Before a computer can analyze a tumor in a CT scan, someone must first draw the boundary around it, a process called tumor segmentation. This step is typically performed manually by radiologists, who trace the outline of the tumor layer by layer through hundreds of CT image slices. It is accurate but also extremely time-consuming and dependent on the radiologist's experience and judgment.
The burden of manual segmentation is one of the biggest bottlenecks in developing and deploying radiomics-based AI tools for cancer diagnosis. If accurate automatic segmentation were possible, it would free radiologists from hours of tedious work and make large-scale AI-based cancer analysis far more practical in clinical settings.
This study compared automatic segmentation using a deep learning model (nnU-Net) against traditional manual segmentation by experienced radiologists, specifically for predicting the WHO/ISUP grade of clear cell renal cell carcinoma (ccRCC) before surgery. The key question: does automatic segmentation work well enough that the resulting AI grading model performs comparably to one built on manual labels?
The study included 405 patients with pathologically confirmed ccRCC treated at a single hospital over an 11-year period (2012 to 2023). Patients were divided into a training cohort of 324 patients and a test cohort of 81 patients in a 4:1 ratio. CT imaging from the corticomedullary phase was selected because it provides the clearest contrast between tumor tissue and surrounding kidney, making boundaries easiest to identify.
The nnU-Net is a self-configuring deep learning framework developed at the German Cancer Research Center. Rather than requiring manual tuning of its network design, nnU-Net automatically adjusts its architecture, training strategy, and image preprocessing pipeline based on the specific dataset it is applied to. The model used in this study was pre-trained on the publicly available KiTS19 kidney tumor segmentation challenge dataset and applied directly to the study data without additional fine-tuning.
Each tumor region was analyzed by Pyradiomics software to extract 1,834 radiomic features from each segmented area, whether automatically or manually delineated. Features included first-order statistics, shape descriptors, and a variety of texture measures. LASSO regression was used to select the most predictive subset of features, followed by training of SVM (Support Vector Machine) and KNN (K-Nearest Neighbors) classifiers to predict high-grade versus low-grade tumors.
The nnU-Net achieved a Dice Similarity Coefficient (DSC) of 0.842, meaning the automatically drawn tumor boundaries overlapped with manually drawn boundaries by more than 84% on average. A DSC above 0.80 is generally considered clinically acceptable for medical image segmentation tasks.
The AutoSeg-SVM model (using automatically segmented regions) achieved the highest diagnostic performance with an AUC of 0.865, accuracy of 79%, sensitivity of 85.7%, and specificity of 77.6% in the test cohort. The comparable ManualLabel-SVM model achieved an AUC of 0.817. The DeLong test, which formally compares two AUC values, found no statistically significant difference (P = 0.431) between the two approaches.
Calibration curves and decision curve analysis further confirmed that automatic segmentation produced models with equivalent probability accuracy and clinical utility to those built on manual segmentation. These additional analyses go beyond simple AUC to confirm that the practical clinical value of both approaches is comparable.
One of the most practically important findings was the dramatic difference in time required for each approach. nnU-Net completed tumor segmentation in an average of 1.29 minutes per tumor, while experienced radiologists required an average of 7.07 minutes. This difference of nearly 6 minutes per patient was statistically significant (P less than 0.01).
At a clinical volume of 200 kidney cancer cases per year, this time saving translates to approximately 19.3 hours of radiologist time saved annually for a single hospital. Across a health system or nationally, the cumulative savings in radiologist workload would be enormous, and that time could be redirected to higher-level diagnostic tasks that require human expertise.
An interesting technical finding was that the automatic and manual methods selected entirely different sets of features as the most predictive, yet still produced comparably accurate grading models. This suggests that deep learning-based automatic segmentation may actually capture different, complementary information about tumor heterogeneity, including subtle texture patterns that human eyes cannot reliably detect.
This study provides strong justification for replacing manual tumor segmentation with automated approaches in radiomics-based kidney cancer grading workflows. Since the final grading performance is equivalent, the only relevant difference is the dramatically lower time investment required for automation.
For patients, this means that AI-based tumor grading tools can be deployed more widely without requiring radiologists to spend hours preparing each case. This makes the technology more scalable, particularly for large multicenter studies and real-world deployment in hospitals that do not have specialist radiologists with dedicated time for research-quality manual segmentation.
The authors also note that automatic segmentation is not influenced by operator experience or fatigue, making it more consistent across cases and over time. Human manual segmentation can vary between different radiologists and even between different sessions for the same radiologist. Automation eliminates this source of variability, potentially improving the reliability of downstream AI diagnoses.
This study concludes that nnU-Net automatic segmentation is a valid and practically superior alternative to manual tumor delineation for the purpose of building radiomics-based kidney cancer grading models. With equivalent accuracy, dramatically reduced time requirements, and improved consistency, automation should be the standard approach going forward.
The finding that different features are selected from automatic versus manual segmentation raises an intriguing possibility: combining both approaches might yield even better grading models than either alone. Future research should explore whether ensemble methods that leverage both human and AI-drawn boundaries offer additive diagnostic value.
Broader validation in multicenter cohorts, inclusion of additional CT phases, and extension to other cancer grading tasks will be important next steps for confirming and extending these findings. If validated more widely, automatic segmentation-based radiomics could become standard practice in radiology departments worldwide.