Pancreatic cancer has a 5-year survival rate of only 12%, and approximately 80% of patients already have advanced or locally advanced disease at the time of diagnosis. Lymph node metastasis — when cancer cells have spread to nearby lymph nodes — is one of the most important factors affecting prognosis and treatment options.
If surgeons know before the operation that lymph nodes are involved, they can tailor the surgery, plan more aggressive adjuvant chemotherapy, or in some cases reconsider whether surgery is the right approach. Currently, lymph node status is only definitively known after surgical pathology, which is too late to change the initial treatment plan.
Standard imaging (CT, MRI, ultrasound) can suggest lymph node involvement, but conventional imaging analysis doesn't capture all the subtle texture and intensity patterns that distinguish metastatic from non-metastatic lymph nodes. Radiomics — extracting thousands of quantitative features from images — combined with machine learning offers a more sensitive approach.
The study enrolled 189 surgically confirmed pancreatic cancer patients (151 training, 38 testing), of whom 50 had lymph node metastasis. From ultrasound images of each tumor, 1,561 radiomic features were automatically extracted — including first-order statistics, shape features, texture descriptors, and wavelet-transformed features that capture different spatial frequencies of the image.
Feature reduction was performed in stages: Pearson correlation testing removed redundant features, principal component analysis (PCA) reduced dimensions, and LASSO regression with cross-validation selected the final 15 most predictive non-zero features. These were then used to compute a 'Rad score' for each patient.
Eight different machine learning algorithms were compared for the radiomics model: logistic regression (LR), SVM, K-nearest neighbors (KNN), random forest (RF), extra trees (ET), XGBoost, LightGBM, and multilayer perceptron (MLP). The best-performing model was then combined with clinical features to build the final joint model.
Among eight machine learning algorithms, logistic regression (LR) performed best on the test set with an AUC of 0.850. Tree-based ensemble models (ET, XGBoost) showed overfitting in the test set and were excluded from the final selection despite high training performance.
Two clinical features were identified as significant predictors of lymph node metastasis: tumor boundary characteristics on ultrasound (whether the tumor margin was regular or irregular) and the serum marker CA19-9. These formed a clinical model with an AUC of 0.875.
The combined model — integrating the LR radiomics model with clinical features — achieved the best overall performance: AUC of 0.872 in training and 0.918 in the test set. Decision curve analysis confirmed that the combined model provided greater net clinical benefit compared to either the radiomics-only or clinical-only models across a wide range of threshold probabilities.
The researchers created a nomogram — a visual scoring tool — that allows clinicians to calculate a patient's probability of lymph node metastasis by reading off values for each predictor variable. This transforms a complex machine learning prediction into a simple, interpretable clinical instrument.
Decision curve analysis showed that the combined model had superior net clinical benefit over the 'treat all' and 'treat none' strategies across probability thresholds from about 20% to 85%, the clinically relevant range for this decision.
Ultrasound is widely available, low-cost, and free from radiation exposure, making this approach practically applicable in most clinical settings — not just in well-resourced hospitals with advanced CT or MRI infrastructure.
This study demonstrates that machine learning applied to ultrasound radiomics features can meaningfully predict lymph node metastasis in pancreatic cancer before surgery — information that is currently only available after pathological examination of the resected specimen.
The combined model's AUC of 0.918 on the test set represents strong discriminative performance. Prospective validation in larger, multi-center cohorts would be needed before clinical implementation, but the results are promising.
Future research should explore whether similar radiomics approaches applied to CT or MRI data, or combined with liquid biopsy biomarkers, can further improve pre-surgical staging accuracy and help individualize treatment planning for pancreatic cancer patients.