Hippocampus Segmentation with Transformer-UNet Fusion and Squeeze-And-Excitation Block

The hippocampus is an essential component of the brain, intimately involved in short-term memory, long-term memory, and spatial navigation. Diseases such as Alzheimer's, epilepsy, schizophrenia, and others are highly correlated with hippocampal atrophy. To delve into the role of the hippocampus in these diseases, it is crucial to segment and quantify the hippocampus. However, the hippocampus accounts for a relatively small proportion of the entire brain and has unclear boundaries with surrounding structures, posing challenges for hippocampal segmentation tasks. In real medical scenarios, models must exhibit real-time and low-latency features, necessitating an improvement in segmentation accuracy while minimizing parameter count. We propose a U-shaped architecture integrating SEFormer and Cascaded Upsampling (CUP) as the encoder and decoder, respectively, within the UNet framework. The SEFormer block utilizes the structure of the Transformer framework, excluding the Multi-head Self-Attention, instantiated with a combination of Squeeze-and-Excitation modules and depth-wise separable convolutions. This architecture significantly reduces parameter count and computational complexity while maintaining good performance. The SE module captures crucial information from input features, enhancing the network's expressive power. We evaluated our experimental model using the widely used benchmark dataset HarP for hippocampal segmentation. The average Dice coefficient achieved was 0.9155. Compared to other models, ours exhibited superior performance with more precise segmentation edges, greatly enhancing its potential for deployment in real medical settings.