Multimodal Survival Prediction Model for Colorectal Liver Metastases Post-Operation

This study is based on a strictly screened multicenter cohort of patients with colorectal liver metastases who underwent surgical intervention. First, high-throughput quantitative features are automatically extracted from the 3D regions of interest (ROIs) segmented based on pre-operative magnetic resonance imaging (MRI). Advanced machine learning dimensionality reduction algorithms are subsequently applied to eliminate redundant variables and select core imaging signatures that deeply reflect tumor micro-heterogeneity, microvascular proliferation, and invasive status. On this basis, these micro-level radiomics features are fused with macro-level clinical parameters, including liver function indexes and tumor load markers. Multivariable survival analysis models are performed to identify independent prognostic factors, which are further used to develop a visual and intuitive predictive nomogram tool. Finally, time-dependent evaluation metrics and an external validation cohort will be utilized to systematically test the dynamic predictive performance and cross-platform generalizability of the model. This multimodal dual-track data paradigm aims to achieve a non-invasive and efficient "digital optical biopsy" approach for prognostic evaluation.

Inclusion criteria

Age >= 18 years old. Confirmed diagnosis of colorectal cancer with liver metastases (CRLM). Underwent surgical resection for colorectal liver metastases. Preoperative liver magnetic resonance imaging (MRI) was performed with high-quality images available for 3D radiomics feature extraction.

Complete baseline clinical, laboratory, and post-operative follow-up survival data.

Exclusion criteria

Patients who received local ablation only (such as radiofrequency ablation) without surgical resection.

Preoperative MRI images with severe artifacts or poor quality that prevent high-throughput radiomics analysis.

Concurrent history of other primary malignant neoplasms. Missing key clinical variables or lost to follow-up immediately after surgical intervention.