Deep learning for automatic detection and classification of pediatric brain tumors: A multi-institutional study
Michelle Han , Wasif Bala, Leo Chen , Jennifer Quon, Ed Lee , Michael Edwards, Gerald Grant, Lily Kim , Kristen Yeom, Katie Shpanskaya
Introduction: Pediatric brain tumors are the most common solid cancer in children. Tumor subtype often dictates prognosis and the optimal treatment regimen including the necessary degree of surgical resection. However, accurate diagnosis based on imaging is limited outside of major academic centers.
Objective: We sought to develop a deep learning model to automatically detect and classify four types of pediatric brain tumors using brain MRIs.
Methods: Pre-intervention T2-weighted brain MRIs were retrospectively collected from 185 normal patients and 617 patients with posterior fossa tumors from 5 academic institutions in North America. Ground truth diagnoses were determined radiographically for 122 diffuse intrinsic pontine glioma (DIPG), and pathologically for 272 medulloblastoma, 135 pilocytic astrocytoma, and 88 ependymoma patients. A residual ResNeXt-50-32x4d deep learning architecture was developed using 493 tumor/148 normal patients, with 124 tumor/37 normal patients as a held out test set. Model performance was compared against a blinded, board-certified neuroradiologist.
Results: On a held-out test set of 161 patients, the model correctly identified the presence of tumor with an AUROC of 0.99. The model’s overall tumor sub-type classification accuracy was 81% with an F1 score of 0.76, compared to a radiologist's accuracy of 65% and F1 score of 0.65 (Table 1).
Conclusion: We present the first deep learning model developed using a large multi-institutional dataset that accurately detects and classifies pediatric posterior fossa tumors using T2-weighted MRIs. This model can serve as a triaging tool and the foundation for ongoing work to automatically characterize genetic subgroups within each tumor type based on imaging.