Predicting microwave ablation results in colorectal carcinoma liver metastases by artificial intelligence using radiomics and clinical data

Predicting Microwave Ablation Results in Colorectal Carcinoma Liver Metastases by Artificial Intelligence Using Radiomics and Clinical Data Purpose: The aim of our study was to investigate the prediction of early local tumor response and the risk of local recurrence (LR) development in patients with Colorectal Carcinoma Liver Metastasis (CCLM) treated with microwave ablation (MDA) by machine learning models obtained magnetic resonance imaging (MRI) based radiomics features and clinical data. Material and Methods: In this single-center retrospective study, 42 patients (67 tumors) who were diagnosed with CCLM between March 2012 and March 2022, underwent MDA procedure with the decision of the tumor council of our hospital, had optimal MRI in the 1st and 3rd months before and after the procedure, and had sufficient clinical and laboratory data were included. A total of 111 radiomics features were extracted for each tumor and for each phase by manual segmentation of MRI T2 fat-suppressed (called Phase 4) and early arterial phase T1 fat-suppressed sequences (called Phase 1) obtained before ablation. A clinical model was constructed using clinical features, 2 combined models were created with feature reduction and machine learning by combining clinical data and independent Phase 1, Phase 4 radiomics features. The sensitivity and selectivity of these models in evaluating early local tumor response and predicting LR after MDA treatment were investigated. Findings: Complete response was detected in the 1st month control MRI in 42 patients (67 tumors) in total, and LR developed in 7 patients (16.6%) and 11 tumors (16.4%) in the 3rd month control MRI. In the clinical model, the presence of extrahepatic metastases was associated with a high probability of LR. (p<0.001) There was no statistically significant difference in the rate of tumor size >3cm in the patient groups with and without LR (p=0.051). Although the median of tumor markers such as carbohydrate antigen 19-9 (Ca 19-9) and carcinoembryonic antigen (CEA) measured before ablation was higher in the LR group than in the non-LR group (p= 0.010 and p=0.020), there was no statistically significant difference between the groups with and without LR in terms of tumor markers obtained before ablation (p>0.05). Patients with LR in both phases had significantly higher rad scores than patients without LR (p=0.001 for Phase 1 and p<0.001 for Phase 4). In the study, the classification performance of the combined model 2, which was created by using radiomics features obtained from Phase 4 MRI and clinical data, achieved the highest discriminative performance in predicting LR (p=0.014). The area under the ROC curve (AUC) value was 0.981(95% CI: 0.948-0.99, p<0.001). It was found that the combined model 1, created using Phase 1 MRI-based radiomics features and clinical data (AUC value 0.927(95%CI: 0.860-0.993, p<0.001)) and the clinical model (AUC value of 0.887 (95% CI: 0.807-0.967, p<0.001)) had similar performance in predicting LR. Conclusion: Combined models using clinical, laboratory data and radiomics features obtained from T2 fat-suppressed and early arterial-phase T1 fat-suppressed MRI are valuable biomarkers in predicting tumor response and local recurrence in patients with colorectal carcinoma liver metastases treated with microwave ablation. Keywords: Colorectal Carcinoma Liver Metastasis, Local Recurrence, Magnetic Resonance Imaging, Microwave Ablation, Radiomics.