A Deep Learning-Based Workflow for Dendritic Spine Segmentation

Abstract

The morphological analysis of dendritic spines is an important challenge for the neuroscientific community. Most state-of-the-art techniques rely on user-supervised algorithms to segment the spine surface, especially those designed for light microscopy images. Therefore, processing large dendritic branches is costly and time-consuming. Although deep learning (DL) models have become one of the most commonly used tools in image segmentation, they have not yet been successfully applied to this problem. In this article, we study the feasibility of using DL models to automatize spine segmentation from confocal microscopy images. Supervised learning is the most frequently used method for training DL models. This approach requires large data sets of high-quality segmented images (ground truth). As mentioned above, the segmentation of microscopy images is time-consuming and, therefore, in most cases, neuroanatomists only reconstruct relevant branches of the stack. Additionally, some parts of the dendritic shaft and spines are not segmented due to dyeing problems. In the context of this research, we tested the most successful architectures in the DL biomedical segmentation field. To build the ground truth, we used a large and high-quality data set, according to standards in the field. Nevertheless, this data set is not sufficient to train convolutional neural networks for accurate reconstructions. Therefore, we implemented an automatic preprocessing step and several training strategies to deal with the problems mentioned above. As shown by our results, our system produces a high-quality segmentation in most cases. Finally, we integrated several postprocessing user-supervised algorithms in a graphical user interface application to correct any possible artifacts.

DeepD3, an Open Framework for Automated Quantification of Dendritic Spines

“…The benchmarking dataset was also segmented using two recently published automated spine detection methods (Vidaurre-Gallart et al, 2022; Singh et al, 2017). NB: centerline extraction and hence differentiation of spine and dendrite signal did not work in our hands using the code provided by Singh et al, (2017) reducing comparability to IoU scores of both spine and dendrite signal.…”

“…To address these concerns, various computational approaches of dendritic spine quantification have been described in the past (Extended Data Table 1) utilizing methods ranging from basic image thresholding to modern deep learning approaches (Koh et al, 2002; Dickstein et al, 2016; Xiao et al, 2018; Vidaurre-Gallart et al, 2022). These efforts improve the throughput of spine quantification and address the issue of reproducibility.…”

“…Using the segmentations of the benchmarking dataset provided by n=3 human experts, we next compared DeepD3 to state-of-the-art (semi-)automated methods for spine detection (Vidaurre-Gallart et al, 2022; Singh et al, 2017). DeepD3 (IoU for intersection: 0.474) outperformed a recently described fully automated deep learning method (Vidaurre-Gallart et al, 2022,; IoU for intersection: 0.278, Extended Data Figure 12), the semi-automated spine segmentation of the IMARIS platform (IoU of intersection: 0.343, Extended Data Figure 13), and a fully automated computational method (Singh et al, 2017, IoU for intersection: 0.422, Extended Data Figure 14). This indicates that DeepD3 is comparable or better than other state-of-the-art spine detection methods.…”

Between neurons and networks: investigating mesoscale brain connectivity in neurological and psychiatric disorders

“…Popular software like Imaris ( Govindan et al, 2021 ), or NeuroLucida ( Dickstein et al, 2016 ), followed by the utilization of semi-automatic measurement tools such as software like SPINEJ ( Levet et al, 2020 ) and NeuronStudio ( Rodriguez et al, 2008 ) have a broad use. To employ deep learning for automated methods, it requires extensive datasets comprising meticulously segmented, high-quality images, known as “ ground truth images ” ( Vidaurre-Gallart et al, 2022 ). However, it’s important to note that even with such datasets, there may still be limitations to achieving precise reconstructions ( Vidaurre-Gallart et al, 2022 ).…”

“…To employ deep learning for automated methods, it requires extensive datasets comprising meticulously segmented, high-quality images, known as “ ground truth images ” ( Vidaurre-Gallart et al, 2022 ). However, it’s important to note that even with such datasets, there may still be limitations to achieving precise reconstructions ( Vidaurre-Gallart et al, 2022 ).…”