Deep Learning Pipeline for Automatic Segmentation, Classification, and Molecular Subtyping of Three Pediatric Posterior Fossa Tumors Using T2-Weighted MRI.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
- Institute of Research and Clinical Innovations, Neusoft Medical Systerns co., Ltd., Beijing, China.
- Philips Healthcare, Shanghai, China.
- Department of Neurosurgery, Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
- Department of Pathology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
- Department of Radiology, the First Affiliated Hospital of USTC, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, Anhui, China.
- Department of Radiology, West China Hospital of Sichuan University, Chengdu, China.
Abstract
Pediatric posterior fossa tumors vary in malignancy, treatment, and prognosis across tumor types and molecular subtypes, yet noninvasive preoperative differentiation remains challenging. To develop a deep learning (DL) pipeline using T2-weighted (T2w) MR images to automatically segment pediatric posterior fossa tumors, differentiate tumor types (medulloblastoma [MB], ependymoma [EP], pilocytic astrocytoma [PA]), and classify molecular subtypes of MB and EP. Retrospective and prospective. 1305 patients (M/F: 828/477; 490 MB, 327 EP, and 488 PA) from three centers. For tumor segmentation and classification, PF-nnU-Net was developed on the training set (n = 880) and validated on a validation set (n = 220), an internal prospective test set (n = 90), and two external independent test sets (n = 68, n = 47). MB-nnU-Net was trained on 338 patients and tested on 63 patients for MB subtyping; a prior developed EP-nnU-Net was tested on 38 patients for EP subtyping. 1.5 T or 3 T MRI, axial T2w images (turbo spin echo). Three nnU-Net-based models: PF-nnU-Net and MB-nnU-Net for development, EP-nnU-Net for validation. Five-fold cross-validation was performed on training sets, followed by testing on independent test sets. Dice similarity coefficient for segmentation. Accuracy, sensitivity, specificity, the area under the receiver operating characteristic curve (AUC), and Cohen's kappa for classification. A two-sided p < 0.05 was considered significant. PF-nnU-Net achieved Dice scores of 0.94-0.96 and overall classification accuracy of 0.824-0.918 (multiclass Cohen's kappa: 0.722-0.873). MB-nnU-Net attained an overall accuracy of 0.794 (multiclass Cohen's kappa: 0.605), and EP-nnU-Net achieved an accuracy of 0.789 (Cohen's kappa: 0.538). A fully automated DL pipeline was developed and validated to accurately segment pediatric posterior fossa tumors, differentiate tumor types (MB, EP, PA), and classify MB and EP molecular subtypes. 3. Stage 2.