18F-FDG PET radiomics model for predicting TARE response in patients with colorectal cancer liver metastases.

PURPOSE: Predicting treatment response in patients with colorectal cancer liver metastases (CRCLM) who have undergone transarterial radioembolization (TARE) based on pre-procedural fluorine-18-fluoro-deoxy glucose positron emission tomography (18F-FDG PET) radiomics and clinical information.

MATERIALS AND

METHODS

Patients with CRCLM who underwent TARE, between March 2015 and May 2025, were consecutively included. Largest tumors were segmented semiautomatically using pre-procedural 18F-FDG PET images. Radiomics features were extracted, clinical information were collected. Two datasets were created comprising radiomics-only and clinico-radiomic features. Datasets were divided 60:40 for training and testing. Top 5 features were selected based on feature importances. Random Forest, Extreme Gradient Boosting, Logistic Regression models were trained. Test-set area under the curves (AUCs) for predicting post-treatment target lesion local progression were calculated and compared using DeLong's test. Sensitivity, specificity, accuracy and F1 scores were calculated at the optimal cut-offs.

RESULTS

Seventy-four patients out of 96 patients were included. Top five selected features in the radiomics-only dataset were Coarseness, IMC1, Zone Entropy, Size-Zone Non-Uniformity, and Strength. In the clinico-radiomic dataset, AST and ALT levels were substituted among the top five features. Radiomics-only features demonstrated AUCs ranging from 0.90 (95% CI 0.71-1) to 0.81 (95% CI 0.51-1) in the test-set while clinico-radiomics dataset AUCs varied between 0.88 (95% CI 0.51-1) and 0.84 (0.62-1).

CONCLUSION

18F-FDG PET radiomics based models can predict the local response to TARE in patients with CRCLM, in this series.