Statistical and Image-Based Deep Learning Modeling to Predict Hearing Preservation in Vestibular Schwannoma Surgery.
Authors
Affiliations (6)
Affiliations (6)
- Department of Internal Medicine, Loyola University Chicago Stritch School of Medicine, Maywood, IllinoisUnited States.
- Department of Neurology, SUNY Downstate Health Sciences University, New York City, New YorkUnited States.
- Department of Neurosurgery, Loyola University Chicago Stritch School of Medicine, Maywood, IllinoisUnited States.
- Department of Neurology, University of Arizona College of Medicine, Phoenix, ArizonaUnited States.
- Department of Otolaryngology, Loyola University Chicago Stritch School of Medicine, Maywood, IllinoisUnited States.
- Department of Radiology, Loyola University Chicago Stritch School of Medicine, Maywood, IllinoisUnited States.
Abstract
Researchers have attempted to predict hearing preservation for vestibular schwannoma (VS) patients, with preoperative hearing scores and tumor size identified as useful predictors for hearing preservation. However, prior studies have not directly included magnetic resonance (MR) images in their predictive models, instead using secondary imaging measurements and other clinical features. Here, we created a deep learning model that directly incorporates preoperative images to predict VS patients' hearing preservation. Using 97 patients who underwent VS hearing preservation surgery at our institution, we modeled hearing preservation with clinical and image feature-based logistic regression (LR) and an image-based convolutional neural network (CNN) model, with or without clinical features added. LR models that included tumor size and preoperative hearing provided good hearing preservation discrimination (American Academy of Otolaryngology-Head and Neck Surgery [AAO-HNS] AUC = 0.79; Word Recognition Score [WRS] Area under curve (AUC) = 0.77). Two-channel MRI inputs were used for a 3D CNN, ResNet50, and a hybrid ResNet50 including tumor size and preoperative hearing. The hybrid model was our most accurate and was comparable with existing models (four-fold mean receiver operating characteristic [ROC] = 0.71; best ROC = 0.85). Previously described MRI features were not significantly correlated with hearing outcomes. We confirm that the first deep learning model to directly incorporate preoperative images and clinical features for VS patients can help predict hearing preservation. This type of model may help risk-stratify VS patients during surgical planning, with the goal of ultimately improving surgical outcomes for patients.