Skip to main content
SIR 2025 Event Main banner
BSCI - Consideration Cases Ad 1
Merit/DPSSME Siege
Back

SIR 2025

Interventional Oncology

Scientific Session

Session: Interventional Oncology 2

Radiomics and Explainable AI for Outcome Prediction in Metastatic Uveal Melanoma with Melphalan Percutaneous Hepatic Perfusion

Monday, March 31, 2025
3:54 PM – 4:03 PM CT
Location: Music City Center, 209 ABC
Purpose: Melphalan percutaneous hepatic perfusion (M-PHP) has shown promise in extending survival for patients with liver metastases from uveal melanoma (mUM). However, approximately 20% of patients experience disease progression despite treatment, making early identification of these individuals crucial. This study aims to predict patient outcomes by combining clinical datasets and baseline CT scan radiomics with machine learning (ML) models.

Materials and Methods: We conducted a multicenter retrospective analysis involving patients from two international institutions (UK and Netherlands), all treated with M-PHP and with available baseline CT scans. Radiomic textural features were extracted from regions of interest in the liver, annotated by radiologists. Several ML classifiers, including RUS Boosted Trees, SVM, Neural Networks, and KNN, were evaluated using 3-fold cross-validation to develop predictive models for identifying patients likely to experience disease progression. Model interpretability was enhanced using SHAP (SHapley Additive exPlanations) and partial dependence analysis.

Results: A total of 144 patients (mean age 58±11.1 years) were included, with complete imaging data available in 102 patients. A combination of clinical features and radiomic features (e.g., Gray Level Co-occurrence Matrix, Gray Level Run Length, Gray Level Size Zone Matrix, Neighbourhood Gray Tone Difference Matrix, and Neighbourhood Gray Level Dependence Matrix) were used to train the models. The RUSBoosted Trees classifier performed best in distinguishing between progressive and non-progressive disease (AUC = 0.84, True Positive Rate = 0.8, False Positive Rate = 0.2). Partial dependence analysis revealed that fewer treatment cycles, higher hepatic disease burden, and a greater number of lesions at baseline were associated with disease progression.

Conclusion: The integration of multimodal data, including clinical and radiomic features, demonstrates the potential to predict which mUM patients undergoing M-PHP are at risk of disease progression. These findings underscore the value of ML in guiding personalised treatment strategies for this rare and aggressive cancer.