Skip to main content

Usformer: A Light Neural Network for Left Atrium Segmentation of 3D LGE MRI

Team: Hui Lin, Santiago Lopez Tapia, Florian Schiffers, Yunan Wu, Huili Yang, Nikolay Iakovlev, Bradley D. Allen, Ryan Avery, Daniel C. Lee, Daniel Kim, Aggelos K. Katsaggelos

Video link:

https://drive.google.com/file/d/1zgG7kxaEWV6zzBZfw8DQ0RaugzH9lmrD/view?usp=sharing

Text

Left atrial fibrosis is an important mediator of atrial fibrillation and atrial myopathy. Late gadolinium-enhancement (LGE) MRI is a proven non-invasive test for the evaluation of left atrial (LA) fibrosis. However, manual segmentation is labor-intensive. Automatic segmentation is challenging due to varying intensities of data acquired by different vendors, low contrast between the LA and surrounding tissues, and complex LA shapes. Current approaches based on 3D networks are computationally expensive and time-consuming due to the large size of 3D LGE MRIs and networks. To address this, most approaches use two-stage methods to first locate the LA center using a down-scaled version of the MRIs and then crop the full-resolution MRIs around the LA center for final segmentation. We propose a light transformer-based model to accurately segment LA volume in one stage, avoiding errors introduced by suboptimal two-stage training. Transposed attention in transformer blocks can capture long-range dependencies among pixels in large 3D volumes without significant computation requirements. Our proposed model achieved a promising dice similarity coefficient of 92.6% in the 2018 Atrial Segmentation Challenge, with only 611k parameters, which is about 1% of the method ranked 3rd in the challenge but with similar performance.

Figure 1. The proposed Usformer for LA segmentation. Usformer incorporates a U-Net architecture with efficient transposed transformer blocks that significantly decrease computation complexity.
Figure 2. Comparative results of the proposed Usformer, nnUnet, and on the testing set in the 2018 atrial segmentation challenge.
Figure 3. Axial view of LA segmentation results by our proposed method Usformer, nnUNet, and UNeXt. Contours of manual and predicted segmentation are denoted in red and green.
Figure 4. 3D visualization of the two worst and best LA segmentation results achieved by the proposed method in terms of 3D dice score. The color on the surface indicates the distance from the prediction to the manual segmentation. The surface distances are scaled between 0 and 10 mm for better visualization. Arrows (1-3) highlight the errors in MV, the regions between LA and RA, and PV, respectively.