Fast and lightweight network improves serial brain section stitching

Abstract

High-resolution three-dimensional brain image reconstruction is crucial for understanding the brain. Light sheet microscopy combined with tissue clearing imaging plays a pivotal role in analyzing the micro-level structure of mammalian brains. However, the complex multi-level stitching process poses challenges such as non-overlapping areas, surface deformation, and tissue loss, resulting in incomplete or discontinuous tissue structures at the junctions. These issues not only impact the precision of the atlas but also complicate subsequent analyses like cell counting and neuron tracing. To address these issues, we propose a rapid deep learning-based image inpainting approach for accurate neuron reconstruction and analysis. Our approach involves initially employing conventional registration algorithms to preliminarily stitch brain sections together, followed by utilizing a neural network to predict and restore missing tissue with a thickness exceeding 10 µm. This process enhances the structural continuity and integrity between adjacent brain slices. Compared to the original 3D U-Net and ResNet models, our approach performs better and has a processing speed that is five times faster than the original 3D U-Net. Moreover, our method enables more accurate cell counting by repairing incomplete cell bodies, leading to an average improvement of 37.37% in the number of cell bodies accurately counted near the slice junction. By integrating this novel 3D image inpainting network into brain reconstruction processes, our research opens new avenues for a more detailed and accurate investigation of neural circuitry and neurological disorders.