This study developed and tested an AI system called TR-Net (Transformer-ResNet) that analyzes CT scans to determine the stage of clear cell renal cell carcinoma (ccRCC), the most common type of kidney cancer. Staging means determining how advanced the cancer is, which directly determines what treatments are appropriate.
The AI was trained on data from 1,148 patients across five medical centers in China, making it one of the largest multi-institutional studies of AI-based kidney cancer staging to date. Testing across multiple centers is critical because AI trained at one hospital may fail at another due to differences in scanning equipment, image settings, and patient populations.
Published in BMC Cancer in 2025, this research also investigated whether AI could help radiologists, the doctors who interpret medical scans, perform better in their staging decisions. The study showed that both junior and senior radiologists improved their accuracy when they reviewed AI results alongside the scan.
Clear cell RCC accounts for roughly 75-80% of all kidney cancer cases. When a patient is diagnosed, one of the first and most important questions is: how far has the cancer spread? The answer, expressed as a TNM stage, guides nearly every treatment decision.
T stage describes the size and local extent of the primary tumor. N stage indicates whether nearby lymph nodes are involved. M stage describes whether the cancer has spread to distant organs. Together, these form the overall stage from I (early, localized) to IV (advanced, spread to distant sites). Getting this classification right matters enormously because treatments ranging from active surveillance to surgery to systemic therapy are all stage-dependent.
Determining stage from CT scans requires expert interpretation and can be difficult, especially when tumors are close to important structures like the renal vein or surrounding fat. Variability between radiologists, particularly between those with more and less experience, is a real clinical problem. AI could provide a consistent, objective staging assessment to complement expert human judgment.
TR-Net combines two powerful neural network designs. The ResNet component (Residual Network) is a well-established architecture for image analysis that uses shortcut connections to help deep networks train effectively without losing information. The Transformer component adds an attention mechanism that allows the model to focus on important regions of the 3D CT volume regardless of where they are spatially located.
The system processes CT images in three dimensions, treating the scan as a volumetric dataset rather than a stack of independent slices. This is important because the extent of kidney cancer, particularly whether it has broken through the kidney's outer fibrous capsule or invaded the renal vein, is best assessed by examining the tumor in 3D context.
TR-Net was trained to classify tumors according to both T stage alone and the combined TNM stage. Training on both tasks simultaneously allows the model to learn shared features that are relevant across different aspects of staging, potentially improving generalization compared to training separate models for each task.
For T stage classification, TR-Net achieved a micro-AUC of 0.939 to 0.954 on external validation datasets, with accuracy ranging from 0.843 to 0.869. These results indicate that the model is highly capable of distinguishing between T1, T2, T3, and T4 tumors based on CT appearance alone.
For the more complex task of TNM staging, the model achieved a micro-AUC of 0.924 to 0.935 and accuracy of 0.807 to 0.856. The slightly lower performance on TNM compared to T stage alone is expected, since incorporating nodal and metastatic assessments adds complexity, and lymph node involvement in particular can be challenging to determine from CT imaging.
Performance was consistent across the five hospital sites, which used different CT scanning equipment and protocols. This cross-site consistency is one of the strongest indicators that the model has learned genuinely generalizable features rather than artifacts specific to one institution's imaging conditions.
One of the most clinically important findings of this study was that TR-Net improved the staging accuracy of radiologists when they used it as a decision-support tool. Junior radiologists improved their accuracy by 5.96 percentage points when reviewing TR-Net's staging suggestion alongside the CT scan. Senior radiologists improved by 4.20 percentage points.
The fact that even experienced radiologists improved suggests that AI is not simply providing obvious guidance that any expert would already know. Instead, the model appears to be capturing nuanced features that are difficult to quantify visually, helping specialists make more consistent and accurate calls in borderline or difficult cases.
This type of human-machine collaboration, where AI serves as an intelligent second opinion rather than replacing human judgment, represents the most promising near-term model for clinical deployment. It preserves the radiologist's expertise while reducing the chance that subtle staging features are missed.
A critical concern with AI in medical settings is interpretability: can doctors understand why the AI reached a particular conclusion? TR-Net addresses this using a technique called Grad-CAM (Gradient-weighted Class Activation Mapping), which generates heatmaps highlighting the regions of the CT scan that most influenced the model's staging decision.
These heatmaps allow radiologists to see whether the AI is focusing on clinically relevant features, such as the tumor boundary or vascular involvement, or whether it is picking up on irrelevant artifacts. This transparency is essential for building appropriate trust in the system: radiologists should neither blindly accept nor reflexively reject the AI's output.
Grad-CAM visualizations showed that TR-Net consistently attended to anatomically meaningful regions when making its staging decisions, supporting the interpretation that the model has learned biologically relevant features rather than spurious correlations in the training data.
For patients with clear cell kidney cancer, accurate staging at the time of diagnosis is life-changing. An incorrect stage can lead to under-treatment, leaving aggressive cancer inadequately addressed, or over-treatment, exposing patients to the risks and side effects of therapies they do not need. AI tools like TR-Net could reduce these staging errors.
The practical model proposed by this study, where AI generates a staging suggestion that radiologists then review and confirm, is realistic and achievable with current technology. It does not require replacing radiologists or overhauling hospital workflows. Instead, TR-Net could be integrated as an additional layer in the standard CT reporting process.
Limitations include the retrospective study design and the restriction to patients from China, which may limit generalizability to other ethnic groups and healthcare systems. Prospective trials across more diverse populations will be important before TR-Net is recommended for routine clinical use globally.