A segmentation method for cardiac MRI that incorporates region constraint guidance and tubular structure awareness...

Cardiac magnetic resonance imaging (CMRI) is a non-invasive medical examination method that provides a comprehensive evaluation of the anatomy, function, blood flow, and histology for cardiovascular diseases. Accurately segmentation of the left and right ventricles and myocardium from CMR images can significantly aid doctors in diagnosing cardiovascular diseases. However, due to the variable shapes of the right ventricle, narrow and tubular myocardial structures, and unclear boundaries caused by small grayscale differences between cardiac substructures, CMR image segmentation remains a challenging task.

To address these challenges, a dual-branch network that combines a region-constrained Siamese encoder for ventricle segmentation and dynamic snake convolutions to enhance the detection of tubular myocardium structures. The fused outputs enable precise segmentation of cardiac substructures.

A multi-target segmentation network named RCSiamTANet which integrates a region constraint guided Siamese network and a tubular structure ware is proposed. The network consists of two sub-networks, i.e., the left-right ventricle segmentation sub-network and the myocardium segmentation sub-network. The encoder of the left right ventricle segmentation sub-network contains two parallel branches, in which the original image slice encoder uses CNN to extract the original image features, while the region constraint slice encoder leverages the Siamese network to focus on the target region, avoiding redundant feature extraction. The information of the adjacent slice region constraint provided by the Siamese network and the features of the original slices are extracted at the same time to capture more local details for the joint segmentation of the left and right ventricles. The myocardium segmentation sub-network employs dynamic snake convolutions to capture the topological information of the myocardium's tubular structure, improving the perception of the slender tubular structure. Finally, the outputs of the two sub-networks are fused to achieve accurate segmentation of the cardiac substructures.

Compared to state-of-art models such as FCN, UNet, TransUNet, SwinUNet, and EMCAD, RCSiamTANet achieves significance improvements in mean Dice scores on the ACDC dataset, with increases of 4.47%, 3.81%, 2.56%, 4.56%, and 0.84%, respectively. On the M&Ms dataset, RCSiamTANet improves mean Dice scores by 3.44%, 3.49%, 2.76%, 3.37%, and 1.21%, respectively. On the M&Ms-2 dataset, improvements in mean Dice scores are 2.33%, 0.40%, 1.10%, 0.33%, and 0.33%, respectively.

The proposed RCSiamTANet significantly enhances the segmentation performance of complex cardiac structural regions and tubular structures, effectively improving both accuracy and generalization.