Deep learning radiomics model for lymph node metastases prediction: a diagnostic study.
Authors
Affiliations (5)
Affiliations (5)
- Division of Pancreatic Surgery, Department of General Surgery, West China Hospital of Sichuan University, Chengdu, China.
- Department of Hepatopancreatobiliary Minimal Invasive Surgery, Chengdu Shang Jin Nan Fu Hospital, Chengdu, China.
- West China School of Medicine, West China Hospital of Sichuan University, Chengdu, China.
- Biomedical Big Data Center, West China Hospital of Sichuan University, Chengdu, China.
- Department of Radiology, Chengdu Shang Jin Nan Fu Hospital, Sichuan University, Chengdu, China.
Abstract
Pancreatic ductal adenocarcinoma (PDAC) frequently exhibits lymph node (LN) metastasis, making early identification of high-risk patients crucial for optimizing treatment strategies. However, robust predictive models and external validation remain limited. This study aimed to develop and validate an artificial intelligence (AI) model for preoperative detection of LN status in PDAC patients. We retrospectively enrolled 233 pathologically confirmed PDAC patients from two high-volume pancreatic surgery centers between July 2015 and July 2022. All patients underwent contrast-enhanced computed tomography (CT) scans and met strict inclusion and exclusion criteria. Two experienced radiologists independently reviewed imaging. Radiomics analysis was performed using MATLAB and Python-based PyRadiomics. Following feature selection, 14 radiomic features were retained to construct a radiomics score. Three predictive models were developed and compared. From 1,652 extracted radiomic features, 14 were selected to build a radiomics score significantly associated with LN metastasis (P<0.001). Multivariate logistic regression identified tumor location, vascular invasion, white blood cell (WBC) count, platelet count, and radiomics score as independent predictors of LN metastasis, and a nomogram was generated. Various machine learning algorithms were evaluated to identify the optimal model combining radiomic and clinical data. Decision curve analysis demonstrated that the extreme gradient boosting (XGBoost) model achieved superior accuracy and stability, with an area under the curve (AUC) of 0.86 in the training set and 0.81 in the validation set. We successfully developed a LN metastasis prediction model integrating preoperative CT radiomics and clinical variables. The XGBoost-based model exhibited excellent accuracy, robustness, and generalizability. Future studies should focus on constructing more reliable models through standardized pathological sampling and prospective multicenter trials, alongside advancing imaging techniques to improve identification of small LNs.