Lymphovascular space invasion (LVSI) refers to cancer cells that have infiltrated into blood vessels or lymphatic channels within or around the tumor. This is one of the most critical prognostic factors in endometrial cancer because it indicates the tumor has already taken the first step toward spreading through the body's circulatory highway - the lymphatic system and bloodstream.
LVSI status profoundly affects surgical planning. Patients with LVSI have a substantially higher risk of lymph node spread and typically need lymph node dissection (lymphadenectomy) during surgery. However, this procedure carries its own risks, including lower limb lymphedema - a painful chronic swelling that can significantly reduce quality of life. For patients without LVSI, guidelines suggest lymphadenectomy can often be safely omitted.
The problem is that LVSI can only be definitively confirmed by the pathologist after examining the surgically removed tumor. Intraoperative frozen sections (rapid mid-surgery tissue analysis) have high error rates for LVSI. A reliable preoperative MRI-based prediction tool could guide surgical planning before the first incision is made.
892 endometrial cancer patients from three hospitals were enrolled. Preoperative MRI included three sequences for each patient: T2-weighted imaging (T2WI), contrast-enhanced T1-weighted imaging (CET1WI), and apparent diffusion coefficient (ADC) maps derived from diffusion-weighted imaging. This multisequence approach captures different aspects of tumor biology - T2WI shows anatomy, CET1WI reveals vascularity, and ADC reflects cellular density.
Two types of features were extracted. Handcrafted radiomic features (RFs) - mathematical measurements of texture, shape, and intensity patterns - were extracted using PyRadiomics software following standardized IBSI guidelines, yielding 3,933 features per 2D analysis and 5,064 per 3D analysis. Deep learning features (DFs) were extracted from pretrained neural networks (VGG11, ResNet101, and DenseNet121) by passing 2D tumor images through the networks and capturing the high-level feature representations from the last convolutional layer.
Feature selection used a rigorous 4-step process: univariate screening, Pearson correlation filtering (removing redundant features), LASSO regularization, and backward stepwise elimination guided by the Akaike information criterion. This process yielded compact, non-redundant feature sets of 3-9 features per model. Six final models were built and compared, including a combined model integrating both radiomic and deep learning features.
In the training cohort, the 2D radiomic model achieved AUC of 0.775 and the 3D radiomic model achieved 0.772 - essentially identical performance, suggesting the simpler 2D approach is adequate. Among deep learning models, DenseNet121 performed best with AUC 0.757, outperforming ResNet101 (0.671) and VGG11 (0.720).
The combined model integrating 2D radiomic features with DenseNet121 deep learning features achieved AUC 0.796 in training, 0.751 in internal validation, and 0.765 in external validation. This consistent performance across three cohorts from different hospitals is the study's key strength - it demonstrates the model does not simply memorize the training data but generalizes to new patients.
Decision curve analysis showed the combined model provided clinically meaningful net benefit over a threshold probability range of 0.20-0.50 - precisely the range where clinical decisions about lymphadenectomy are made. Calibration was excellent across all cohorts (Hosmer-Lemeshow test, all P greater than 0.05), confirming the model's predicted probabilities are reliable rather than systematically biased.
DenseNet121 (Dense Convolutional Network) has a distinctive architecture: instead of passing information only forward layer by layer like conventional networks, it creates dense connections where each layer receives inputs from all previous layers. This promotes feature reuse across the network and enables efficient gradient flow during training, which helps prevent the "vanishing gradient" problem that causes very deep networks to stop learning effectively.
This dense connectivity means DenseNet can extract both low-level features (edges, textures) and high-level semantic features (complex tumor patterns) simultaneously, without losing information as it flows through the network. For a complex medical imaging task like LVSI prediction - where the relevant signal may be subtle - this feature preservation is particularly valuable.
The finding that 2D segmentation performed as well as 3D is practically important. 3D tumor contouring is highly labor-intensive, requiring a radiologist to outline the tumor on every imaging slice. If 2D segmentation of the single largest tumor slice provides equivalent performance, this dramatically reduces the workload needed to apply the model in clinical practice, making real-world deployment more feasible.
An LVSI prediction tool with AUC near 0.77 would not replace pathology, but it could meaningfully guide preoperative decision-making. Surgeons could use the model's probability output to plan surgical scope: patients with high predicted LVSI probability could be scheduled for more extensive lymph node sampling, while those with low predicted probability could potentially proceed with less extensive surgery.
The clinical benefit extends beyond individual patients to health systems. Unnecessary lymphadenectomy is costly and causes morbidity. A validated preoperative LVSI tool could reduce the rate of unnecessary lymph node dissections performed on LVSI-negative patients who were incorrectly assumed to be at high risk. Conversely, it might increase lymphadenectomy rates in patients who appear clinically low-risk but have imaging features suggesting occult LVSI.
Limitations include the retrospective design, manual tumor segmentation (which introduces variability and requires radiologist time), and the use of CNN architectures pretrained on natural images (ImageNet) rather than medical images. Future work should prioritize automated segmentation algorithms and domain-specific pretraining to further improve performance and reduce implementation barriers.
This study stands out for its rigorous multicenter design - using data from three independent hospitals in different cities - and for demonstrating consistent performance across training, internal validation, and external validation cohorts. Many radiomics studies report excellent training-set performance that fails to replicate externally; this study shows that the combined radiomic-deep learning approach is genuinely generalizable.
The convergence of findings - that combining handcrafted and deep learning features outperforms either alone, that DenseNet121 is the optimal CNN architecture for this task, and that 2D analysis is sufficient - provides a practical blueprint for others building similar tools. These are replicable, testable conclusions rather than just a black-box prediction score.
The authors call for prospective multicenter studies that would constitute true clinical validation. The ideal next step would enroll patients prospectively at the time of diagnosis, apply the model blinded to the treating team, then compare surgical outcomes between guided and unguided decision-making groups. That level of evidence would be needed before the model could be formally recommended in clinical guidelines.