Clear cell renal cell carcinoma (ccRCC) is the most common type of kidney cancer. While many patients do well after surgery, some - even those at the same cancer stage - have cancer return or spread. Doctors currently struggle to identify which patients face the highest risk.
The standard tools for predicting outcomes - TNM staging (which measures tumor size and spread) and ISUP grading (which assesses how abnormal the tumor cells look under a microscope) - have real limitations. Two patients with identical TNM and ISUP scores can have very different long-term outcomes.
This gap exists because traditional microscope reviews capture only a limited view of the tumor. A pathologist looking through a standard microscope can assess individual cells, but misses the bigger picture of how tumor tissue is arranged and how cancer cells interact with surrounding tissue.
This study asked whether AI analysis of whole slide digital images - scanning the entire tumor specimen at high resolution - could extract hidden patterns that predict which patients will do well and which are at high risk of cancer recurrence.
Pathomics is a field that applies computational and AI methods to analyze the enormous amount of visual information contained in cancer tissue slides. It goes far beyond what a human eye can process - extracting hundreds of measurable features from thousands of cells at once.
Traditional pathology uses standard glass slides viewed through a microscope. Whole slide imaging (WSI) digitizes these slides at extremely high resolution, producing images that capture both individual cell details and the overall tissue architecture simultaneously.
AI algorithms can then scan these digital images and measure features like cell shape, nuclear size, how densely cells are packed, how deeply the tumor invades surrounding tissue - all at a scale and consistency no human observer could achieve.
Studies in other cancers (lung, breast, skin) have already shown that AI-based analysis of tissue slides can predict patient outcomes as well as or better than traditional grading. This study applies the same approach to kidney cancer for the first time at this scale.
The study analyzed 483 whole slide images from ccRCC patients across three separate datasets: 59 patients from the Clinical Proteomic Tumor Analysis Consortium (CPTAC), 146 patients from Shanghai General Hospital, and 278 patients from The Cancer Genome Atlas (TCGA).
Using three completely separate patient populations was a key strength of the study. The AI model was developed using one cohort (CPTAC) and then tested independently in the other two - this tests whether the tool works on patients it has never seen before, which is essential for real-world clinical use.
All tumor tissue was prepared as H&E (hematoxylin and eosin) stained slides - the standard tissue preparation used in pathology labs worldwide. This means the AI approach requires no special equipment beyond what hospitals already have.
Patients were followed for an average of 2-5 years after surgery, tracking disease-free survival (DFS) - the time until cancer returned, spread, the patient died, or their last follow-up visit.
Each whole slide image was broken into smaller tiles and analyzed using QuPath, a specialized digital pathology software. The software automatically detected and measured 43 different features for every individual cell in the slide - measuring things like cell size, nuclear shape, how round or elongated cells were, and how intensely they absorbed the H&E staining dyes.
With 43 potential features, the challenge was identifying which ones actually predict survival. The researchers applied a statistical technique called LASSO (Least Absolute Shrinkage and Selection Operator) - an AI method that tests all features simultaneously and selects only those that genuinely add predictive value while discarding redundant information.
This process identified just five key features out of 43 as the most predictive: nucleus circularity (how round the cell nucleus is), nucleus minimum width (the smallest diameter of the nucleus), and three measures of how intensely the nucleus and cytoplasm absorbed the hematoxylin and eosin stains.
These five features were combined into a single score called the Machine Learning-based Pathomics Signature (MLPS). Each patient receives an MLPS score, and patients are then classified as either high-risk or low-risk based on whether their score falls above or below the median.
In the training cohort, patients classified as high MLPS had dramatically worse outcomes than low MLPS patients. Their risk of cancer recurrence or death was over 15 times higher (hazard ratio 15.05). This is an extremely strong effect - for comparison, TNM staging itself had a hazard ratio of about 7-10 in the same cohort.
When the model was independently tested in the Shanghai General Hospital cohort (146 patients who were never part of training), high-MLPS patients still had over 4 times the risk of worse outcomes. In the large TCGA cohort (278 patients), the risk was 1.65 times higher - smaller but still statistically significant and meaningful.
Importantly, Cox regression analysis - a statistical method that accounts for multiple factors at once - confirmed that MLPS was an independent predictor of survival. This means the MLPS score added predictive value beyond what staging and grading alone could tell doctors.
The differences in hazard ratios between cohorts (15 vs. 4.5 vs. 1.65) likely reflect the different disease severity across the groups. The Shanghai hospital cohort had mostly stage I patients, who generally do well, so distinguishing high and low risk in that group is harder.
