When a patient with kidney cancer has surgery, two types of data are generated: the CT scan (taken before surgery, showing the whole tumor in 3D) and the pathology slide (a microscopic view of a small slice of removed tissue). These two data types are complementary - the CT shows the overall tumor landscape, while the pathology reveals microscopic details like tumor grade and cell architecture.
Researchers have long wanted to connect these two data types computationally - to learn what microscopic tumor features correspond to specific CT appearances. This is called radiology-pathology correlation. The challenge is that CT images and pathology slides are at very different scales and from different perspectives, making direct spatial alignment extremely difficult and often impractical.
This study introduces CorrFABR (Correlation-based Foundation Alignment and Bridging Representation), a novel method that learns radiology-pathology correspondences without requiring spatial registration - solving a major practical barrier in this field.
CorrFABR builds on two powerful AI foundation models - large pre-trained networks that have learned rich representations from massive datasets. It uses DinoV2 (trained on natural images) for processing CT scan features and CONCH (a pathology-specific foundation model) for processing histology slide features.
The key innovation is that CorrFABR does not try to align CT and pathology spatially (matching exactly which part of the CT corresponds to which part of the slide). Instead, it learns statistical correlations between CT-derived features and pathology-derived features at the patient level, capturing the relationship without requiring precise spatial matching.
This registration-free approach is important because in real clinical practice, pathology sampling is random - the pathologist selects representative areas, not necessarily areas that correspond to a specific CT region. CorrFABR's approach respects this reality while still extracting meaningful correlations between the two data types.
CorrFABR was applied to the task of classifying ccRCC aggressiveness - distinguishing tumors that are likely to behave aggressively and spread versus those that are likely to remain indolent (slow-growing and low-risk). This is a clinically important distinction because it guides how intensively a patient should be monitored and treated.
The model achieved a ROC-AUC of 0.855 for this classification task. This means that in about 85.5% of cases where you randomly picked one aggressive and one indolent tumor, the model correctly identified which was which. This level of performance is competitive with methods that require spatial alignment and is achieved using only preoperative CT data.
The foundation model approach proved its value here: by leveraging representations learned from massive datasets, CorrFABR extracted informative features even from relatively small kidney cancer datasets, avoiding the overfitting problems that plague many specialized medical AI models trained from scratch.
The most compelling clinical use case for CorrFABR is preoperative risk stratification. If a patient's CT scan can reliably predict whether their tumor is aggressive or indolent before surgery, this information can directly influence the surgical plan and post-operative management.
For patients with small kidney tumors, there is often uncertainty about whether to operate, observe ("active surveillance"), or ablate. A tool that predicts aggressiveness from the CT scan - without requiring biopsy - could help resolve this uncertainty, enabling more personalized decisions about treatment timing and intensity.
CorrFABR represents a broader trend in medical AI: using large pre-trained foundation models as starting points rather than building specialized models from scratch. This approach is particularly valuable in medicine, where labeled training data is scarce and expensive to generate.
The registration-free design also makes CorrFABR more practically deployable than methods requiring precise CT-pathology alignment, which often demands significant manual effort from radiologists and pathologists. By removing this burden, the approach is more compatible with real-world clinical workflows.
Future work will need to validate CorrFABR in larger, multi-institutional datasets and explore whether the radiology-pathology correlations it learns can provide biologically interpretable insights into tumor biology, not just classification labels.