Neighbouring-slice Guided Multi-View Framework for brain image segmentation

Abstract

Automated segmentation of brain images is a critical task in neuroscience for brain registration, atlas generation, etc. Deep learning techniques have been widely investigated for segmenting brain images, where most existing methods only consider each brain slice independently in a single view, limiting their ability to explore the correlation among adjacent slices and spatial brain structures. This paper proposes a Neighbouring-slice Guided Multi-view Framework for automated segmentation of brain images. To fully utilize the information between neighbouring slices, we design a dual-decoder network to segment targets (e.g., regions/tumors) in brain images and the edge of each brain targets simultaneously, by calculating the difference between adjacent slices. Considering the fact that some brain images cannot be fully recognized in a single view, we integrate the neighbouring-slice strategy in a multi-view segmentation framework to fully explore spatial structures in 3D brains. The proposed framework is validated on the automated segmentation of diverse brain tumors and brain regions including CTX (cerebellar cortex), CP (caudoputamen), HPF (hippocampal formation), BS (brain stem), CB (cerebellum), and CBX (cerebellar cortex), in LSFM (Light-sheet Fluorescence Microscopy) and MRI (Magnetic Resonance Imaging) modalities, demonstrating superior performance in comparison with state-of-the-arts. The codes are released at: https://github.com/NeuronXJTU/NGMV.