Advanced Diffusion MRI Combined With Clinical Factors Shows Promise for Predicting Lymph Node Spread in Endometrial Cancer

Researchers in China have reported that a quantitative MRI technique capable of measuring how water molecules move through tumors at a microscopic level may help identify which patients with endometrial cancer harbor hidden lymph node metastases before surgery ever begins. In a prospective study published in Frontiers in Medicine, a team at The First Affiliated Hospital of Henan Medical University combined whole-tumor histogram features from five advanced diffusion models with routine clinical information and built a prediction model that achieved an area under the curve of 0.924, with 90.5% sensitivity and 82.7% specificity for detecting lymph node metastasis. The work, led by Dongming Han and colleagues, addresses one of the most consequential unanswered questions in gynecologic oncology: whether a tumor’s microscopic architecture, captured noninvasively on a scanner, can reveal its tendency to spread.

The clinical stakes are considerable. Endometrial cancer is the sixth most common malignancy in women worldwide, and its incidence is steadily rising. Most patients, roughly 80% to 90%, have endometrioid adenocarcinoma, which generally carries a favorable prognosis. But non-endometrioid carcinoma, although it accounts for only about 10% of cases, drives more than half of all recurrences and deaths. Lymph node metastasis is one of the strongest prognostic factors in the disease and directly shapes decisions about adjuvant radiotherapy, chemotherapy, or combined treatment. Standard surgical staging involves hysterectomy, removal of both ovaries and fallopian tubes, and systematic pelvic and para-aortic lymphadenectomy, an operation associated with well-recognized complications including vascular injury, lymphocele, and lower-limb lymphedema that can substantially impair quality of life.

To reduce surgical morbidity, many centers have adopted sentinel lymph node biopsy in place of full lymphadenectomy for patients with early-stage, low-risk disease. Yet nodal status uncertainty remains a concern in some patients, and the false-negative rate of sentinel node biopsy alone is approximately 3%. For elderly patients, those with major comorbidities, or those at low preoperative risk, systematic lymphadenectomy may do more harm than good. Conventional MRI, the preferred modality for preoperative staging, relies primarily on size-based criteria for lymph node assessment, most commonly a short-axis diameter greater than 10 millimeters, which cannot reliably distinguish metastatic from inflammatory nodes. The result is a persistent diagnostic gap that the new study set out to close using physics rather than anatomy.

The key innovation lies in how the researchers characterized water diffusion. Conventional diffusion-weighted imaging summarizes water motion with the apparent diffusion coefficient, a mono-exponential model that assumes diffusion follows a Gaussian, or bell-curve, pattern. In highly heterogeneous tumors such as endometrial cancer, however, water diffusion markedly departs from this Gaussian behavior, and the simple ADC model cannot account for that complexity. The team instead acquired diffusion-weighted images with 16 b-values ranging from 0 to 3,600 s/mm² on a 3.0T scanner, a single acquisition from which five non-Gaussian models could be derived simultaneously: intravoxel incoherent motion, diffusion kurtosis imaging, the stretched exponential model, fractional-order calculus, and continuous time random walk. Together these yielded 13 parametric maps capturing microvascular perfusion, diffusion restriction, and spatial heterogeneity.

Each model contributes distinct quantitative parameters. The IVIM framework separates the perfusion fraction contributed by microvascular flow, the pseudo-diffusion of blood within capillaries, and pure water diffusivity. The DKI kurtosis parameter quantifies the degree of microstructural complexity, while its diffusivity term adjusts for non-Gaussian water motion. The stretched exponential model’s exponent measures the departure from Gaussian diffusion, and its distributed diffusion coefficient characterizes the overall rate of non-Gaussian diffusion. The fractional-order calculus model’s β parameter reflects spatial heterogeneity of diffusion, and the continuous time random walk model describes the waiting time and step length distributions of moving water molecules, capturing spatiotemporal heterogeneity. Rather than sampling single slices, the researchers delineated the entire tumor volume slice by slice and extracted 16 histogram features per map, yielding 208 features per patient, an approach designed to reduce the sampling bias inherent in small regions of interest.

The prospective cohort comprised 96 consecutive patients with suspected endometrial cancer enrolled between September 2024 and December 2025, of whom 73 with histologically confirmed disease and known nodal status formed the final analysis: 21 with lymph node metastasis and 52 without. All underwent multi-b-value diffusion MRI within two weeks before surgery, and the metastasis reference standard was postoperative pathological examination after systematic pelvic and para-aortic lymphadenectomy, with isolated tumor cells and micrometastases counted as positive. Two radiologists with 15 and 10 years of pelvic MRI experience, blinded to clinical and pathological data, independently segmented the tumors, and inter-observer agreement for the diffusion parameters ranged from moderate to excellent, with intraclass correlation coefficients between 0.706 and 0.993.

