Abstract
BACKGROUND: Ki-67 is a key marker of tumor proliferation. This study aimed to develop machine learning models using single- and multi-parameter MRI radiomic features for the preoperative prediction of Ki-67 expression in primary central nervous system lymphoma (PCNSL), aiding prognosis and individualized treatment planning. METHODS: A retrospective analysis of 74 patients was conducted using MRI scans, including T1, contrast-enhanced T1, T2, T2-FLAIR, DWI, and ADC sequences. Patients were categorized into high-expression (Ki-67 > 70%) and low-expression (Ki-67 ≤ 70%) groups. Tumor volumes of interest (VOIs) were manually delineated by radiologists, and 851 radiomic features were extracted using 3DSlicer. After preprocessing, including bias field correction and normalization, feature selection was performed using SelectKBest and ANOVA. Eight machine learning classifiers, including Logistic Regression, Random Forest, and SVM, were applied to single- and multi-parameter datasets. RESULTS: Multiparameter models, particularly Naive Bayes and Logistic Regression, demonstrated superior predictive performance (AUC: 0.78, 0.73; AP: 0.90, 0.83) compared to single-parameter models. Decision curve analysis highlighted that Logistic Regression provides the highest net benefit, followed by Naive Bayes. CONCLUSION: Multiparameter MRI models are more accurate and stable for predicting Ki-67 expression in PCNSL, supporting clinical decision-making.