A Fully Automated Deep-Learning Model for Differentiating Diffuse Gliomas and Circumscribed Astrocytic Gliomas: A Multi-center Study

Motivation: The WHO classification divides gliomas into diffuse and circumscribed types, where accurate preoperative differentiation is crucial for treatment planning but remains challenging with conventional methods. Goal(s): To develop and validate a deep learning framework for automated preoperative segmentation and classification of diffuse gliomas and circumscribed astrocytic gliomas using multimodal MRI. Approach: A multi-center study implementing a two-stage deep learning method combining 3D U-Net segmentation and ResNet50 classification, trained and validated on CE-T1 and T2FLAIR sequences. Results: The model achieved high performance in both internal (AUC: 0.872) and external validation (AUCs: 0.735-0.794), significantly improving radiologists' diagnostic accuracy. Impact: The integrated deep learning framework demonstrates robust performance in segmenting and differentiating diffuse gliomas and circumscribed astrocytic gliomas across multi-institutional datasets. Notably, the system significantly enhanced the preoperative diagnostic performance of radiologists across all experience levels.