Radiogenomics is an emerging field that combines two types of information doctors already collect: imaging data from MRI or CT scans, and molecular data from tumor tissue samples. The core idea is that the visible appearance of a tumor on a scan - its shape, texture, and internal structure - reflects the underlying biology of that tumor. If this connection is strong enough, radiomics (the mathematical analysis of medical images) could reveal information about tumor genetics without requiring additional biopsies or molecular tests.
For endometrial cancer, accurate risk stratification before surgery is critical. Women with low-risk disease may be managed with minimally invasive surgery, while high-risk patients need more aggressive treatment. Currently, preoperative risk assessment relies on endometrial biopsy and imaging to assess how deeply the tumor has grown into the uterine wall. However, a significant proportion of cases are misclassified, leading to either under-treatment or over-treatment.
A 2021 study published in Communications Biology (Nature Publishing Group) analyzed 866 endometrial cancer patients from a single Norwegian cancer center, integrating preoperative MRI with gene expression data to develop a radiogenomics approach. The study's key outputs were: (1) three radiomic risk clusters that predict survival, and (2) an 11-gene signature derived from those clusters that can be measured in tumor tissue to identify high-risk disease - validated in four separate patient cohorts.
The study enrolled 866 endometrial cancer patients treated at Haukeland University Hospital in Norway. From these, 487 patients had preoperative pelvic MRI available, and 554 had transcriptome (gene expression) data from resected tumors. A subset of 51 patients had both MRI and gene expression data, forming the key discovery set where imaging patterns could be directly linked to molecular patterns.
For the MRI analysis, two approaches to tumor outlining (segmentation) were compared. In the training cohort (138 patients), experienced radiologists manually outlined each tumor in three dimensions on MRI. In the validation cohort (336 patients), automated machine learning-based segmentation was used, running without any human involvement. From each tumor outline, 53 mathematical features were extracted covering volume, surface geometry, signal intensity statistics, and texture patterns - including how homogeneous the tumor appears internally and how irregular its shape is.
Gene expression data came from three platforms: L1000 (a high-throughput gene expression assay), Agilent microarray, and RNA sequencing data from TCGA. This multiplicity of platforms is important because it allowed the researchers to validate their gene signature across independent cohorts using completely different measurement technologies - a much more rigorous test than validating with the same technology in a different set of patients from the same center.
Unsupervised clustering of the 53 radiomic features in the manually segmented training cohort divided patients into three groups. Cluster 1 (70 patients) showed generally favorable tumor characteristics. Cluster 2a (44 patients) and Cluster 2b (24 patients) showed progressively more aggressive features. Compared to cluster 1, patients in clusters 2a and 2b were significantly more likely to have deep tumor invasion into the uterine muscle wall (77.8% vs. 24.3%, p less than 0.001), lymph node spread (85.7% vs. 14.3%, p = 0.020), high-grade histology, non-endometrioid tumor type, and advanced FIGO staging.
Survival analysis confirmed these clinical differences translate into meaningfully different outcomes. Disease-specific survival (survival from cancer specifically, not other causes) was significantly worse for patients in clusters 2a and 2b compared to cluster 1 (p less than 0.001). Pairwise comparisons showed that cluster 2b patients had the worst outcomes, cluster 2a patients had intermediate outcomes, and cluster 1 patients fared best. The same three-cluster separation in survival was reproduced in the larger automated segmentation validation cohort (336 patients, p less than 0.0001), confirming that machine learning can replace manual expert segmentation for this analysis.
Among the 53 radiomic features, no single feature alone could discriminate between all three clusters - the predictive information was distributed across multiple interacting features. However, normsurfvolratio (a measure of how irregular the tumor's surface shape is, independent of overall tumor size) was ranked as the most important discriminating feature between the best-prognosis and worst-prognosis clusters, suggesting that irregularly shaped tumors are a particularly important predictor of high-risk disease.
