Renal cell carcinoma (RCC) is the most common malignant kidney tumor, and its incidence is rising globally. The tumor grade, a measure of how abnormal cancer cells appear under the microscope, is one of the most critical pieces of information for deciding how aggressively to treat the disease. Stage I RCC has a 5-year survival rate of 80 to 95%, but Stage III and IV drop to 40% and 10 to 20% respectively, making accurate preoperative grading essential.
CT scanning is the standard non-invasive tool for evaluating kidney cancer, but it has a specific limitation: most existing AI systems for grading RCC analyze only a single CT phase, meaning a single snapshot taken at one moment during a contrast-dye injection. A tumor's appearance can vary meaningfully across the plain scan, arterial phase, venous phase, and delayed phase, and relying on just one of these phases risks missing small lesions or producing a one-sided evaluation that does not reflect the tumor's true biological character.
Clinical variables such as patient age, sex, TNM staging, and tumor size also carry prognostic information that CT images alone cannot provide. This study proposed a deep learning algorithm that combines all four CT phases with structured clinical variable data, creating a richer, more complete representation of each patient's disease that better supports accurate pathological grade prediction before surgery.
The overall framework has three stages: feature extraction from clinical variables, feature extraction from each CT phase, and fusion of all extracted features. For clinical data, an Embedding-MLP module was designed to handle the heterogeneous mix of text-based variables (converted to dense vectors using Word2Vec) and numeric variables (processed by a multi-layer perceptron). The output is a single clinical feature vector that encodes the patient's demographic and medical context in a form the neural network can use.
For each of the four CT phases (plain scan, arterial, venous, and delayed), a separate 3D ResNet-50 model was used to extract spatial features. Unlike 2D networks that analyze individual CT slices, 3D ResNet-50 processes the full volumetric scan, capturing the three-dimensional shape, density, and architectural patterns of the tumor across its entire extent. Residual connections in the network prevent information loss during training. Each phase produces its own feature vector, which is then independently fused with the corresponding clinical feature vector.
The final fusion step uses cross-self-attention, a mathematical mechanism that computes how much each CT phase's features should borrow information from all other phases. For each phase, attention scores are calculated by comparing its query vector against the key vectors of all other phases, and the resulting weighted sum produces an updated feature representation that has absorbed complementary information from all four time points. Eight stacked attention layers perform this fusion progressively, and the final output passes through a classification network to predict whether the tumor is low grade (WHO/ISUP Grade 1 or 2) or high grade (Grade 3 or 4).
The dataset included 1,229 patients with surgically confirmed RCC treated at the Second Xiangya Hospital of Central South University between 2015 and 2024. All patients had complete multi-phase CT data (plain, arterial, venous, and delayed phase), TNM stage I or II disease, and WHO/ISUP pathological grading available. Three CT scanner models were used: Siemens SOMATOM Force, Siemens SOMATOM Flash, and GE Revolution, ensuring the dataset reflected real-world variability in equipment.
All tumor regions were manually segmented by a radiologist with 5 years of CT experience using 3D Slicer software. To ensure accuracy, a senior radiologist with more than 10 years of experience reviewed a random 10% sample of the segmentations, and any that did not pass review were re-done. Importantly, the segmentation radiologist had no access to clinical or pathological information during this process, ensuring the outlines were based purely on visual CT appearance.
The dataset was divided into a training set of 861 patients (70%), a validation set of 124 patients (10%), and a test set of 244 patients (20%). Statistical checks confirmed no significant differences between the three groups in age, sex, tumor size, location, grade, or stage, ensuring a fair comparison. Data augmentation was applied during training, including rotations up to 15 degrees, random scaling between 0.8 and 1.2, horizontal and vertical flips, brightness and contrast adjustments, Gaussian noise, and random patch erasure.
The proposed method was compared against nine alternatives, including traditional radiomics approaches (SVM, logistic regression, Random Forest, MLP) and more recent deep learning methods (RCNN, ResNet34, sECANet, RepVGG-A0, and DC-Net). The proposed multi-phase algorithm achieved the best overall performance with 83.87% accuracy, 72.46% precision, 95.04% recall, and an F1 score of 82.23%. The AUC values for both the ROC curve and the precision-recall curve reached 0.97, the highest among all methods tested.
The most striking result was the recall rate of 95.04%, which exceeded the next-best method by more than 13.5 percentage points. In the context of cancer grading, recall measures how often the model correctly identifies truly high-grade cases. A high recall rate means fewer high-risk patients are missed, which is especially important because high-grade kidney cancer requires more aggressive treatment and closer follow-up. The precision of 72.46% was the lowest among the tested models, indicating a trade-off toward catching all high-grade cases even at the cost of some false positives.
Ablation experiments confirmed that combining all four CT phases with clinical variables was essential. Using only clinical data achieved 82.26% accuracy, and each individual CT phase (plain, arterial, venous, or delayed) alone performed between 75% and 82% accuracy. The integrated model combining all five data streams outperformed every individual component, directly validating the core hypothesis that multi-phase and multi-modal data complement each other in ways that no single source can replicate.
To test whether the model would work on genuinely new patients, the researchers applied it to 20 patients admitted to the same hospital between January and March 2025, after the original training and testing period. These patients met the same inclusion criteria as the training set. Among them, 12 had low-grade and 8 had high-grade tumors, with a mean age of 55 years.
On this prospective validation set, the model achieved 85% accuracy, 77.78% precision, 87.5% recall, and an F1 score of 82.35%. These results are consistent with the test set performance, suggesting the model generalizes effectively to new patients rather than simply memorizing patterns from its training data. The recall of 87.5% in particular confirms that the model continues to reliably flag high-grade cases when encountering new clinical scenarios.
The study acknowledges that this generalization test used only 20 patients, limited by the difficulty of obtaining multi-phase CT data from outside the study institution due to privacy regulations and the logistical complexity of collecting all four CT phases in a standardized format. Nevertheless, the consistency of results between the original test set and the prospective validation set provides encouraging initial evidence of real-world clinical applicability.
This study presented a meaningful advance in automated kidney cancer grading by addressing the core limitation of prior methods: reliance on a single CT phase that captures only a static, one-dimensional view of a dynamic disease process. By combining four imaging phases with structured clinical data and a cross-self-attention fusion mechanism, the model captures complementary information that no single source provides alone.
The most important clinical implication is the model's high recall rate. In oncology, the cost of missing a high-grade tumor, and therefore underestimating the need for aggressive treatment, is far greater than the cost of over-calling a low-grade case. A system that identifies 95% of high-grade tumors preoperatively could help surgeons and oncologists plan more targeted interventions, improving outcomes for the patients most at risk.
Future work should focus on validating this approach across multiple hospitals and scanner types to confirm that it generalizes beyond a single institution. The model could also be extended to include additional clinical variables, tumor molecular markers, or pathology report data as further inputs. As multi-phase CT becomes more routinely available, integrating this type of AI grading tool into clinical decision support systems could become a standard part of preoperative kidney cancer workup.