![]() Based on observations of a vast number of samples (see Supplementary material, Section 1), we identified several types of error: missed reconstruction and erroneous extra reconstruction due to entanglement, noise, or other artifacts. ![]() We closely studied the errors in reconstruction results from several automated algorithms, including ENT (Wang et al., 2017), APP2 (Xiao and Peng, 2013), and ST (Chen et al., 2015). In morphology reconstruction systems, therefore, the manual annotation time should be reduced to achieve high throughput, which means the errors resulting from automated reconstruction must be reduced. To obtain gold-standard morphology reconstruction, researchers need to curate reconstruction results with manual reconstruction platforms such as Vaa3D (Peng et al., 2014), TeraVR (Wang et al., 2019), or FNT (Gao et al., 2022) however, such manual annotation is labor-intensive and time-consuming, limiting the throughput of the morphology reconstruction workflow. However, even with these pre-processing and advanced deep learning–based approaches, the results of automated reconstruction still contain complex errors and cannot be used directly in analysis. Among them, weakly supervised learning (Huang et al., 2020) and false negative mining (Liu et al., 2022) are proposed to rescue and connect the weak and broken neurites in the segmentation step for reconstruction subgraph connection (SGC) method (Huang et al., 2022) starts from prediction map obtained by CNN to link the broken reconstruction crossover structure separation (CSS) method (Guo et al., 2021) is proposed to detect the crossover structures and generate deformed separated neuronal fibers in the images to eliminate entanglements in reconstruction. Deep learning–based approaches have been investigated for neuron tracing. Pre-processing algorithms, including multi-scale enhancement (Zhou et al., 2015), CaNE (Liang et al., 2017), and filtering-based enhancement (Guo et al., 2022) aim to enhance images by reducing background noise and improving image contrast. Moreover, for data sets with a low signal-to-noise ratio and dense neuron distribution with neuron fiber entanglement, the existing reconstruction algorithms do not show satisfactory performance. ![]() Owing to the complexity of the images and the limitations of automated reconstruction algorithms, these algorithms are unsuitable for whole-brain images. Existing automated reconstruction algorithms are generally effective only for a few specific data sets. The wide variety of brain images in terms of background noise, complicated branching patterns, and clutter of neuron fibers presents challenges for automated neuron reconstruction. Nevertheless, neuron morphology reconstruction remains an unsolved problem (Li S. For example, the 3D Visualization-Assisted Analysis software suite Vaa3D (Peng et al., 2014) has more than 32 plugins, including ENT (Wang et al., 2017), APP (Peng et al., 2011), APP2 (Xiao and Peng, 2013), NeuTube (Zhao et al., 2011), MOST (Wu et al., 2014), and ST (Chen et al., 2015). A large number of automated neuron reconstruction algorithms exist. Research institutions have also held competitions and established worldwide projects, such as the DIADEM competition (Liu, 2011) and BigNeuron (Peng et al., 2015 Manubens-Gil et al., 2023). Researchers have developed various manual, semi-automated, and automated neuron reconstruction tools for digital reconstruction of neuron morphology (Meijering, 2010). With these techniques, reconstruction of single-neuron morphology from optical microscopy images has become possible and now has an essential role in neuron science. In recent years, there has been considerable development of techniques, including sparse, robust, and consistent fluorescent labeling of a wide range of neuronal types (Peng et al., 2021) and fluorescence micro-optical sectioning tomography (fMOST Gong et al., 2016). Neuron morphology is considered to be a critical component of neuron cell type identification (Ascoli et al., 2008). ![]() Characterization of neuron cell type is an international research frontier in neuron science (Zeng and Sanes, 2017).
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |