Author: Sahaja Acharya, Matthew Ladra, Junghoon Lee, Lina Mekki, Bohua Wan 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Department of Biomedical Engineering, Johns Hopkins University, Department of Computer Science, Johns Hopkins University 🌍
Purpose: Cerebellar mutism syndrome (CMS) is the most frequently observed complication in children undergoing surgical resection of posterior fossa tumors. Previous work explored lesion to symptom mapping with statistical tools to predict CMS and few have investigated deep learning-based prediction model. In this work, we present our preliminary work in adopting 3D Residual Convolutional Neural Network (3D-RCNN) to predict CMS based on preoperative multiparametric MRI (mpMRI).
Methods: Preoperative T1 contrast-enhanced and T2 FLAIR MRIs, and tumor contours of 87 patients were collected and used as the input to train a 3D-RCNN model to predict CMS. The data were divided into 51 for training, 18 for validation, and 18 for test. Random contrast adjustment, horizontal flip, rotation, scaling, translation, and shear transform are used for data augmentation. Gradient-weighted Class Activation Mapping (Grad-CAM) is used to generate heat map highlighting important regions in the images for model prediction.
Results: For 18 independent test sets, the model achieved prediction performance with area under the curve (AUC) of 0.79, accuracy 0.72, sensitivity 0.50, specificity 0.79, accuracy 0.72, negative predictive value 0.85, and positive predictive value 0.4, showing that the 3D-RCNN model is a promising approach to predict CMS even if the number of data is limited. The CAM highlights the tumor as well as surrounding regions, implying that the model primarily pays attention to the tumor characteristics in relation to the surrounding structures.
Conclusion: A 3D-RCNN model is developed to predict postoperative CMS based on preoperative mpMRI and tumor contour. The model shows promising preliminary results and the CAM highlights regions around tumor showing the prediction is based on tumor textures and locations, which is consistent with previous studies. In the future, we will conduct comprehensive optimization and testing of the model with a larger cohort of data.