Multi-scale Segmentation- and Error-guided Generative Adversarial Network for Neuron Segmentation

Accurate instance segmentation of neurons is one of the core issues in brain science, and is the key to quantitatively analyzing information such as the number, morphology, and distribution of brain neurons. Accurate neuron semantic segmentation can provide accurate foreground neuron pixels, further improving the accuracy of neuron instance segmentation. However, the structure of the brain is complex, the intra-class differences among neurons is large, the inter-class differences among neurons is small, the neurons in certain anatomical regions are highly aggregated, and the distribution of neurons is uneven in different anatomical regions, causing new problems when existing semantic segmentation methods are applied to actual data, resulting in limited accuracy of semantic segmentation of neurons throughout the brain. In this paper, a multiscale segmentation- and error-guided generative adversarial network (MSEG-GAN) is proposed for neuron segmentation. This method applies the dual attention gate-based multiscale segmentation- and error-guided framework as the generator to predict neuron segmentation results, false negative errors and false positive errors. According to the three prediction results, the neuron semantic segmentation results are iteratively improved. Then, the generated segmentation results and the ground truth are, respectively, multiplied by the original images as the two inputs of the discriminator. The difference between the two branches is calculated by applying multiple convolutional operations to distinguish the generated segmentation results and the ground truth. The semantic segmentation results obtained by the proposed model in the macaque brain dataset were compared with seven reference methods (U-Net, AttU-Net, UNet++, MSEG-iCNN, SSAN, SegAN, SETR). Compared with the reference methods, the value of criterion Dice of the semantic segmentation performance of the whole brain neurons was, respectively, increased by 2.8%、0.5%、2.5%、1.3%、16.8%、1.4%、10.1%. Especially for light-stained neurons, the segmentation performance was effectively improved, and the Dice value was, respectively, increased by 20.3%、1.2%、15.5%、8.2%、159.2%、0.8%、13.1%. The proposed multiscale segmentation- and error-guided generative adversarial network can automatically and accurately segment the whole brain neurons, and can be further used for neuron instance segmentation and neuron morphological analysis.