Multiparametric-MRI-Based Radiomics Model for Differentiating Primary Central Nervous System Lymphoma From Glioblastoma: Development and Cross-Vendor Validation

Abstract

Background: Preoperative differentiation of primary central nervous system lymphoma (PCNSL) from glioblastoma (GBM) is important to guide neurosurgical decision-making.

Purpose: To validate the generalization ability of radiomics models based on multiparametric-MRI (MP-MRI) for differentiating PCNSL from GBM.

Study type: Retrospective.

Population: In all, 240 patients with GBM (n = 129) or PCNSL (n = 111).

Field strength/sequence: 3.0T scanners (two vendors). Sequences: fluid-attenuation inversion recovery, diffusion-weighted imaging (DWI), and contrast-enhanced T₁ -weighted imaging (CE-T₁ WI). Apparent diffusion coefficients (ADCs) were derived from DWI.

Assessment: Cross-vendor and mixed-vendor validation were conducted. In cross-vendor validation, the training set was 149 patients' data from vendor 1, and test set was 91 patients' data from vendor 2. In mixed-vendor validation, a training set was 80% of data from both vendors, and the test set remained at 20% of data. Single and multisequence radiomics models were built. The diagnoses by radiologists with 5, 10, and 20 years' experience were obtained. The integrated models were built combining the diagnoses by the best-performing radiomics model and each radiologist. Model performance was validated in the test set using area under the ROC curve (AUC). Histological results were used as the reference standard.

Statistical tests: DeLong test: differences between AUCs. U-test: differences of numerical variables. Fisher's exact test: differences of categorical variables.

Results: In cross-vendor and mixed-vendor validation, the combination of CE-T₁ WI and ADC produced the best-performing radiomics model, with AUC of 0.943 vs. 0.935, P = 0.854. The integrated models had higher AUCs than radiologists, with 5 (0.975 vs. 0.891, P = 0.002 and 0.995 vs. 0.885, P = 0.007), 10 (0.975 vs. 0.913, P = 0.029 and 0.995 vs. 0.900, P = 0.030), and 20 (0.975 vs. 0.945, P = 0.179 and 0.995 vs. 0.923, P = 0.046) years' experiences.

Data conclusion: Radiomics for differentiating PCNSL from GBM was generalizable. The model combining MP-MRI and radiologists' diagnoses had superior performance compared to the radiologists alone.

Level of evidence: 4 TECHNICAL EFFICACY STAGE: 2.

Keywords: GBM; MP-MRI; PCNSL; machine learning; radiomics.