The study findings indicate that radiomics features, particularly peritumoral and tumoral signatures, serve as more precise indicators for grading and predicting survival. Additionally, the study demonstrated that the RM and nomogram performed better than the CM.

Invasion of tumor cells often disrupts the normal structure of the surrounding parenchymal tissues and leads to alterations in the peritumoral microenvironment. Unfortunately, the significance of this peritumoral microenvironment is sometimes overlooked in research that primarily focuses on the internal aspects of the tumor [20, 24]. Conventional imaging techniques struggle to accurately depict the microenvironment around a tumor. However, peritumoral characteristics offer the potential for quantitatively analyzing the heterogeneity of the microenvironment [25, 26]. In the present study, we constructed RMs for IAT and PAT with the goal of predicting the grading of ccRCC and investigating its association with survival prognosis. Our results revealed that the PAT model (PAT 3 mm and PAT 5 mm) had higher predictive values than that of the IAT model. This finding confirms that incorporating incremental information from both the internal tumor components and peritumoral features into a combined model can significantly enhance predictive performance. Interestingly, our findings align with those of previous reports showing that all peritumoral-feature models yielded grading performance superior to that based on features solely within the tumor [22]. Pathological research has demonstrated the crucial role of the peritumoral microenvironment in evaluating tumor invasiveness [27]. Pathological grading is correlated with tumor invasiveness [28]. Tumors of higher grades show increased invasiveness, leading to the invasion of surrounding tissues. As a consequence, the tumor’s surrounding environment undergoes heterogeneous alterations. Consequently, the evaluation of intratumoral grading heavily relies on peritumoral characteristics, which capture the heterogeneity within the tumor and its surroundings [17]. Our study revealed an association between the histological characteristics of perirenal fat invasion and radiomics features. This finding provides evidence that the radiomics features obtained from the PAT region can reflect the biological behavior of tumors. Therefore, when delineating ROIs, it is important not to overlook areas within the tumor and those outside the tumor, as this is associated with tumor proliferation and heterogeneity. Hence, the PAT model developed in this study automatically extends to the surrounding region based on the delineation of the IAT, not only eliminating the need for time-consuming operations to remove the internal tumor mask but also improving the clinical workflow.

The analysis of feature selection revealed that the first-order texture feature “ exponential_firstorder_Kurtosis,” which was obtained by exponentiating the “firstorder_Kurtosis,” had the strongest correlation with the grade in the peritumoral 5 mm area. The relationship between kurtosis and tumors can be inferred from the morphological characteristics of tumors [29, 30]. A higher kurtosis value indicates greater cell density within the tumor, implying a uniform and densely structured tumor. Kurtosis is frequently employed in radiomics to assess tumor heterogeneity [31]. Higher values of kurtosis may indicate increased tumor heterogeneity, which has a potential correlation with tumor malignancy. Our research supports this relationship, as we have discovered a close association between higher kurtosis and tumor malignancy grade. “Entropy” and “JointEntropy” are two terms that measure the uncertainty or disorder in a random variable or joint probability distribution, respectively. Higher values indicate greater heterogeneity [32]. Previous studies [33, 34] have also highlighted the close association between entropy and tumor invasiveness, which aligns with the findings of our research. From a pathologic perspective, the grades of ccRCC are determined by nuclear diameter, nuclear shape, and nucleoli. A higher grade is characterized by a larger nuclear diameter, a more irregular nuclear shape, and greater irregularity in the arrangement of histological internal components in pathological sections. “Busyness” is an additional crucial feature in the PAT 5 mm model, which measures the intensity change between a pixel and its neighboring pixels. A high value for busyness indicates a “busy” image with rapid changes in CT intensity between a pixel and its neighboring pixels, often signifying the rate of tumor growth or progression. Busyness is linked to tumor prognosis [35, 36]. Our study indicates a positive correlation between this parameter and tumor grade, as well as a close association with perirenal or perirenal sinus fat invasion. To summarize, quantitative analysis of radiomics features is a non-intrusive strategy to better understand the biological properties of tumors. This strategy can provide a substantial foundation for achieving personalized precision treatment of ccRCC.

We then analyzed the clinical risk factors, which have the potential to offer supplemental data and enhance the accuracy of the predictive model [37, 38]. Previous research has prioritized the examination of radiomic characteristics while neglecting clinical risk factors [9,10,11,12,13,14,15,16]. In this investigation, we performed multivariable logistic regression analysis and identified the T stage as an independent predictive factor. This finding aligns with the results of Zheng et al [17] who demonstrated a strong association between T stage and renal cancer grade. The AUC of the nomogram was improved relative to that of the RM in the training set (0.93 vs. 0.80). However, there was no improvement observed in the internal validation (0.78 vs. 0.80) and external validation (0.61 vs. 0.71), indicating that clinical risk factors may not have a strong predictive value for grade. Nevertheless, the nomogram-gained score was able to predict OS, highlighting the importance of both the RM and the nomogram in clinical applications.

Our study had several limitations. Initially, this study was conducted retrospectively and did not encompass chromophobe or papillary tumors and those who will not have nephrectomy, thus potentially introducing selection bias. We collected data on a limited number of cases of chromophobe and papillary RCC (approximately 20–30 cases). The small sample size could potentially impact the reliability of the model and statistical results, affecting the statistical power. In the future, we will expand the sample size and add more cases of other types. To assess the robustness and consistency of our models, we used a dataset from a different institution to evaluate their effectiveness. However, it is important to note that the predictive accuracy of the external validation cohort may be limited by the small sample size and relatively short follow-up period. Consequently, further validation through additional studies incorporating larger sample sizes is a necessary step for future research.

To summarize, the performance of the PAT RM outperformed the IAT model for predicting ccRCC grade before surgery. Additionally, both nomograms and RMs improved predictive accuracy and survival prediction. This advancement can help in adjusting treatment strategies promptly and establishing prognostic risk stratification.