MRI-Based Machine Learning for Prediction of Clinical Outcomes in Primary Central Nervous System Lymphoma

Area of Science:

• Neuro-oncology
• Medical Imaging
• Machine Learning

Background:

• Primary central nervous system lymphoma (PCNSL) can present with early relapse or refractory disease after initial treatment.
• Predicting treatment response is crucial for optimizing patient management and outcomes.

Purpose of the Study:

• To investigate the utility of MRI-based radiomics for predicting early relapse or refractory (R/R) disease in PCNSL patients.
• To evaluate the performance of machine learning algorithms in forecasting R/R PCNSL using radiomic features.

Main Methods:

• Retrospective analysis of pretreatment brain MRIs (T1WI, T2WI, T2 FLAIR) from 46 PCNSL patients.
• Extraction of 107 radiomic features (shape, statistical, texture) and calculation of apparent diffusion coefficient (ADC) values.
• Development of predictive models using five machine learning algorithms (SVM, k-NN, LDA, Naïve Bayes, Decision Trees) to identify R/R cases.

Main Results:

• 43.5% of patients experienced R/R disease.
• Higher predictive model scores and lower ADC values were significantly associated with R/R PCNSL (p < 0.05).
• The support vector machine model achieved the highest performance with 83% accuracy, 80% precision, and 0.78 AUC. Elevated SVM and Naïve Bayes scores correlated with reduced progression-free survival.

Conclusions:

• Preoperative MRI-based radiomics shows promise as a non-invasive tool for predicting R/R in PCNSL.
• Machine learning models, especially SVM, can effectively forecast treatment outcomes, potentially guiding personalized treatment strategies.