Inside every kidney tumor, specific genes and molecular regulators are either more active or less active than in healthy tissue. These molecular differences influence how aggressively the cancer behaves, whether it will spread, and how it responds to treatment. Normally, measuring gene activity requires a biopsy and laboratory analysis. But what if CT scan images already contain hidden clues about a tumor's molecular makeup?
This study focused on a small RNA molecule called miR-15a, a type of microRNA that acts as a molecular regulator suppressing the activity of cancer-driving genes. When miR-15a levels are low in a kidney tumor, the tumor tends to behave more aggressively. Previous research had linked miR-15a to tumor size, growth patterns, and patient outcomes. However, measuring miR-15a still required invasive tissue sampling.
The researchers asked whether CT scan features, which can be obtained non-invasively, could predict the level of miR-15a expression in a kidney tumor. This type of approach, linking imaging findings to underlying biology, is called radiogenomics. A reliable radiogenomic link would mean that a standard CT scan could provide information about tumor biology without the need for a separate biopsy.
Sixty-four patients with confirmed renal cell carcinoma were included in the study. For each patient, CT scans were analyzed to extract both conventional radiological features, such as tumor size, enhancement pattern, and the presence of necrosis or calcification, and quantitative radiomic features capturing texture, shape, and intensity distributions.
Tissue samples from the same patients were also analyzed to measure miR-15a expression levels directly. Patients were divided into high and low miR-15a expression groups based on whether their measured values fell above or below the median. This binary classification was the outcome the machine learning models were trained to predict from the imaging features alone.
A Random Forest classifier was trained on the combined set of conventional and radiomic imaging features to distinguish high-expression from low-expression miR-15a tumors. Random Forest builds many independent decision trees and combines their predictions, which makes it robust against overfitting and able to handle the many correlated features typical in radiomics studies. Model performance was assessed using the area under the ROC curve (AUC).
Hierarchical clustering was also applied to group patients based on their imaging feature profiles and examine whether the clusters aligned with known differences in tumor aggressiveness. This unsupervised approach does not use the miR-15a labels during grouping, making it a stronger test of whether imaging features genuinely reflect meaningful biological differences.
The single strongest predictor of miR-15a expression was tumor size, which by itself explained 82.81% of the variation in miR-15a levels, expressed as an adjusted R-squared of 0.8281. This remarkably high explanatory power from a single measurement suggests a strong biological link: as tumors grow larger, miR-15a activity tends to fall, potentially reflecting how the cancer progressively disrupts normal molecular regulation.
The full Random Forest classifier achieved an AUC of 1.0, meaning it perfectly classified every patient in the test set as high or low miR-15a expression. While a perfect AUC on a small dataset of 64 patients should be interpreted cautiously and validated in a larger cohort, it demonstrates that the combination of conventional imaging features, especially tumor size, with radiomic texture measures provides very strong discriminatory power.
Tumors with low miR-15a expression were significantly more likely to show necrosis (areas of dead tissue inside the tumor) and nodular enhancement patterns (irregular contrast uptake) on CT. These are visual features that experienced radiologists already note clinically and that are associated with more aggressive tumor behavior, strengthening the biological plausibility of the radiogenomic link.
Hierarchical clustering of patients based on their CT imaging features, performed without any knowledge of miR-15a levels, produced two distinct patient groups. When the miR-15a measurements were mapped onto these clusters after the fact, the groups aligned closely with high and low miR-15a expression, confirming that imaging features capture genuine biological signal rather than statistical noise.
The cluster associated with low miR-15a expression contained tumors with more aggressive radiological characteristics: larger size, more necrosis, and more irregular enhancement. The cluster associated with high miR-15a expression, meaning more active tumor suppression by this microRNA, contained tumors that appeared more homogeneous and less aggressive on imaging. This biological consistency across two independent analytical approaches strengthens confidence in the findings.
These results connect to clinical outcomes because low miR-15a is associated with higher-grade tumors that have worse prognosis. If imaging-based clustering can reliably separate aggressive from indolent tumors without molecular testing, it could support clinical decision-making for patients who present without the opportunity for biopsy, such as those who are not surgical candidates.
For patients, the most direct implication of this research is the possibility that a standard CT scan, which they would receive anyway as part of their kidney cancer evaluation, could provide information previously only obtainable through additional invasive testing. This is especially relevant for patients who are older, have significant other health conditions, or whose tumor location makes biopsy technically difficult.
Clinicians currently use tumor size and CT appearance to guide treatment decisions, but molecular information like miR-15a expression could refine these decisions further. For example, a large tumor with low predicted miR-15a expression might warrant more aggressive surgical planning or earlier systemic treatment consideration compared to a same-sized tumor with high predicted miR-15a expression.
The limitation of 64 patients is the most important caveat. Small datasets can produce models that appear highly accurate but fail to generalize. Independent validation in a larger, multicenter patient cohort is essential. Future work should also explore whether miR-15a expression prediction adds prognostic value over and above what tumor size and grade already provide to ensure the additional molecular information is truly actionable.