A method for magnetic resonance splenomegaly image segmentation based on large-kernel multi-scale attention mechanism.

The abnormal enlargement of the spleen plays a critical role in the diagnosis and treatment of infectious and hematologic diseases. In the clinical and pharmacological research for the proper diagnosis, monitoring, and surgical planning of diseases, an accurate assessment of the Speen size possesses gargantuan significance. In MRI images, achieving high-quality automatic segmentation can effectively quantify splenomegaly. However, automatic segmentation techniques face multiple challenges due to the low contrast between spleen boundaries and surrounding tissues and the extremely variable size and shape of the spleen. Hence, this study proposes an improved neural network structure LMA-Net (Large-kernel Multi-scale Attention Net) to efficiently address the problems of splenomegaly image segmentation in MRI images. The proposed framework adopts a U-shaped architecture and employs a large-kernel multi-scale feature fusion attention mechanism to enhance the ability to identify and segment the spleen, focusing on its key features and regions. Furthermore, the proposed model utilizes a lightweight feature fusion technique in the decoder stage to enhance segmentation accuracy and processing efficiency. The presented frameworks were experimentally validated on the dataset of MRI splenomegaly and the MSD Spleen that achieved the mIoU values of 96.13% and 94.41%, respectively, proving it as a promising approach.