Machine Learning Methods for differentiation of Primary Central Nervous System Lymphoma and Glioma

Abstract

Abstract Purpose and Background. Accurate differentiation of primary central nervous system lymphoma (PCNSL) and glioma on Magnetic Resonance Imaging (MRI) is an important task because the two diseases have similar imaging features, but treatment options differ vastly. This purpose of this study was to develop various machine learning methods based on radiomics features extracted from contrast-enhanced T1-weighted, T2 and the two modalities fusion to predict PCNSL and glioma types and compare the performance of different models. Materials and Methods. A total of 82 patients from five Chinese medical centers with pathologically confirmed PCNSL and glioma were analyzed retrospectively, including 38 PCNSL and 44 glioma. Region of interest (ROI) was manually segmented on contrast-enhanced T1-weighted and T2 scans. 572 radiomics features of each patient on each modality were extracted. We explored six machine learning methods on single modality and multi-modality fusion radiomics features. The area under the receiver operating characteristic curve (AUC) and accuracy (ACC) were used to evaluate different models. Results The cohort was split into a training (57, 70% patients) and validation cohort (25,30% patients) according to stratified sampling strategy. For the T1-weighted images, among all models, the KNN performed best, with an accuracy of 0.947, 0.880, sensitivity of 0.885, 0.833, specificity of 1.00,0.909, and AUC of 0.993, 0.949 in the training and validation cohort, respectively. For the T2 images, among all models, the SVM performed best, with an accuracy of 0.982,0.880, sensitivity of 1.00,0.833, specificity of 0.978, 0.923, and AUC of 0.981, 0.917 in the training and validation cohort, respectively. The multi-modality fusion model achieved an accuracy of 1.000, 0.920, sensitivity of 1.00, 1.00, specificity of 1.00, 0.846, AUC of 1.00,0.994 in the training and validation cohort, which was significantly better than the single modality models. Conclusion The established machine learning prediction models based on single or multi-modality radiomics features were feasible and achieved high performance in the PCNSL and glioma differentiation, which showed the potential to help clinical decision-making.