Accurately 3D neuron localization using 2D conv‐LSTM super‐resolution segmentation network

Abstract

Abstract Neuron localization is a fundamental step in the neuron morphology reconstruction and quantitative analysis of cell counting and spatial distribution. The recent development of labelling and imaging techniques has resulted in growing demand for automatic neuron localization methods of 3D neuronal microscopy images. However, accurately localizing touching neurons remains a challenge. To address this issue, a novel method utilizing a super‐resolution segmentation network based on 2D Conv‐LSTM and a region growing method are proposed. This network can map and detect individual neurons in higher resolution space, allowing for the separation of closely touching neurons with reduced resource consumption. Subsequently, a region growing method is employed to localize neurons accurately. This method is evaluated using neuronal images generated by TDI‐fMOST. This method achieved neuron localization with an F1 score of 0.91. In comparison, other automatic localization methods achieved F1 scores lower than 0.85. It is also demonstrated that our network has fewer computational requirements compared to 3D neural networks. This method is promising for accurately localizing neurons in large‐scale neuron images.

Keywords

• biomedical optical imaging
• brain
• medical image processing
• optical microscopy
• supervised learning