W. Liao1, X. Luo2, and S. Zhang3; 1department of Radiation Oncology, Radiation Oncology Key Laboratory of Sichuan Province, Sichuan Clinical Research Center for Cancer, Sichuan Cancer Hospital& Institute. Sichuan Cancer Center, Affiliated Cancer Hospital of University of Electronic Science, Chengdu, China, 2School of Mechanical and Electrical Engineering, University of Electronic Science and Technology of China, Chengdu, China, 3Department of Radiation Oncology, Sichuan Cancer Hospital & Institute, Sichuan Cancer Center; Cancer Hospital affiliate to University of Electronic Science and Technology of China, Chengdu, China
Purpose/Objective(s): Deep learning-based algorithms can automatically segment brain tumors in magnetic resonance imaging (MRI), mostly previous works were based on a single-center dataset or just focused on a few subtypes of brain tumors. This work aimed to propose a fundamental deep learning model for segmenting nine types of brain tumors on heterogeneous multi-center datasets. Materials/
Methods: A total of 2826 patients from center A were retrospectively collected, and randomly split into 2283 patients and 543 patients for model training and internal testing, respectively. 269 patients from center B and 92 patients from center C were used to validate the model externally. Each dataset contained nine types of brain tumors. These datasets contained T1Gd MRIs acquired on diverse scanners using different pulse sequences and acquisition parameters. A 3D unified segmentation model was developed by extending the widely-used baseline nnUNet and trained on the large-scale dataset. The Dice similarity coefficient (DSC) and Normalized surface Dice (NSD) were calculated to evaluate the model performance. An observer study was performed in which two experienced radiation oncologists evaluated both the model and human-generated delineations. Results: In the internal testing dataset, the average DSC for glioma, meningioma, acoustic neuroma, ependymoma, craniopharyngioma, medulloblastoma, germ cell tumor, pineal gland tumor, and brain metastases (BMs) was 0.79, 0.86, 0.88, 0.86, 0.89, 0.87, 0.83, 0.72, and 0.72. The average NSD for the corresponding brain tumors was 0.76, 0.87, 0.91, 0.86, 0.91, 0.86, 0.88, 0.70, and 0.80. Comparable segmentation results were obtained in the two external testing datasets. And no significant difference was found among the three testing datasets for six subtypes of brain tumors (all p-values ? 0.05) except for BMs, germ cell tumor, and pineal gland tumor. The observer study indicated that the model resembled human delineations in 85%–90% of cases (n = 50) randomly selected from an internal testing dataset. Conclusion: The proposed model can automatically segment nine types of brain tumors with reasonable accuracy on multi-center datasets with highly variable MRI sequences. The model is robust and may be potentially used in real-world situations.
