Acoustic neuroma segmentation using ensembled convolutional neural networks

Acoustic neuroma (AN) is a noncancerous and slow-growing tumor that influences the human hearing system. Magnetic resonance images (MRIs) are routinely utilized to monitor tumor progression. Quantifying tumor growth in an automated manner would allow more precise studies, both at the population level and for the clinical management of individual patients. In recent years, deep learning methods have shown excellent performance for many medical image segmentation tasks. However, most current methods do not work well on heterogeneous datasets where MRIs are acquired with vastly different protocols. In this paper, we propose a deep learning framework with ensembled convolutional neural networks (CNNs) to segment acoustic neuromas even in heterogeneous datasets. We ensemble a 2.5D CNN model and a 3D CNN model together, with augmentations added to the model for better inter-dataset segmentation performance. We test our methods on two datasets: the publicly available dataset from the crossMoDA challenge and an in-house dataset. We examine our method with supervised learning on the crossMoDA dataset and directly apply the trained model to the in-house dataset. We use the Dice score, average surface distance (ASD), and 95-percent Hausdorff distance (95HD) as evaluation metrics. Our method has better performance than the baseline methods, not only on intra-dataset segmentation accuracy but also on inter-dataset generalizability.