A 2.5D assembly framework to segment high-dimensionality medical images by Bayesian aggregation of parallel 2D CNNs

Abstract

Deep learning neural networks have been widely used in general 2D image processing tasks. However, its application to process high-dimensional medical images is impeded by the need to tailor the network for specific considerations on imaging systems and/or biological characteristics, and the high memory/computation cost as the image dimensionality increases. This study aims to design an assembly 2.5D image segmentation framework based on native 2D convolutional neural network (CNN), which is naturally adaptable to problem dimensionality changes in an economical way and intrinsically amicable to parallel processing. In particular, we perform soft segmentation along each 2D fiber using one native 2D CNN, aggregate such decisions based on Bayesian rule, and apply an (optional) polish step to geometrically regularize the raw segmentation. Validation experiments on volumetric CT liver segmentation demonstrate higher segmentation accuracy with pronounced cost-saving benefit, compared to the state-of-the-art 3D CNN and triplanar approaches.