Development and validation of a convolutional neural network for automatic differentiation of primary central nervous system lymphoma and glioblastoma.
Authors
Affiliations (5)
Affiliations (5)
- Department of Neuro-Oncology, Cancer Center, China National Clinical Research Center for Neurological Diseases, Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
- National Institute for Data Science in Health and Medicine, Capital Medical University, Beijing, China.
- Department of Neuro-Oncology, Cancer Center, China National Clinical Research Center for Neurological Diseases, Beijing Tiantan Hospital, Capital Medical University, Beijing, China. [email protected].
- Department of Neuro-Oncology, Cancer Center, China National Clinical Research Center for Neurological Diseases, Beijing Tiantan Hospital, Capital Medical University, Beijing, China. [email protected].
- National Institute for Data Science in Health and Medicine, Capital Medical University, Beijing, China. [email protected].
Abstract
Primary central nervous system lymphoma (PCNSL) and glioblastoma (GBM) are two distinct types of malignant brain tumors, each requiring specific therapeutic approaches. Accurate differentiation between these tumors is crucial for selecting appropriate treatments. We developed and validated a 3D DenseNet264 convolutional neural network (CNN) to automatically differentiate PCNSL and GBM. A total of 141 patients initially admitted to Tiantan Hospital underwent preoperative T1Gd-MRI and were confirmed by histopathology. These patients were randomly divided into training and validation groups at a 7:3 ratio. Subsequently, the DenseNet264 was trained and validated using these datasets. External validation was performed using additional datasets from the Radiological Society of North America (RSNA) and patients previously admitted to Tiantan Hospital. Standardized image preprocessing was conducted following the RSNA-ASNR-MICCAI BraTS 2021 guidelines. A total of 623 patients (Tiantan Hospital: 535, RSNA: 88) were initially enrolled, of whom 316 patients (Tiantan Hospital: 228 [141 patients enrolled between December 2015 and December 2021, and 87 patients enrolled before November 2015], RSNA: 88; GBM: 159, PCNSL: 157) met the inclusion criteria. The DenseNet264 achieved optimal classification performance in the training set (AUC: 0.98) and validation set (AUC: 0.90). In held-out data from RSNA and patients enrolled earlier at Tiantan Hospital, the model showed similarly consistent performance (C-statistic: 0.77). We successfully developed and validated a robust deep-learning model capable of accurately differentiating PCNSL from GBM. This model provides a reliable, efficient, and cost-effective clinical decision-support tool for differential diagnosis.