CT Radiomics and Deep Learning-Based Non-Invasive Model for Tracing Primary Sites of Adenocarcinoma Liver Metastases: Automatic Differentiation of Multiple Primary Digestive Tumors

Brief summary public domain (ClinicalTrials.gov)

This is a retrospective observational study. Liver metastasis occurs frequently in digestive-system cancers. Identification of the primary tumor source of liver metastasis may pose clinical challenges. Biopsy is an invasive procedure associated with potential sampling errors. This study aims to develop and validate a non-invasive artificial-intelligence model based on contrast-enhanced CT images. The model integrates CT radiomics and deep learning to automatically determine the primary origin of adenocarcinoma liver metastases, including gastric adenocarcinoma, colorectal adenocarcinoma, biliary/gallbladder adenocarcinoma, and pancreatic ductal adenocarcinoma. The investigators will retrieve historical medical records and de-identified abdominal contrast-enhanced CT images from patients diagnosed at West China Hospital, Sichuan University between January 2015 and August 2026. No new intervention or treatment will be administered to participants in this study.

Detailed description public domain (ClinicalTrials.gov)

Liver metastasis represents the most common form of distant metastasis for digestive adenocarcinomas. Approximately 22% of patients with colorectal cancer and 52% of patients with pancreatic cancer present with liver metastasis at initial diagnosis. Two challenging clinical scenarios are recognized: participants with liver metastases derived from occult primary tumors (carcinoma of unknown primary, CUP), and participants with multiple primary malignancies in whom newly-detected hepatic lesions cannot be distinguished as metastatic deposits versus de-novo primary tumors. Pathological biopsy serves as the gold-standard for differential diagnosis; however, its invasiveness, sampling bias, and limited repeatability restrict routine clinical application. Prior radiomics and deep-learning investigations predominantly addressed binary-classification tasks and lacked multi-class external validation. This retrospective study will enroll approximately 500-700 eligible participants with pathologically-confirmed adenocarcinoma liver metastases. The investigators will extract radiomic features from pretreatment portal-venous-phase contrast-enhanced CT images, and construct and validate an integrated radiomics-deep-learning model for four-class primary-site differentiation. All imaging and clinical datasets will undergo full de-identification prior to analysis. No participant will receive novel clinical interventions for the purpose of this research.