The renal fibrous capsule is the outermost fibrous covering of the kidney, and its invasion by clear cell renal cell carcinoma (ccRCC) is a recognized adverse pathological feature associated with higher local recurrence risk and upstaging of the tumor. Preoperative knowledge of capsule invasion status could critically influence surgical planning, particularly decisions about resection margins.
Currently, definitive determination of capsule invasion requires pathological examination of the surgical specimen, meaning surgeons must make intraoperative decisions without this information. A reliable preoperative imaging-based prediction tool would allow better preparation for potentially more aggressive resection when invasion is likely.
CT-based radiomics offers a non-invasive approach to extracting quantitative morphological and textural features from tumor imaging that correlate with underlying histopathological characteristics. This study applied radiomics combined with machine learning classifiers to predict renal fibrous capsule invasion preoperatively.
The study enrolled 163 ccRCC patients divided into two independent batches: batch 1 (103 patients) for training and primary validation, and batch 2 (60 patients) for external temporal validation. This two-batch design tests whether the model generalizes across different time periods and clinical contexts.
Tumor segmentation, essential for radiomics feature extraction, was performed using a transfer learning-based deep learning model pretrained on the KiTS19 public kidney tumor segmentation dataset. This semi-automatic pipeline substantially reduces the manual annotation burden while maintaining high segmentation accuracy.
Segmentation performance was validated using the Dice similarity coefficient (DSC), a standard measure of spatial overlap between predicted and ground truth segmentations. The model achieved DSC values of 0.98 and 0.96 for kidney segmentation and 0.94 and 0.86 for tumor segmentation across the two batches, confirming highly reliable automated delineation.
After segmentation, a comprehensive set of radiomics features was extracted from the delineated tumor volumes. Features spanned multiple categories including first-order statistical features (reflecting intensity distribution), shape and morphology features (capturing geometric properties), and texture features from gray-level co-occurrence matrices (GLCM), gray-level size zone matrices (GLSZM), and gray-level dependence matrices (GLDM).
Ten machine learning classifiers were evaluated: AdaBoost, logistic regression, random forest, support vector machine, gradient boosting, decision tree, k-nearest neighbors, Gaussian Naive Bayes, linear discriminant analysis, and ExtraTree classifier. Feature selection was performed to identify the subset of radiomic features most consistently predictive of capsule invasion across both batches.
Nine features were identified as consistently significant across both patient batches, ensuring that selected features were not batch-specific artifacts. These features included flatness, sphericity, and maximum 3D diameter from the shape category, alongside multiple texture descriptors from GLCM, GLSZM, and GLDM.
AdaBoost was the best-performing classifier, achieving AUC of 0.83 in batch 1 and 0.74 in batch 2. The performance drop between batches is typical in radiomics studies due to differences in scanner settings, patient demographics, and imaging protocols, and the maintained discriminative ability in batch 2 confirms meaningful generalizability.
AdaBoost is an ensemble boosting method that combines multiple weak learners (typically decision trees) iteratively, with each subsequent learner focused on correcting the errors of its predecessors. This iterative error-correction mechanism is well-suited to the high-dimensional, often noisy feature space of radiomic data.
The AUC of 0.83 in batch 1 represents strong discriminative ability for a preoperative imaging biomarker. Comparison with other classifiers showed that ensemble methods consistently outperformed single-model approaches, underscoring the benefit of model aggregation for radiomic prediction tasks.
Flatness and sphericity emerged as the most discriminative shape features. Tumors invading the fibrous capsule tend to be less spherical and flatter, reflecting irregular growth patterns as the tumor extends through and distorts the capsular boundary. Maximum 3D diameter correlates with overall tumor burden, which is also associated with invasion propensity.
Texture features from GLCM, GLSZM, and GLDM captured heterogeneity patterns within the tumor. Capsule-invading tumors showed more heterogeneous internal textures, consistent with the known histopathological complexity of aggressively growing ccRCC that has breached its natural containment barrier.
The convergence of morphological and textural features as predictors aligns with the biological expectation that tumors with irregular shape and heterogeneous internal architecture are more likely to have locally invasive behavior, providing biological plausibility to the radiomic findings.
A validated preoperative radiomics model for fibrous capsule invasion prediction could be integrated into surgical planning workflows, flagging patients at high invasion risk for more extensive resection planning and closer intraoperative margin assessment. This could reduce positive margin rates and potentially improve local recurrence outcomes.
The semi-automatic segmentation pipeline using transfer learning from KiTS19 substantially reduces the time burden associated with manual contouring, which is a major barrier to clinical radiomics adoption. Automated segmentation combined with automated feature extraction creates a fully automated prediction pipeline that could be deployed with minimal radiologist effort.
Limitations include the relatively small cohort size, single-institution data for each batch, and the need for prospective validation. Future work should focus on multicenter prospective studies with standardized imaging protocols, integration with other preoperative biomarkers, and development of user-friendly clinical decision support interfaces.