.png){kind=icon}
.png){kind=icon}
Return
Neuron image segmentation algorithm based on dual encoder and Hadamard product
.png){kind=icon}
OA
AI
DOI：10.1007/s40747-026-02395-0.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Neuron electron microscopy images have the problems of unclear neuronal edges and textures, as well as complex and diverse neuronal structures. Existing neuron segmentation algorithms are computationally complex, have a relatively long inference time, and there is still room for improvement in terms of segmentation accuracy. Therefore, we propose a neuron image segmentation model based on dual encoder and Hadamard product (DH-UNet). The model is improved on the basis of the U-Net as follows. First, we use two encoding branches in the model to extract neuron features. In the first encoding branch, the fine-tuned MobileNetV3 Small is used to extract the rough features of neurons, and this structure is referred to as the ME structure in the paper. In the second encoding branch, a feature reconstruction unit is used to extract the complex features of neurons. The two encoders extract features respectively to obtain feature representations at different levels and abstraction degrees, thereby enhancing the model’s ability to model complex relationships. Secondly, we introduce a feature group fusion structure at the skip connection to fully integrate the features of the two encoding branches and reduce information loss. Immediately afterwards, in the decoding stage of the model, we use a module composed of the Hadamard product and depthwise separable convolution to further refine the features and then restore the semantic features. In addition, we explore the impacts of three loss functions on the model’s segmentation results. Finally, we conduct a large number of comparative experiments and ablation experiments on three datasets to further verify the effectiveness and feasibility of this method.
Keywords:
Convolutional neural networks
Neuron image segmentation
U-Net
Dual encoding structure
Journal
C
IF:
4.6
Papers:
271
Citations:
0