Machine Learning and Deep Learning CT-Based Models for Predicting the Primary Central Nervous System Lymphoma and Glioma Types: A Multicenter Retrospective Study

PURPOSE AND BACKGROUND: Distinguishing primary central nervous system lymphoma (PCNSL) and glioma on computed tomography (CT) is an important task since treatment options differ vastly from the two diseases. This study aims to explore various machine learning and deep learning methods based on radiomic features extracted from CT scans and end-to-end convolutional neural network (CNN) model to predict PCNSL and glioma types and compare the performance of different models. METHODS: A total of 101 patients from five Chinese medical centers with pathologically confirmed PCNSL and glioma were analyzed retrospectively, including 50 PCNSL and 51 glioma. After manual segmentation of the region of interest (ROI) on CT scans, 293 radiomic features of each patient were extracted. The radiomic features were used as input, and then, we established six machine learning models and one deep learning model and three readers to identify the two types of tumors. We also established a 2D CNN model using raw CT scans as input. 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 (70, 70% patients) and validation cohort (31,30% patients) according to the stratified sampling strategy. Among all models, the MLP performed best, with an accuracy of 0.886 and 0.903, sensitivity of 0.914 and 0.867, specificity of 0.857 and 0.937, and AUC of 0.957 and 0.908 in the training and validation cohorts, respectively, which was significantly higher than the three primary physician's diagnoses (ACCs ranged from 0.710 to 0.742, p < 0.001 for all) and comparable with the senior radiologist (ACC 0.839, p = 0.988). Among all the machine learning models, the AUC ranged from 0.605 to 0.821 in the validation cohort. The end-to-end CNN model achieved an AUC of 0.839 and an ACC of 0.840 in the validation cohort, which had no significant difference in accuracy compared to the MLP model (p = 0.472) and the senior radiologist (p = 0.470). CONCLUSION: The established PCNSL and glioma prediction model based on deep neural network methods from CT scans or radiomic features are feasible and provided high performance, which shows the potential to assist clinical decision-making.