JointSwinUNETR: an Efficient Feature-enhanced Architecture for Small Intestine Cine MRI Segmentation

Abstract

The Cine MRI of the small intestine is a dynamic magnetic resonance imaging technique used to observe and evaluate small intestine motility. It captures sequential images of the organ in motion over time through rapid imaging. The Transformer architecture is highly effective at capturing long-range dependencies, enabling it to extract more comprehensive temporal information. As the image resolution provided by medical devices continues to improve, images now contain far more intricate details. However, most existing Transformer-based 3D segmentation models, constrained by computational resources, cannot directly handle high-resolution sequential Cine MRI data. In this paper, we propose an efficient joint training architecture for high-resolution 3D small intestine segmentation, consisting of a Local Sequential Encoder (LSE), a Global Multi-Scale Encoder (GMSE), and a Spatial Sequential Hybrid Decoder (SSHD). The LSE partitions high-resolution temporal images into multiple patches and uses a parameter-sharing mechanism to extract local detail features with a low computational cost. To prevent the loss of global features due to patch partitioning, the GMSE complements the LSE through joint training. Both encoders are trained simultaneously, the LSE captures local details, while the GMSE focuses on global, multi-scale features. Their combined outputs enable the model to balance fine-grained and large-scale information, improving segmentation accuracy. Furthermore, to address the periodic nature of small intestine peristalsis, we propose a dynamic consistency loss, which effectively quantifies motion amplitude differences between predicted and ground truth sequences. Experimental results on our dataset demonstrate that the proposed architecture significantly improves prediction accuracy compared to existing advanced 3D segmentation methods, while also delivering superior performance with low computational cost than recent advanced methods.