Clear cell renal cell carcinoma (ccRCC) is a common and aggressive form of kidney cancer. Even after surgery that appears successful, 20 to 30% of patients experience recurrence (cancer comes back) or metastasis (cancer spreads to other organs), usually within 3 years of the operation.
Accurately predicting who will and will not experience recurrence is critical for tailoring follow-up care. Patients identified as high-risk might benefit from more intensive surveillance or additional treatment, while low-risk patients could avoid unnecessary interventions.
Current methods rely on clinical staging (how advanced the tumor is) and pathology reports after surgery. However, these tools do not capture the full complexity of tumor biology. Combining information from CT scans and pathology slides could provide a more complete picture -- and that is exactly what this study set out to do.
This retrospective study included 274 patients with confirmed ccRCC who underwent surgery at Jiangmen Central Hospital between January 2008 and March 2021. All patients had preoperative CT scans taken in three phases and had their pathology slides digitally scanned as whole slide images (WSIs).
Patients were split into a training group (164 patients) to build the model and a test group (110 patients) to validate it. The primary outcome was whether the patient experienced local recurrence or metastasis versus no recurrence, with a minimum follow-up time of 3 years after surgery.
Three CT imaging phases were used: precontrast phase (PCP) taken before contrast dye is injected, corticomedullary phase (CMP) taken shortly after injection when the kidney cortex enhances, and nephrogenic phase (NP) taken later when the whole kidney fills with contrast. Each phase reveals different aspects of tumor biology.
The study developed a novel deep learning system called CPNet (CT and Pathology Mutual Guidance Fusion Diagnostic Network). Unlike simpler approaches that just merge features from different data types, CPNet uses a bidirectional attention mechanism to allow the two image types to actively guide each other.
Here is how it works: CT features, while analyzing the CT scan, are simultaneously guided by pathology features to look for CT regions that correspond to microscopic tumor patterns. In return, pathology features are guided by CT features to emphasize tissue regions that relate to what is visible on the CT scan. This back-and-forth creates a richer, more informative combined representation.
Both the CT and pathology image data were processed using ResNet18, a well-established deep learning architecture for image recognition. Transfer learning was applied -- meaning the model started with knowledge learned from analyzing millions of general images before being fine-tuned for cancer images. This approach helps the model generalize better with limited medical data.
Among all models tested, the PCP-Pathology Fuse model (using precontrast CT combined with pathology in CPNet) achieved the highest AUC of 0.8363 and accuracy of 75.45% in the test group. This outperformed models using other CT phases and models using each image type alone.
Compared to models using only a single type of image, the CPNet fusion models were dramatically better: on average, they achieved 15.85% higher AUC than pathology-only models and 15.71% higher AUC than CT-only models. They also had far lower misdiagnosis rates (up to 39.59% lower).
The PCP-Pathology Fuse model also performed comparably to models using all three CT phases combined with pathology (AUC 0.8341), even though it uses only one CT phase. This is significant because using fewer imaging phases means less radiation exposure and lower cost for patients.
It was surprising that precontrast phase (PCP) CT images, taken before contrast dye is injected, outperformed the contrast-enhanced phases. The researchers identified three reasons for this advantage.
First, precontrast CT images capture the full tumor without the distraction of contrast enhancement patterns, giving a more complete view of internal tissue heterogeneity. Second, PCP images are more standardized across imaging equipment and protocols, making the AI model more generalizable. Third, eliminating contrast phases reduces radiation exposure and avoids contrast-related complications for vulnerable patients.
The model's visualization showed it focused on tumor nuclei, cytoplasm, and the surrounding microenvironment in pathology patches, and assigned higher scores to regions with higher malignancy. In CT images, recurrence cases showed the model attending to a broader range of tumor tissue, suggesting these features contain clinically meaningful recurrence signals.
If validated in larger studies, CPNet could be integrated into clinical practice to help surgeons and oncologists stratify patients by recurrence risk before or immediately after surgery. Patients identified as high-risk could be offered adjuvant therapy, closer follow-up, or enrollment in clinical trials.
The model does not require additional testing beyond what patients already undergo. CT scans and pathology slides are standard parts of kidney cancer care. The AI would analyze existing images rather than requiring new, expensive tests -- making it practical to implement.
The researchers emphasize that using only precontrast CT reduces unnecessary radiation and eliminates contrast agent risks, making the tool potentially safer and more equitable for patients with reduced kidney function who cannot tolerate contrast dye.
The study has important limitations to consider. It was conducted at a single hospital, which means the patient population may not represent the diversity of real-world kidney cancer patients globally. Results may differ across different ethnicities, geographic regions, or imaging equipment brands.
The model does not currently incorporate other clinically relevant information such as patient comorbidities, smoking history, or molecular biomarkers like VHL mutation status. Including these additional data sources could further improve predictions.
External validation at multiple institutions using diverse patient populations is essential before CPNet can be considered for clinical integration. The authors recommend a multicenter prospective study incorporating data from different CT scanner brands and hospital settings. Despite these limitations, the PCP-Pathology Fuse CPNet model represents a meaningful advance in non-invasive kidney cancer prognosis prediction.