Accurately discerning primary central nervous system lymphoma (PCNSL) from glioblastoma (GBM) on preoperative MRI is hard but crucial, because early steroid use can impair PCNSL biopsy yield and delay targeted therapy. In our iScience study, we trained a DenseNet169 model on a balanced cohort (68 PCNSL, 69 GBM) using only standard sequences (T1, T1-CE, T2, FLAIR). On a held out 20-patient test set the network achieved 80% accuracy with an AUC of 0.9, slightly outperforming clinicians (79%).
Importantly, human and AI tended to err on different cases; by using the AI only where humans disagreed, the combined approach reached 95% accuracy. Saliency analyses indicated the model focused on tumor regions, and performance did not depend on specialized imaging or heavy manual annotation, pointing to practical deployability in real-world workflows.