Automatic Segmentation and Classification of Glioblastoma and Solitary Brain Metastasis Using a Deep Learning Model on Multiparametric MRI.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiology, The Second Hospital of Tianjin Medical University, Tianjin, China.
- Department of Radiology, Liaocheng People's Hospital, Liaocheng, China.
- Department of Radiology, Qilu Hospital of Shandong University, Jinan, 250012, China.
- School of Medicine, Nankai University, Tianjin, China.
- Institute of Research and Clinical Innovations, Neusoft Medical Systems Co., Ltd., Beijing, China.
- Department of Radiology, Liaocheng People's Hospital, Liaocheng, China. [email protected].
- Department of Radiology, The Second Hospital of Tianjin Medical University, Tianjin, China. [email protected].
Abstract
Differentiating between glioblastoma (GBM) and solitary brain metastasis (SBM) before surgery is challenging yet essential for clinical decisions. We aimed to construct a three-dimensional (3D) deep learning (DL) model for the automated segmentation and classification of GBM and SBM based on multiparametric MRI. A total of 314 patients from one medical center and two public datasets were recruited. Tumor segmentation was performed using No-new-UNet (nnU-Net), followed by the development and comparison of 3D DL, 2D DL, and radiomics models for classification. Additionally, three radiologists with varying experience levels conducted a two-round classification analysis, with and without a 3D DL model as a reference. The nnU-Net segmentation of GBM and SBM achieved a Dice score of 0.917 and 0.915 on the training and testing sets. The performance of the 3D DL model, with an area under the curve (AUC) of 0.842, was superior to that of the 2D DL model, radiomics model, trainee radiologist, and experienced radiologist (AUC, 0.687, 0.720, 0.527, and 0.709, respectively) and was comparable to that of expert radiologist (AUC, 0.862). The performance of trainee, experienced, and expert radiologists improved significantly with the 3D DL model than without it as a reference (AUC, 0.724 vs. 0.527, 0.845 vs. 0.709, and 0.911 vs. 0.862, respectively). The MRI-based 3D DL model shows promising performance in automatically segmenting and classifying GBM and SBM. Reading with the 3D DL model improves the diagnostic accuracy of radiologists. Given its promising performance, it has potential for clinical translation as a decision-support tool.