A novel deep learning graph attention network for Alzheimer’s disease image segmentation

Abstract

Neuronal cell segmentation identifies and separates individual neurons in an image, typically to study their properties or analyze their organization in the nervous system. This is significant because neurological problems and diseases can only be treated effectively when the structure and function of neurons are understood. The proposed method is based on convolutional neural networks (CNNs) and graph attention networks (GATs) for segmenting biomedical images. A contracting path built upon a couple of convolution layers and max pooling is included in the architecture to capture context. After that, the GATs are applied to the captured context. In GATs, each node in the graph is associated with a vector of hidden features, and the model calculates attention coefficients between pairs of nodes. These attention coefficients are learned during training and can be used to weigh the contribution of each node’s features to the representation of the graph. An expanding path that utilizes the outputs generated by GATs paves the way for exact segmentation. The dataset comprises 606 microscopic images, mainly categorized into different cell types (astrocytes, cortex, and SHSY5Y). By implementing our proposed U-GAT algorithm, we obtained the highest accuracy of 86.5% and an F1 score of 0.719 compared to the CNN, U-Net, SegResNet, SegNet VGG16, and GAT benchmarking algorithms. This proposed method could help researchers in the biotech industry develop novel drugs since a more accurate deep-learning method is essential for segmenting complex images like neuronal images.