DeformFlowNet: A deformable transformer with optical flow constraint for interventricular septum segmentation in echocardiography videos

Qihui Guo, Zhou Xu, Weiwei Tan, Yiwen Shen, Shupeng Liu, Bo Zhang, Qi Zhang

Abstract: Transthoracic echocardiography (TTE) is often used to assess left ventricular hypertrophy (LVH) but is limited by specialist availability. An automated system for accurate LVH diagnosis is needed, focusing on segmenting the interventricular septum from TTE videos. In this study, we propose a hybrid network that uses U-shaped attention networks and optical flow constraints for interventricular septum segmentation in TTE videos, named the DeformFlowNet. The encoder is composed of the deformable transformer, which is integrated with the skip connection of the U-shaped network, and it balances computational and spatial complexities while modeling long-range dependencies of features. The optical flow information between frames in the TTE videos is mapped to the corresponding mask outputs of the decoder, establishing long-range relationships across successive frames. By enforcing optical flow consistency constraints, the segmentation results are further refined and improved. Enhanced by a mixed loss function supervision, the model achieves notable accuracy in interventricular septum segmentation with a Dice similarity coefficient of 0.914 and a Jaccard coefficient of 0.842. Our DeformFlowNet outperforms other advanced 3D segmentation networks, offering a reliable foundation for subsequent LVH etiological diagnosis. Code is available on-line at https://github.com/treyguo15/DeformFlowNet.

External IDs:dblp:journals/bspc/GuoXTSLZZ25