Comparative evaluation of deep learning architectures, including UNet, TransUNet, and MIST, for left atrium segmentation in cardiac computed tomography of congenital heart diseases

Figures

• 

  Fig. 1.. Model architectures. (A) UNet (Ronneberger et al. [1], 2015), (B) TransUNet (Chen et al. [2], 2021), and (C) MIST (Rahman et al. [3], 2023). A brief overview explains the evolution of transformer and UNet hybrid models. The UNet architecture introduced symmetric encoder-decoder paths with skip connections. TransUNet is built upon UNet by adding transformer attention mechanisms to the encoder, and MIST, a recent model, further integrates multiple attention mechanisms and optimized skip connections to improve performance.

• 

  Fig. 2.. Qualitative segmentation outcomes across model architectures. The ground truth (blue) is presented in the second column, while the predicted segmentation (green) is shown in the left column. (A) UNet, (B) TransUNet, and (C) MIST. Detailed results, including dice scores and 95th percentile of the Hausdorff distance (HD95) values for individual computed tomography scans, are provided in Supplements 3 and 4.

• 

  Fig. 3.. Dice scores for 11 left atrial cardiac computed tomography images segmented by UNet, TransUNet, and MIST models.

• 

  Fig. 4.. 95th percentile of the Hausdorff distance (HD95) values for 11 left atrial cardiac computed tomography images segmented by UNet, TransUNet, and MIST models.

• 

  Fig. 5.. Dice scores for 11 left atrial cardiac computed tomography images segmented by MIST with or without spatial squeeze-and-excitation attention module (SSAM) or sum output heads.

• 

  Fig. 6.. 95th percentile of the Hausdorff distance (HD95) values for 11 left atrial cardiac computed tomography images segmented by MIST with or without spatial squeeze-and-excitation attention module (SSAM) or sum output heads.

• 

  Fig. 7.. Qualitative outcomes for MIST variants. The ground truth (blue) is presented in the second column, while the predicted segmentation (green) is shown in the left column. Starting from left: spatial squeeze-and-excitation attention module (SSAM) without summation (A), SSAM with summation (B), and no attention mixer within the decoder and no summation (C).