ID #183 Preoperative prediction of cerebellar mutism syndrome using deep learning on multi-institutional MRI data
Bohua Wan, Lina Mekki, Gilbert Vezina, Alan Cohen, Kenneth Cohen, Jeff Michalski, Karin Walsh, Sahaja Acharya, Junghoon LeeAbstract
Background
Cerebellar mutism syndrome (CMS) occurs in approximately 20% of children following posterior fossa tumor resection and is associated with significant morbidity. While previous studies have explored statistical lesion-to-symptom mapping to predict this condition, robust preoperative patient-level prediction remains limited. We developed a deep learning model to predict the risk of postoperative CMS using preoperative multiparametric MRI (mpMRI).
Methods
Preoperative T1 contrast-enhanced, T2-weighted, and T2 FLAIR MRIs were collected from 385 patients with posterior fossa tumors, including 299 medulloblastoma patients from ACNS0331 and 86 patients from Johns Hopkins. Images were preprocessed to remove the skull and segment the tumor, brainstem, and cerebellum and input into a 3D Residual Convolutional Neural Network (3D-RCNN) to predict CMS. The dataset was split into a 10-fold cross-validation cohort (n = 350) and an independent test cohort (n = 35). An ensemble of four top-performing models from cross validation was used for final predictions and test set evaluation. Gradient-weighted Class Activation Mapping ++ (Grad-CAM ++) was used to visualize important regions in the images for model prediction.
Results
On the independent test set of 35 cases, the model achieved prediction performance with area under the curve of 0.92, accuracy of 0.83, sensitivity of 0.75, specificity of 0.85, negative predictive value of 0.92, and positive predictive value 0.60. Additionally, Grad-CAM ++ highlighted tumor-brainstem interface in 89% (31/35) cases, with the rest highlighting regions in the cerebellum-tumor interface, brainstem-cerebellum interface and/or contralateral cerebellum.
Conclusion
A 3D-RCNN model can predict postoperative CMS risk based on preoperative mpMRI, tumor, brainstem, and cerebellum contours. The model consistently implicates the tumor-brainstem interface, suggesting that spatial relationships to critical structures play a key role in CMS development. Preoperative identification of high-risk regions contributing to CMS at the individual patient-level may support surgical planning and patient counseling.