<p>Semantic segmentation of the hippocampus aims to enhance the efficiency and accuracy of medical image analysis by precisely segmenting the hippocampus, thereby assisting doctors in making more accurate diagnoses and developing effective treatment plans. However, in 3D medical image segmentation tasks realized through downsampling using transformers within conventional U-shaped networks, the sequence-to-sequence prediction problem often neglects the connection between channel information and spatial information, leading to partial loss of channel information. To address this issue, we propose a network for hippocampal 3D image segmentation based on a hybrid attention mechanism with cross-dimensional interactions, termed CDI-Unet. The cross-dimensional interaction block (CDIB) is first introduced to address the separation of channel attention and spatial attention by capturing both spatial and channel dimensions of the input tensor, thereby enabling more comprehensive extraction of key segmentation semantic information from 3D medical images. The multilayer fusion block (MLFB) is then employed to replace the jump connections in the Unet in order to tackle the problems of category imbalance and feature loss. Additionally, a new space and channel reconstruction convolution block (SCRC) is designed to eliminate the substantial redundancy that arises when CNNs extract features, thus reducing computational load and redundant features. Experimental results on current mainstream datasets demonstrate that our CDI-Unet outperforms existing methods in all metrics, achieving better scores and performance.</p>