Comparison of Radiomic Features from Different MRI Sequences for Predicting Synchronous Liver Metastases after Rectal Cancer

introductionSynchronous liver metastases (SLM) critically influence prognosis in rectal cancer, highlighting the need for accurate preoperative detection. This study aimed to compare the predictive performance of radiomic features extracted from T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) MRI sequences and to develop machine learning-based predictive models for the early detection of SLM in rectal cancer patients.

methodsThis retrospective study included 160 rectal cancer patients confirmed by pathology at our institution between September 2018 and June 2023. After screening, 137 patients were enrolled, comprising 71 patients with SLM and 66 without SLM. Clinical characteristics such as age, gender, tumor (mrT) staging, lymph node (mrN) staging, tumor size, tumor distance from the anal verge, location, and circumferential range were analyzed, with mrT and mrN staging showing statistical significance (p < 0.012). Radiomic features were extracted from regions of interest (ROIs) on T2WI and DWI using Pyradiomics after manual segmentation in ITK-SNAP. A total of 3,452 radiomic features (1,726 each from T2WI and DWI) were extracted, of which 14 features (4 from T2WI and 10 from DWI) were selected using the LASSO. Predictive models were developed using three machine learning algorithms: Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF), with a five-fold cross-validation strategy.

resultsAmong the machine learning algorithms, the RF consistently outperformed LR and SVM across all models. The Optimal model yielded the highest predictive performance, with RF achieving an AUC of 0.82 (95% CI: 0.66-0.93), an accuracy of 0.71, and an F1-score of 0.74. RF also showed superior performance in the Combined-Optimal model (AUC = 0.76, accuracy = 0.71). In contrast, models built using LR and SVM algorithms demonstrated moderate performance, with lower AUC values ranging from 0.68 to 0.70. Confusion matrix analysis confirmed RF's superior classification ability, accurately predicting SLM and non-SLM cases. DISCUSSION: The incorporation of radiomics and RF-based models conveys a promising, non-invasive approach for enhancing early detection and risk stratification of SLM, which could help with more reliable clinical decision-making and individualized treatment planning for patients with rectal cancer.

conclusionThe optimal feature set-based predictive model demonstrated the highest accuracy for SLM detection, with the RF algorithm outperforming LR and SVM by consistently achieving the best AUC and balanced diagnostic performance.