Machine Learning-Based Prediction of Consistency and Histological Characteristics in Renal Cell Carcinoma Venous Tumor Thrombus Through Volumetric Radiomics

Background. Renal cell carcinoma (RCC) possesses a distinctive inclination to infiltrate the inferior vena cava, resulting in the formation of a venous tumour thrombus (VTT). Accurately assessing the consistency of the VTT prior to surgery is essential for effective treatment strategizing and favourable results. The study aimed to investigate the performance of volumetric radiomic MRI analysis in prediction of consistency and histomorphological vascular patterns of RCC venous tumour thrombus (VTT) with the assistance of machine learning. Methods. Twenty-four RCC patients with VTT underwent nephrectomy and thrombectomy in this study. Preoperatively abdominal DW-MRI was conducted, followed by the creation of 3D model of the thrombus. First-order radiomic features were computed from the complete thrombus volume utilizing ADC maps. The immunohistochemical staining of VTT was performed using CD34, SMA and VEGFR. The machine learning analysis was employed to develop predictive models for VTT histologic features. Patients were grouped based on the thrombus consistency into either solid or friable categories. Results. The solid and friable thrombi were detected in 13 (54.2%) and 11 (45.8%) cases, respectively. Large vessels were predominantly observed in solid VTTs (73.3%; p=0.015). Rich vascularization was a main pattern in solid VTT at 51.5%, contrasting with the friable at 9.1% (p=0.008). There was a strong association between thrombus vessel size and following radiomic features: entropy (r=0.722), skewness (r=0.635), and ADC mean (r=0.610). ADC entropy outperformed in distinguishing between VTT with large and small vessels, achieving the highest performance (AUC 0.930; p<0.001). In distinguishing between VTT with rich and poor vascularization, ADC median showed the best performance (AUC = 0.881; p < 0.001). Using machine learning analysis, we've developed two models predicting crucial histologic traits of VTT with prognostic accuracies of 89% for consistency and 75% for vessel size. Conclusions. Leveraging volumetric radiomic data from MR-DWI, along with machine learning models, we identified unique vascular patterns in VTTs among RCC patients. These models were developed to predict VTT consistency and vessel size using volumetric ADC data from DWI.