When endometrial cancer spreads to nearby lymph nodes, the prognosis worsens considerably. Surgically removing these nodes (lymphadenectomy) is valuable for accurate staging and guides decisions about additional treatment -- but the operation carries risks of its own, including lymphedema (chronic leg swelling), blood clots, and wound complications. Because only about 10% of endometrial cancer patients actually have lymph node metastasis, routine lymphadenectomy for all patients exposes the majority to unnecessary surgical risk.
The current standard for identifying suspicious lymph nodes on MRI is to flag any node with a short-axis diameter of 10mm or more -- a simple size threshold that achieves only about 48% sensitivity, meaning it misses more than half of metastatic nodes. The ADC value from diffusion-weighted imaging has also been investigated but yields contradictory results across studies. This study took a different approach: combining multiple MRI-derived features with clinical parameters in a machine learning decision tree to build a more accurate, clinically interpretable prediction tool.
The study prospectively enrolled 236 women with confirmed endometrial cancer at Chang Gung Memorial Hospital in Taiwan between July 2010 and July 2018. All patients underwent pelvic MRI at 3 Tesla before surgery, followed by hysterectomy and pelvic lymphadenectomy as standard care. The 236 patients were randomly split 70/30 into training (165 patients) and testing (71 patients) sets. With an average of 27 nodes harvested per patient, the pathology reports provided definitive answers for 5,078 individual lymph node regions.
A key innovation was the computer-aided lymph node segmentation: a region-growing algorithm was implemented that automatically expands outward from a selected starting pixel, including neighboring pixels with similar ADC values until the boundary is reached, providing fully reproducible segmentation. This solved a major reproducibility problem since pelvic lymph nodes are often less than 100 pixels in size on MRI and are extremely difficult to segment manually. From the segmented nodes and tumors, 109 parameters were extracted -- including 96 ADC-related histogram features plus anatomical measurements and 7 standard clinical variables such as tumor grade, tumor size, myometrial invasion depth, and CA-125 level.
The machine learning algorithm identified six parameters that best predicted lymph node metastasis and combined them into a decision tree called the RadSignature. The imaging-based component (RadScore) used four MRI features: (1) mean ADC value of the tumor (cutoff 1.1 x 10-3 mm2/s), (2) skewness of the relative ADC value (cutoff 1.2), (3) short-axis diameter of the lymph node (cutoff 1.7 mm), and (4) skewness of the lymph node ADC value (cutoff 0.072). These were combined with two clinical features: tumor grade (Grade 1-2 vs Grade 3) and tumor size (cutoff 20 mm).
On the independent testing set, the RadSignature achieved 86% sensitivity, 78% specificity, and AUC of 0.85. In direct comparison, the conventional ADC model achieved AUC 0.54 (barely better than random chance) and the standard lymph node short-axis diameter model achieved AUC 0.62. Both comparisons were statistically highly significant (p less than 0.0001). The negative predictive value of 98% on the training set is particularly important clinically: a negative RadSignature result would correctly identify the overwhelming majority of patients who do not have lymph node spread, potentially sparing them from unnecessary lymphadenectomy.
The researchers tested whether the RadSignature was useful across the full spectrum of patients, divided into low-risk, intermediate-risk, and high-risk groups based on tumor stage, grade, and histological type. In the low-risk group (Stage IA, Grade 1-2 endometrioid type), the RadSignature's specificity was significantly higher than both the ADC model (p = 0.0005) and size model (p less than 0.0001), meaning fewer low-risk patients would be unnecessarily referred for lymph node removal. In the high-risk group, the model achieved 100% sensitivity -- catching every metastatic case -- while maintaining superior specificity to the size model (p = 0.0313).
The only cases the RadSignature missed were two instances of microscopic tumor deposits: one measuring just 0.8mm and another measuring 3.3mm. These micrometastases are extremely small and represent the fundamental biological limit of any imaging-based tool, as no MRI technique currently has sufficient resolution to directly visualize deposits of this size. Both false-negative cases had otherwise unremarkable imaging appearances that would challenge any classification system.
A key insight from the study is that while ADC values of metastatic lymph nodes are indeed lower than those of benign nodes on average, this difference is not consistent enough on its own to be diagnostically reliable -- a finding that explains the contradictory results in prior literature. The computer-aided segmentation technique in this study reduces the measurement variability that comes from manual node tracing, which may partly explain why the ADC-related features became more useful when combined in the machine learning model rather than used alone.
From a surgical planning perspective, the decision tree format has practical advantages over a "black box" neural network: each split is a yes/no question that clinicians can follow and verify. A patient with Grade 1 endometrioid cancer smaller than 20mm falls into a low-risk branch that reliably excludes nodal spread -- matching the existing Mayo Clinic criteria. The model can also provide information about which side of the pelvis is more likely to be affected, potentially guiding surgeons to perform targeted node removal rather than complete bilateral dissection. The authors acknowledge that external validation in different hospital systems is needed before broader clinical adoption.