BACKGROUND: Different from typical primary central nervous system lymphoma (PCNSL), early-stage atypical PCNSL usually presents as patchy signal abnormalities without evident mass effect or significant contrast-enhancement and is prone to confusion with low-grade glioma (LGG). This study aims to develop an MRI-based radiomics model to differentiate early-stage atypical PCNSL from LGG. METHODS: Two cohorts consisting of early-stage atypical PCNSL patients, as well as LGG patients with similar radiological manifestations, were retrospectively recruited from West China Hospital of Sichuan University (WCHSU
PCNSL=75, LGG=138) and Chengdu Shangjin Nanfu Hospital (CSNH
PCNSL=35, LGG=72) to serve as the training set and external validation set, respectively. Within the training set, there were additional early-stage atypical lesions from 19 typical or advanced-stage PCNSL patients included as a supplement. MRI-based radiomics models were developed and validated based on these two cohorts. RESULTS: Nine radiomic features were selected as significant features, most of which are wavelet radiomic features. The best radiomics model achieved an area under the curve (AUC) of 0.929 (0.901-0.957) and an accuracy of 91.6% on the independent external validation set. The inclusion of 19 additional PCNSL patients improved the model's performance. CONCLUSIONS: The MRI-based radiomics model can accurately differentiate early-stage atypical PCNSL from LGG with similar radiological manifestations, allowing early-stage atypical PCNSL patients to receive timely and appropriate radiotherapy or chemotherapy while avoiding unnecessary surgical resection.