The researchers then built an integration nomogram - a visual prediction chart that combines the MLPS score with tumor stage and tumor grade. A nomogram translates multiple risk factors into a single overall risk estimate, and is commonly used in cancer care for treatment planning discussions.
The combined nomogram's accuracy was measured using AUC (area under the curve) values, where 1.0 is perfect prediction and 0.5 is no better than random chance. The AUC values for predicting whether a patient would remain cancer-free were 89.5% at 1 year, 90.0% at 3 years, 88.5% at 5 years, and 85.9% at 10 years.
These accuracy levels are substantially better than what staging or grading alone typically achieves, and competitive with the best existing kidney cancer prediction tools. Predicting outcomes over 10 years with nearly 86% accuracy represents a significant advance in long-range survival estimation.
High accuracy was maintained across all three separate patient cohorts, confirming this is not just a result that looks good in the dataset it was built on - it generalizes to new patients from different hospitals and countries.
The five features selected by the AI - all related to cell nucleus shape and staining intensity - make biological sense. In cancer, the nucleus (the control center of the cell containing DNA) typically becomes more irregular, enlarged, and abnormally structured as the cancer becomes more aggressive.
Nuclear circularity measures how round the cell nucleus is. More irregular, non-circular nuclei tend to indicate more aggressive cancer behavior. Nucleus minimum caliper reflects the smallest width of the nucleus, capturing another dimension of nuclear shape abnormality.
The hematoxylin staining intensity of the nucleus reflects how much DNA is present and how it is organized - which changes with the cell cycle activity and genetic instability common in aggressive cancers. The eosin staining of the cytoplasm (the cell body surrounding the nucleus) reflects protein content and cell metabolism.
Together, these features capture a detailed quantitative picture of nuclear heteromorphism - the degree to which cancer cell nuclei have become abnormal. This is the same biological principle behind the ISUP grading system, but measured far more precisely and consistently than a human pathologist can do by eye.
Current TNM staging focuses on macroscopic features - how large the tumor is and whether it has spread to nearby lymph nodes or distant organs. It does not capture the microscopic biological characteristics of the tumor cells themselves.
ISUP grading does capture microscopic features, but relies on a pathologist's subjective assessment of a limited field of view, and only examines nuclear morphology at the cellular level without considering tissue architecture or how cancer infiltrates surrounding tissue.
WSI-based pathomics addresses both limitations at once: it captures the full tissue architecture across the entire tumor slide (not just a small area), and it quantifies microscopic cellular features objectively and consistently - the same measurement every time, regardless of which pathologist performs the analysis or which hospital the patient is treated at.
For patients with the same TNM and ISUP classification who nevertheless have different outcomes, the MLPS may help explain why - and help doctors offer more personalized follow-up schedules, earlier surveillance imaging, or consideration of additional treatments for high-risk patients.
For kidney cancer patients, this research points toward a future where the pathology report after surgery includes not just a stage and grade, but also an AI-generated risk score that more precisely predicts long-term outcomes.
A high MLPS score would signal to doctors that a patient needs closer monitoring - more frequent imaging scans to catch any recurrence early, or possibly referral to clinical trials exploring additional treatments after surgery (called adjuvant therapy).
A low MLPS score would provide reassurance that a patient is likely in a lower-risk group, potentially avoiding unnecessary treatments and reducing the anxiety and cost of over-surveillance.
Because this tool uses standard H&E tissue staining that is already performed in virtually every pathology lab in the world, it would require no new laboratory equipment or tissue preparation - only software to analyze the digital scan, making it potentially widely accessible if validated in prospective studies.
The study has important limitations. All patient data was collected retrospectively - meaning the researchers analyzed historical records rather than following patients prospectively from the start. Retrospective studies can introduce selection bias and are considered less definitive than prospective trials.
The cut-off value used to divide patients into high and low MLPS groups was set at the median of each individual cohort rather than a single universal threshold. For clinical use, a standardized, fixed cut-off value would need to be established.
Different hospitals use slightly different H&E staining protocols, which can affect how colors appear in the digital slides. This color variation between institutions is a known challenge for AI image analysis tools and was acknowledged as a potential source of error.
The next step is a large, prospective, multi-center study that follows patients from the time of surgery with a pre-defined analysis plan. Only such a study can fully confirm whether MLPS should be incorporated into routine clinical practice for kidney cancer patients.