Non-invasive predictive model for incidental gallbladder carcinoma based on multimodal features: Integrating clinical data, MRI radiomics, and deep transfer learning features.

Gallbladder carcinoma, among the most prevalent malignancies of the biliary system, often presents with insidious early symptoms. Delayed diagnosis of incidental gallbladder carcinoma frequently leads to therapeutic delays, significantly compromising patient prognosis. Clinical data,preoperative Magnetic Resonance Imaging examinations, and histopathological records from incidental gallbladder carcinoma patients were retrospectively enrolled. An integrated noninvasive predictive model was developed by combining deep learning features extracted from Magnetic Resonance Imaging radiomics with key clinical predictors. Data from 299 patients with benign gallbladder disease and 106 incidental gallbladder carcinoma cases were analyzed. Multimodal deep learning algorithms identified an optimal predictive model. Multivariate analysis revealed hemoglobin, direct bilirubin, age, distance of the cystic duct from the confluence of right and left hepatic ducts, and diameter of the common bile duct as independent predictors of incidental gallbladder carcinoma (all P<0.01). The multimodal combined feature-based prediction model (AUC=0.894[95%CI 0.829-0.960]) surpassed unimodal models (Clinic:0.862 [ 95%CI 0.787-0.936]; Rad: 0.797 [ 95%CI 0.716-0.878]; DLR: 0.844[ 95%CI 0.775-0.912]) in test cohort. Haemoglobin, direct bilirubin,age, distance of the cystic duct from the confluence of right and left hepatic ducts, and diameter of the common bile duct constitute independent predictors of incidental gallbladder carcinoma. The multimodal combined feature-based prediction model significantly outperforms single-modality models, offering a robust tool for preoperative risk stratification.