A multimodal prediction framework for colorectal cancer peritoneal metastasis: CT-based tumor and adipose tissue analysis.

Accurate preoperative assessment of CT-occult peritoneal metastasis (PM) in colorectal cancer is critical for clinical decision-making. This study demonstrates that radiomic features derived from visceral adipose tissue (VAT) provide substantial predictive value for occult PM. Accordingly, we developed a multimodal prediction model integrating deep learning-based tumor features, VAT radiomics, and clinical variables. Specifically, an attention-based dual-branch deep convolutional neural network (DenseNet121-MARNet) was employed to extract high-dimensional deep features from tumor regions on CT images. In parallel, quantitative VAT radiomic features were extracted from L3 vertebral-level CT slices using high-throughput radiomics analysis. These heterogeneous imaging features were subsequently integrated with key clinical characteristics selected via logistic regression. Multiple machine learning classifiers, including random forest, logistic regression, and support vector machine (SVM), were systematically evaluated, with SVM identified as the optimal classifier. The proposed model achieved an area under the receiver operating characteristic curve (AUC) of 0.958 (95% CI: 0.915-1.000) in the internal test set (IntTS) and demonstrated robust generalization in the external test set (ExtTS), with an AUC of 0.913 (95% CI: 0.867-0.960). Decision curve analysis indicated a superior net clinical benefit across a wide range of threshold probabilities. Furthermore, in a reader study, the model significantly outperformed radiologists in diagnostic performance, exhibiting notably higher sensitivity for identifying occult PM-positive cases. These findings confirm that a multimodal predictive model incorporating VAT radiomic features offers a clinically valuable tool for preoperative individualized risk stratification in patients with colorectal cancer.