical and medullary) of contrast-enhanced CT images and ultrasound images were further fused to construct a machine learning model for predicting LNM in ccRCC patients. The AUC values for the training set, testing set 1, and testing set 2 were 0.870, 0.836 and 0.804, respectively. The outcomes obtained from this model surpassed those of models solely relying on a single contrast-enhanced CT scan or ultrasound images. More importantly, the results indicated that integrating more patient examination information could enhance the predictive performance of the model.
To explore clinical use, this study further incorporating clinical-radiological characteristics, CT and ultrasound radiomics features to construct a easy-to-use combined model based on multimodal data. Hutterer et al[38]. recruited RCC patients from 12 clinical centers and devloped rosettes that demonstrated a 78.4% accuracy rate in forecasting distant LNM. Capitanio and co-workers[39] also developed a predictive model for LNM in kidney cancer patients, with a high accuracy of 86.9%. Marconi and colleagues[40] developed a prognostic model for anticipating survival rates in distant metastases patients, and the AUC values were 0.68 (95%CI = 0.62–0.74) and 0.73 (95%CI = 0.68–0.78) for the preoperative and postoperative models, respectively. Bai[41] used MRI images to develop an radiomics nomogram for predicting outcomes in patients with distant metastasis, and the AUC value of the external validation cohort was 0.816. Similarly, Zhao and co-workers [42] employed CT images and radiomics to forecast the likelihood of distant metastases in patients, and the AUC in the radiogenomics validation cohort was 0.843. In this study, the AUC values of the training set, testing set 1 and testing set 2 were 0.924, 0.877 and 0.849, respectively. Compared with the radiomics model alone or the clinical radiology model alone in this study and the previous studies, the combined model achieved better prediction performance, indicating that the combined model could potentially lead to improved clinical outcomes when devising treatment plans. Thus, the combined model represents a promising supportive tool for estimating the risk of metastasis in ccRCC patients.
It has been widely accepted that the outcomes of cancer can differ significantly depending on the level of economic development, as well as social and lifestyle factors, across various regions. To minimize the impact of regional factors, this study recruited patients from three distinct regions in two provinces in China, including a coastal city. This represents a minor strength of this study.
Although the results of this study are encouraging, it is important to address its limitations. Firstly, the retrospective design of the study may have introduced selection bias, which could affect the accuracy of the prediction model. Hence, prospective trials are required to validate the findings of this study. Secondly, despite the inclusion of clinical-radiological, ultrasound images, and three-phase images of contrast-enhanced CT scans, some important information such as pathological images was not incorporated in this study. In the future, we plan to integrate these data into our prediction model to enhance its accuracy. Lastly, although our study enrolled patients from three hospitals, the sample size remains relatively small. Further studies with larger sample sizes are warranted to validate the accuracy and reliability of the model.