Neuronal activity remodels the F-actin based submembrane lattice in dendrites but not axons of hippocampal neurons

Abstract

The nanoscale organization of the F-actin cytoskeleton in neurons comprises membrane-associated periodical rings, bundles, and longitudinal fibers. The F-actin rings have been observed predominantly in axons but only sporadically in dendrites, where fluorescence nanoscopy reveals various patterns of F-actin arranged in mixed patches. These complex dendritic F-actin patterns pose a challenge for investigating quantitatively their regulatory mechanisms. We developed here a weakly supervised deep learning segmentation approach of fluorescence nanoscopy images of F-actin in cultured hippocampal neurons. This approach enabled the quantitative assessment of F-actin remodeling, revealing the disappearance of the rings during neuronal activity in dendrites, but not in axons. The dendritic F-actin cytoskeleton of activated neurons remodeled into longitudinal fibers. We show that this activity-dependent remodeling involves Ca 2+ and NMDA receptor-dependent mechanisms. This highly dynamic restructuring of dendritic F-actin based submembrane lattice into longitudinal fibers may serve to support activity-dependent membrane remodeling, protein trafficking and neuronal plasticity.

MICRA-Net: MICRoscopy Analysis Neural Network to solve detection, classification, and segmentation from a single simple auxiliary task

“…Figure 1 shows that image-level annotation reduced by more than 19 folds the time required by an expert to generate the training dataset compared to precise identification of the structure boundaries that would be required for fully-supervised DL approaches. This also corresponds to a reduction of the annotation time of more than 3 folds compared to the tracing of polygonal bounding boxes, which were recently used for weakly-supervised training of the U-Net architecture on this dataset [2].…”

“…The next question that needed to be addressed was the applicability of our approach for super-resolution microscopy image segmentation, for which precisely annotated datasets are rarely available. The auxiliary task was the semantic segmentation of STimulated Emission Depletion (STED) microscopy images of two nanostructures of the F-actin cytoskeleton in neurons: 1) a periodical lattice structure and 2) longitudinal fibers (Figure3a,b) [2]. The F-actin nanostructure segmentation task is challenging since the morphology of neurons is highly variable throughout the dataset, and there are many distractors around the structures of interest [2].…”

“…The auxiliary task was the semantic segmentation of STimulated Emission Depletion (STED) microscopy images of two nanostructures of the F-actin cytoskeleton in neurons: 1) a periodical lattice structure and 2) longitudinal fibers (Figure3a,b) [2]. The F-actin nanostructure segmentation task is challenging since the morphology of neurons is highly variable throughout the dataset, and there are many distractors around the structures of interest [2]. Figure 1 shows that image-level annotation reduced by more than 19 folds the time required by an expert to generate the training dataset compared to precise identification of the structure boundaries that would be required for fully-supervised DL approaches.…”

“…To alleviate the annotation process, weakly-supervised DL methods were introduced [11][12][13][14]. Bounding box annotations are commonly used for weakly-supervised segmentation tasks as they are simple, allow the task to be spatially constrainted [2,13,[15][16][17], and were shown to decrease the annotation phase by 15-fold compared to precise identification of structure boundaries [18]. Methods for training with binary, image-level targets, reducing even further the complexity and duration of the annotation task, have been proposed when multiple instances are displayed on a single image [19].…”

“…, each segmentation performance evaluation consisted of: i) MICRA-Net, ii) multi-expert boundingbox annotations (6 experts on 25 images of each structure: user-study), and iii) the U-Net trained with polygonal bounding boxes[2] (see Methods and Supplementary Note 2 for specific details). MICRA-Net achieved equivalent or superior segmentation performance on the precisely annotated dataset in comparison to both the user-study and the U-Net (SupplementaryFigure 5).…”