EBTNet: Efficient Bilateral Token Mixer Network for Fetal Cardiac Ultrasound Image Segmentation

Fetal cardiac ultrasound apical 4-chamber (A4C) view segmentation with deep learning technique is a crucial auxiliary to diagnosing congenital heart disease. Due to the complexity and variety of morphology in ultrasound images, the accurate segmentation of the A4C view is still a tricky task. Also, the existing methods ignore the correlation among feature maps, resulting in inefficient capture of context information. To address these issues, we propose an Efficient Bilateral Token Mixer Network (EBTNet) with global-local branches, a novel solution for fetal cardiac ultrasound segmentation. Specifically, the global feature branch, with global receptive fields to obtain shallow features, makes the segmentation boundaries smoother and more complete. The local feature branch, with local receptive fields to extract deep feature information, distinguishes similar tissues in ultrasound images well. Furthermore, a novel token mixer called MixFormer is proposed to model the interrelated information of channel tokens and space tokens in the feature maps. We conducted extensive experiments on the CAMUS dataset and the FHSET dataset, demonstrating that our model outperforms existing state-of-the-art deep segmentation models. On these datasets, our model achieved a mean Intersection over Union (mIoU) of 71.78% and 94.07%, Precision scores of 84.98% and 96.95%, mean Pixel Accuracy (pixAccMean) of 98.74% and 98.56%, Recall rates of 80.38% and 96.89%, and Dice coefficients of 81.11% and 96.91%, respectively. Our experimental code is available at https://github.com/KOOKOKOK/EBTNet .