Multi-algorithm radiomics machine learning models integrating ultrasound imaging and inflammation-immune features for...

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Differentiating hepatic metastases (HM) from primary liver malignant tumors (PLMC) is challenging due to similar pathological foundations and imaging findings.
• Ultrasound imaging alone has limitations in distinguishing between HM and PLMC.

Purpose of the Study:

• To explore the diagnostic value of multi-algorithm radiomics machine learning models.
• To integrate ultrasound imaging features with inflammation-immune markers for improved HM identification.

Main Methods:

• Retrospective analysis of 104 patients with hepatic malignancies, divided into training (60%) and internal validation (40%) cohorts, with an external validation cohort of 45 patients.
• Extraction of 107 radiomic features using PyRadiomics and identification of significant features via Wilcoxon test.
• Development of 98 machine learning models integrating radiomic and inflammation-immune features (platelet-to-lymphocyte ratio [PLR] and platelet-to-albumin ratio [PAR]), evaluated using ROC curves and SHAP analysis.

Main Results:

• 15 radiomic features significantly associated with HM were identified (P < 0.001).
• The optimal radiomics model (random forest + gradient boosting machine) achieved an AUC of 0.77 (training), 0.82 (internal validation), and 0.70 (external validation).
• Integrating radiomic features with PLR/PAR improved identification in the training cohort (AUC=0.84) and external validation (AUC=0.74), with PAR being a key predictor.

Conclusions:

• Radiomics features and inflammation-immune markers demonstrate potential clinical utility in identifying HM.
• Machine learning models integrating imaging and blood-based biomarkers offer a promising approach for HM diagnosis.
• Further research is warranted to validate and refine these integrated models for clinical application.