Development and independent validation of explainable radiomics-based machine learning models for prognosis in colorectal liver metastases.

Introduction: Colorectal cancer frequently leads to liver metastases (CRLM), posing a major challenge to long-term survival. Prognosis remains heterogeneous, and traditional clinical risk scores often lack biological depth and spatial information. Advances in radiomics and machine learning (ML) offer the potential for improved, explainable outcome prediction; however, robust and interpretable prognostic models for CRLM remain an unmet need. This study aimed to develop and validate explainable ML models based on radiomic features extracted from both metastatic lesions and background liver tissue, enhancing the prediction of recurrence and overall survival (OS) status in patients with CRLM. Materials and methods: Patient data and contrast-enhanced CT images from two independent cohorts were analysed: a publicly available TCIA-CRLM series, employed as the discovery set, and a real-life clinical cohort, used as an external validation set. Segmentation focused on the largest liver metastasis (L-MAX) and surrounding healthy liver tissue (L-BKG), extracting radiomic features from both areas and their ratios (L-MAX/L-BKG). An end-to-end pipeline for data preprocessing and classification was designed. Multiple ML and Deep Learning (DL) classifiers were trained and validated. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) analysis to identify key predictive radiomic determinants. Performances were compared to recognized clinical models. Results: For recurrence prediction, the best-performing classifier was a soft-voting ensemble of a multilayer perceptron (MLP) optimized via a Genetic Algorithm (GA); for OS status classification, the best performance was obtained by a hard-voting ensemble of a GA-optimized MLP. Both classifiers demonstrated robust discrimination capabilities in external validation, with AUCs of 0.78 and 0.68, respectively. The explainability analysis performed with SHAP revealed the most relevant radiomic determinants in the classification. These features retained prognostic significance in the independent cohort, supporting their use for clinical risk stratification. Discussion: Explainable ML models leveraging both lesion-centric and contextual liver radiomics offer clinically transparent prediction of recurrence and survival in CRLM. SHAP highlighted clinically plausible, reproducible imaging determinants, enabling risk stratification. The validation of specific radiomic determinants suggests the potential practical utility of this approach, laying out the groundwork for integrating with DL and multi-omic data in future oncology strategies.