Abstract
BACKGROUND: Hepatic metastases (HM) and primary liver malignant tumor (PLMC) share partially similar pathological foundations, which can make it difficult to differentiate them based on visual imaging findings. This study will explore the value of multi-algorithm radiomics machine learning models that integrate ultrasound imaging and inflammation-immune features in the identification of HM. METHODS: Patients with hepatic malignancies who had undergone ultrasound-guided biopsy and image acquisition were retrospectively included. A total of 104 patients were randomly divided into training and internal validation cohorts at a 6:4 ratio, while other 45 patients were assigned as the external validation cohort. The PyRadiomics package was utilized to extract 107 radiomics original features. The Wilcoxon test was employed to identify high-value features significantly associated with HM. Univariate analysis was employed to identify inflammation-immune risk features related to HM for model development. 12 machine learning algorithms were integrated to combine radiomics features with inflammation-immune features. Model performance was comprehensively evaluated through receiver operating characteristic curves, and Shapley additive explanations (SHAP) method. RESULTS: The Wilcoxon test identified 15 radiomics features significantly associated with HM (P < 0.001), which were subsequently presented to 12 machine learning algorithms to develop 98 models using two or more features. Among these, 68% of the models achieved moderate identify performance [area under the curve (AUC), 0.73-0.82] in the internal validation cohort. The optimal algorithm radiomics model (random forest + gradient boosting machine) was developed using 4 selected features, yielding AUCs of 0.77 in the training cohort and 0.82 in the internal validation cohort, 0.70 in the external validation cohort, respectively. Univariate analysis further identified platelet-to-lymphocyte ratio (PLR) and platelet-to-albumin ratio (PAR) as high-risk inflammation-immune features for HM. While integrating radiomic features with PLR/PAR enhanced identification performance in the training cohort (AUC = 0.84), this improvement was not significantly replicated in the internal validation cohort (AUC = 0.82), but there was an improvement in the external validation cohort (AUC = 0.74). SHAP analysis revealed that PAR contributed most to the integrated model's predictions, followed by the aforementioned 4 radiomic features. CONCLUSION: This study revealed the potential clinical utility of radiomics features and inflammation-immune features in HM identification. CLINICAL TRIAL NUMBER: Not applicable.