Renal medullary carcinoma (RMC) and clear cell renal cell carcinoma (ccRCC) are two very different kidney cancers that can appear strikingly similar on CT scans. This creates a serious diagnostic challenge because the two diseases require completely different treatments and have very different outlooks.
RMC is an extremely rare and aggressive cancer that almost exclusively affects people with sickle cell trait, a genetic condition affecting red blood cells. It tends to strike younger patients and progresses quickly. ccRCC, by contrast, is the most common form of kidney cancer in the general population and follows a different disease course with different targeted therapies available.
Doctors need to accurately tell these two cancers apart as early as possible. A wrong diagnosis could mean a patient receives the wrong treatment, wasting precious time. This study, conducted at MD Anderson Cancer Center, asked whether artificial intelligence applied to CT scan measurements could reliably distinguish between them.
The researchers collected CT scans from 87 patients with RMC and 93 patients with ccRCC who had all been treated at MD Anderson Cancer Center. Each tumor was carefully outlined on the scan images to create a precise three-dimensional region of interest, and then a technique called radiomics was applied.
Radiomics works by mathematically analyzing the patterns of brightness, texture, and shape within tumor pixels. From each outlined tumor, the researchers extracted 949 distinct radiomic features. These features capture subtle properties invisible to the naked eye, such as how evenly the brightness is distributed, how grainy or smooth the texture appears, or whether certain patterns repeat in particular directions.
The team then trained an XGBoost machine learning model on these features to learn which patterns consistently differ between RMC and ccRCC tumors. XGBoost is a powerful algorithm that builds many small decision rules and combines them into a single strong classifier. Feature selection was performed using LASSO regression to identify the most informative radiomic measurements, reducing complexity and preventing the model from overfitting to noise.
In one analysis the model used only radiomic features. In a second analysis the researchers added a single piece of clinical information: whether the patient had sickle cell trait. This combination test explored whether pairing imaging data with this known RMC risk factor could further improve accuracy.
The radiomics-only XGBoost model achieved an AUC of 0.915, meaning it correctly distinguished RMC from ccRCC in about 91.5% of cases. AUC, or area under the receiver operating characteristic curve, is a standard measure of how well a model separates two groups: a value of 1.0 is perfect and 0.5 is no better than random guessing.
When the researchers added sickle cell trait status to the model, performance jumped to a perfect AUC of 1.0. This means the combined model correctly classified every single patient in the test set. Although sickle cell trait is not always documented, when it is available, including it dramatically sharpens diagnostic accuracy.
The three most important radiomic features the model relied on were RunEntropy, DependenceEntropy, and ZoneEntropy. All three are entropy-based texture measures that capture how randomly or uniformly the pixel intensities are organized within the tumor. RMC tumors tend to have distinctly different textural randomness compared to ccRCC tumors, a difference the model learned to exploit.
Patients presenting with a kidney tumor who also have sickle cell trait should be considered at risk for RMC, and clinicians should communicate this information clearly when ordering imaging and requesting pathology review. This study shows that knowing about sickle cell trait is not just medically relevant but can be the deciding factor in an AI-assisted diagnosis.
For patients without known sickle cell trait, the radiomics model still provides strong support. An AUC of 0.915 is clinically meaningful and could help radiologists prioritize which cases need urgent biopsy or expedited specialty consultation rather than routine follow-up.
One important limitation is that this model was built and tested at a single large academic center. Real-world performance at community hospitals with different CT scanner types or imaging protocols may differ. Validation studies at other institutions are needed before widespread clinical adoption.
This study demonstrates that machine learning applied to CT scan texture measurements can reliably separate two kidney cancers that appear visually similar to radiologists. The approach is entirely non-invasive, requiring only a scan that most patients would receive anyway during diagnosis and staging.
The entropy-based features that drove model performance suggest that RMC tumors have a distinctly different internal tissue organization compared to ccRCC tumors, detectable even before a tissue biopsy is performed. This opens the door to earlier, more confident clinical decision-making.
Future work should focus on validating this model across multiple hospitals, expanding the dataset to capture more rare RMC cases, and exploring whether similar radiomic approaches can track how these tumors respond to treatment over time.