Clear cell renal cell carcinoma (ccRCC) carries the worst prognosis among the three most common RCC histological subtypes, with an estimated 5-year survival rate of 75% after surgery. Despite this sobering outlook, risk stratification within the subtype remains an ongoing challenge, and current clinical tools do not fully capture individual patient risk.
AI-based image analysis of hematoxylin and eosin (H&E) stained histopathological sections has emerged as a low-cost method for predicting genetic mutations and clinical outcomes in oncology. One major limitation has been the requirement to train these models on large quantities of labeled data, which is scarce in medical imaging contexts.
Self-supervised learning addresses this gap by training models on unlabeled data through the creation of pretext tasks. The DINO method (self-distillation with no labels), when combined with a Vision Transformer (ViT) backbone, can identify recurring structures in images without needing any expert annotations, making it well-suited for medical imaging applications.
Vision Transformers (ViTs) differ from conventional convolutional neural networks by considering the position and relationships of objects within an image through an attention mechanism. Multiple attention heads each focus on different image structures, and their outputs are merged to produce a rich feature representation that is highly transparent and interpretable.
The DINO self-supervised method trains the ViT by presenting the model with different transformations of the same image, including cropping and rotation, and tasking it to find consistent underlying features. This forces the model to learn robust representations of recurring structures in histological ccRCC images, even without any outcome labels during training.
Once trained, the model extracts a 384-dimensional feature vector for every 512-pixel patch from each whole-slide image. These vectors capture the histological variety of ccRCC tissue. A Cox proportional hazard model with LASSO regularization is then fitted to these vectors to predict overall survival and disease-specific survival, with patch-level scores averaged to produce a slide-level risk estimate.
The training cohort was 443 ccRCC patients from The Cancer Genome Atlas (TCGA-KIRC), with external validation performed on 266 patients from the University Medical Centre Mannheim who underwent nephrectomy between 2006 and 2011. The median risk score from the training set was used as a fixed threshold to classify patients into low- and high-risk groups.
Significant risk stratification was achieved in the training cohort for both overall survival (OS) and disease-specific survival (DSS), with log-rank test p-values below 0.001 for both endpoints. In the external validation set, the model maintained a significant separation between low- and high-risk groups for OS (p less than 0.005) and DSS (p less than 0.001).
In multivariable analysis including age, metastatic status, tumor size, and grading, the DINO-ViT risk group was an independent predictor of OS in the training set with a hazard ratio of 3.03 (95% CI 2.11 to 4.35). For DSS, the model remained an independent predictor in both the training (HR 4.90; 95% CI 2.78 to 8.64) and the external validation sets (HR 2.31; 95% CI 1.15 to 4.65; p equals 0.02).
A subgroup analysis restricted to patients with metastases showed that the model achieved significant risk stratification for OS and DSS in the TCGA metastatic subgroup, even though the risk threshold was not adapted for this population. This suggests potential utility in high-risk patients for whom tailored follow-up or treatment intensification may be warranted.
Visualization of the six attention heads trained by DINO-ViT revealed that different heads specialized in distinct tissue structures. Heads 1, 2, and 5 focused predominantly on the peritumoural stroma, while heads 3 and 4 highlighted different aspects of the cytoplasm, specifically clear cell morphology versus stained cytoplasm. Head 6 consistently focused on cell nuclei.
This spatial specialization emerged purely from self-supervised training without any prior knowledge of pathological features. The consistency of these learned structures across four different patient slides demonstrates the model's ability to reliably identify the tissue elements that pathologists consider relevant for prognosis, without being explicitly taught to do so.
The alignment between learned attention patterns and established pathological features such as nuclear morphology, cytoplasmic characteristics, and stromal organization suggests that the model captures biologically meaningful variation. Subtle differences in cell size and shape that fall outside the established grading categories may contribute to the additional prognostic information the model provides.
The fact that DINO-ViT remained an independent predictor of DSS after adjusting for grading, tumor size, and metastatic status suggests that it captures information not encoded in current clinical classifications. Known prognostic factors such as WHO/ISUP grade describe discrete categories, whereas the model extracts continuous, fine-grained morphological variation that may reflect underlying biology more precisely.
Certain genetic changes in ccRCC, such as TSC2 or MTOR mutations, manifest as morphologically distinct phenotypes in the cytoplasm and stroma. The model's attention to these compartments suggests it may be implicitly detecting gene expression patterns visible in tissue morphology without direct genomic measurement.
Compared to prior work using pre-defined handcrafted features, the DINO-ViT approach is data-driven and potentially able to identify novel prognostic structures that have not yet been considered clinically relevant. The use of a fixed threshold from the training set for external validation is a methodological strength, as cohort-specific thresholds risk overfitting and inflate apparent generalizability.
If prospectively validated, the DINO-ViT model could serve as a clinical assessment tool to guide therapy selection and follow-up intervals. For example, patients stratified as high-risk by this model might be candidates for adjuvant therapy or more frequent surveillance, while low-risk patients could be managed conservatively.
The retrospective study design is the primary limitation of this work. The TCGA-KIRC cohort had a median follow-up of only 37 months compared to 108 months in the Mannheim validation set, which limits the accuracy of the Cox hazard model trained on the TCGA data. The approximation used for disease-specific survival in TCGA patients is a further source of imprecision.
Larger datasets, particularly with longer follow-up and more complete clinical data, are needed to strengthen the model's robustness. The benefit of adding an AI model over clinical data alone also requires further evaluation, given that established clinical variables such as grading, tumor size, and metastatic status are already strong predictors in the validation cohort.