Laplacian-guided hierarchical transformer: A network for medical image segmentation.

Abstract

Background and objective: Accurate medical image segmentation is crucial for diagnosis and treatment planning, particularly in tumor localization and organ measurement. Despite the success of Transformer models in various domains, they still struggle to capture high-frequency features, limiting their performance in medical image segmentation, especially in edge texture extraction. To overcome this limitation and improve segmentation accuracy, this study proposes a novel model architecture aimed at enhancing the Transformer's ability to capture and integrate both high-frequency and low-frequency features.

Methods: Our model combines the extraction of high-frequency features using a Laplacian pyramid with the capture of low-frequency features through a Local-Global Feature Aggregation Module. A Feature Interaction Fusion module is employed to integrate these features, focusing on target areas. Additionally, a new bridging module facilitates the transfer of spatial information between the encoder and decoder via layer-wise attention mechanisms. The model's performance was evaluated using the Synapse dataset with statistical measures such as the Dice Similarity Coefficient and Hausdorff Distance. The code is available at https://github.com/chenyuxiao123/LGHF.

Results: The proposed model demonstrated state-of-the-art performance in 2D medical image segmentation, achieving a Dice Similarity Coefficient of 84.10% and a Hausdorff Distance of 12.78. The evaluation metrics indicate significant improvements compared to existing methods.

Conclusion: This novel model architecture, with its enhanced capability to capture and integrate both high-frequency and low-frequency features, shows significant potential for advancing medical image segmentation. The results on the Synapse dataset demonstrate its effectiveness and suggest its application could improve diagnosis and treatment planning in clinical settings.