Preoperative prediction of metachronous liver metastasis in colorectal cancer using a deep learning-based radiomics model with automatic segmentation: a multicenter study.

Abstract

Objectives: To develop and validate an integrated clinical-radiomics nomogram predicting the risk of metachronous liver metastasis (MLM) in patients with colorectal cancer (CRC).

Methods: In this multicenter retrospective study, 518 CRC patients underwent nnU-Net-based automatic segmentation of CT images. Radiomic features from CRC and liver regions were extracted and selected using LASSO regression. Independent clinical predictors were identified through multivariate logistic regression. Clinical, radiomic, and combined clinical-radiomic models were developed and independently validated in the training, validation, and test cohorts. Model performance was evaluated using the area under the curve (AUC), concordance index (C-index), and calibration. Feature importance was interpreted with SHapley Additive exPlanations (SHAP), and survival differences were analyzed using Kaplan-Meier curves.

Results: The nnU-Net demonstrated high segmentation accuracy (CRC DSC: 86.2% validation, 81.0% test; liver DSC: 96.5% validation, 94.7% test). AFP, dissected lymph nodes, perineural invasion, and tumor nodules were independent predictors. The combined model outperformed individual models (AUC: 0.972/0.875/0.814; C-index: 0.819/0.728/0.690; P < 0.05). SHAP analysis highlighted lymph node count, AFP, tumor nodules, and multiregional radiomics features as key contributors. High-risk patients exhibited a markedly reduced MLM-free survival (log-rankP < 0.0001) with good calibration.

Conclusion: An automatic segmentation model based on deep learning enables accurate delineation of CRC and liver regions. Furthermore, clinical-radiomic nomogram demonstrates high accuracy in identifying patients at elevated risk of metachronous liver metastasis, thereby supporting clinicians in optimizing therapeutic strategies and facilitating personalized treatment planning.