In 51 patients with both MRI and gene expression data available, the researchers identified genes that were differentially expressed between the best-prognosis radiomic cluster and the worst-prognosis cluster. From this analysis, 11 genes were selected to form a signature: 3 upregulated genes in the most aggressive tumors (HSPA5, GATA3, HSP90AA1) and 8 downregulated genes (SCGB2A1, GSTK1, MMP7, GDF15, ANXA1, SAT1, CNDP2, and PBX1). Each patient's signature score was calculated as the sum of the upregulated gene expression values minus the sum of the downregulated gene expression values.
The two upregulated heat shock proteins (HSPA5 and HSP90AA1) are molecular chaperones that help cells survive stress - and in cancer, elevated heat shock protein levels are associated with activation of the PI3K/AKT signaling pathway, the most commonly altered pathway in endometrial cancer. High HSPA5 and HSP90AA1 expression has previously been reported in endometrial cancer by immunohistochemistry, validating the biological plausibility of their inclusion. Among the downregulated genes, reduced SCGB2A1 (a protein called secretoglobin family 2A member 1, sometimes called mammaglobin-B) has specifically been linked to poor survival in endometrial cancer in prior work.
Patients with high signature scores were strongly associated with loss of estrogen receptor (ER), progesterone receptor (PR), and androgen receptor (AR) protein expression - all well-established markers of aggressive endometrial cancer biology. They were also predominantly classified as the TCGA copy-number-high (p53 abnormal) subtype, which has the worst prognosis in the TCGA molecular classification system. Conversely, patients with low signature scores were more often POLE-mutated - the best-prognosis TCGA subtype.
The 11-gene signature was validated for its prognostic value in four independent datasets. In the full L1000 dataset (392 patients), high signature score predicted significantly worse disease-specific survival (p less than 0.0001). In the Agilent microarray dataset (256 patients), the same pattern was reproduced (p less than 0.0001). In the external TCGA RNA sequencing dataset (298 patients), high signature score again predicted poor survival (p = 0.0019). In each case, the signature retained its prognostic impact after accounting for standard clinical variables including tumor grade, stage, and histologic type.
Crucially, the signature also showed prognostic value in a subset of 296 patients who had been classified as low-risk based on preoperative biopsy (endometrioid tumors, grade 1 or 2). This subgroup is clinically important because these are exactly the patients who might be expected to do well and are sometimes undertreated. The finding that the gene signature identified a subset of these apparently low-risk patients with poor prognosis (p = 0.00012) suggests it could help avoid under-treatment in patients who look favorable by standard criteria but have aggressive underlying biology.
The automated ML segmentation cohort also partially validated the gene signature link: in 98 patients with overlapping MRI and gene expression data, there was a trend (though not statistically significant) for patients in the low-risk radiomic cluster to have lower gene signature scores, consistent with the finding in the manually segmented discovery set. The somewhat weaker result may reflect the larger group sizes and slightly different patient mix in the validation cohort.
This study demonstrates a working radiogenomics pipeline for endometrial cancer: preoperative MRI features can cluster patients into groups with distinct biological characteristics and survival outcomes, and those imaging clusters are linked to a measurable gene expression signature that validates independently across multiple platforms and patient populations. The fact that automated machine learning segmentation reproduces the same clusters as expert manual segmentation is particularly important for practical deployment - it means the imaging analysis does not require scarce radiology expert time for each patient.
The most clinically impactful finding may be the ability of the 11-gene signature to identify high-risk patients within the low-risk histological category. Current standard of care uses preoperative biopsy showing low-grade endometrioid cancer as a reason to recommend less extensive surgery (avoiding lymph node removal and limiting surgical extent). If the gene signature can identify the subset of these patients with aggressive underlying biology, it could prompt more thorough surgical staging and potentially adjuvant therapy for a group currently being undertreated.
Important limitations include the retrospective single-center design, the use of two different MRI field strengths (1.5 T and 3 T scanners) requiring separate normalization, and the relatively small size of the discovery set (51 patients with both MRI and gene expression data). The gene expression datasets also had some patient overlaps. The authors conclude that prospective validation in larger independent patient cohorts is needed before this radiogenomics approach can be incorporated into clinical decision-making. The code and model weights are publicly available for other researchers to build on.