Feature selection proceeded in stages. A univariate Mann-Whitney U test retained 29 features with P values below 0.01, and LASSO regression with 10-fold cross-validation distilled these to three: DKI_K_energy, FROC_μ_10th percentile, and DKI_K_total energy. Patients with nodal metastasis showed markedly higher DKI_K_energy, with a median of 1,800.544 versus 415.358 in node-negative tumors, and correspondingly higher DKI_K_total energy. Collinearity diagnostics revealed that the two DKI energy measures were perfectly correlated, with a Spearman coefficient of 1.000 and variance inflation factors exceeding 135, so the total energy variant was excluded as redundant. Cervical involvement, despite univariate significance, yielded an area under the curve of only 0.618 with a P value of 0.116 and was also dropped. The surviving diffusion star, DKI_K_energy, achieved a univariate area under the curve of 0.805, with 61.9% sensitivity and 92.3% specificity.

Three logistic regression models were then built. The clinical model combined age, pathological type, and preoperative MRI-based local stage, reaching an area under the curve of 0.837. The radiomics model paired DKI_K_energy with the FROC_μ_10th percentile, reaching 0.856. The combined model integrated DKI_K_energy with age and pathological type and achieved the highest apparent performance at 0.924, with 90.5% sensitivity and 82.7% specificity. Notably, DeLong tests found no statistically significant differences among the models, with P values ranging from 0.090 to 0.140, meaning the combined model was not significantly superior in discrimination. The authors are careful on this point: the reported metrics are apparent estimates, since no optimism-corrected internal validation was performed, and the bootstrap analysis was used only to estimate the variability of model intercepts, not to validate performance.

Where the combined model did stand out was in calibration and clinical utility. Its Brier score of 0.092 was lower than the clinical model’s 0.129 and the radiomics model’s 0.120, indicating better agreement between predicted probabilities and actual outcomes. Decision curve analysis showed that across threshold probabilities of 0.1 to 0.3, the combined model consistently delivered higher net benefit than either individual model or the treat-all and treat-none strategies. The authors note this pattern echoes earlier radiomics studies in which combined models improved net benefit without significant gains in area under the curve, suggesting that discrimination metrics alone may not fully capture the value of adding imaging features when the clinical model already performs well.

The biological interpretation is equally nuanced. Higher kurtosis reflects greater microstructural complexity, and energy quantifies gray-level uniformity, with lower energy indicating higher heterogeneity. The elevated DKI_K_energy in node-positive tumors likely reflects dense, uniform high-cellularity rather than necrosis or cysts, a pattern the authors liken to malignant testicular tumors showing higher energy than benign ones, and they emphasize that the meaning of energy depends on tumor type, sequence, and diffusion model. The limitations are candid: the single-center study was not registered, had no a priori sample-size calculation, involved a small and imbalanced sample drawn from patients selected for nodal staging, and lacked motion and distortion correction at high b-values. Conventional ADC histograms were not analyzed, so the superiority of non-Gaussian models over conventional diffusion remains unproven. The authors present the findings as exploratory, requiring independent external validation before clinical use, while arguing that multi-model diffusion imaging combined with clinical indicators may ultimately refine risk stratification and guide individualized treatment decisions in endometrial cancer.

Subject of Research: Non-Gaussian diffusion MRI histogram analysis combined with clinicopathological indicators for assessing lymph node metastasis in endometrial cancer

Article Title: Whole-tumor histogram analysis of non-Gaussian diffusion MRI models combined with clinicopathological indicators for assessing lymph node metastasis in endometrial cancer

Article References: Zang, Y., Wang, H., Ge, W., Yin, H., Xie, L., Qin, R., Liu, W., & Han, D. (2026). Whole-tumor histogram analysis of non-Gaussian diffusion MRI models combined with clinicopathological indicators for assessing lymph node metastasis in endometrial cancer. Frontiers in Medicine, 13, Article 1959334. https://doi.org/10.3389/fmed.2026.1959334

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1959334

Keywords: endometrial cancer, lymph node metastasis, diffusion MRI, diffusion kurtosis imaging, IVIM, fractional-order calculus, continuous time random walk, stretched exponential model, histogram analysis, radiomics, predictive modeling, gynecologic oncology