JAS-GAN: Generative Adversarial Network Based Joint Atrium and Scar Segmentations on Unbalanced Atrial Targets.

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Accurate segmentation of left atrium (LA) and atrial scars from late gadolinium-enhanced cardiac magnetic resonance (LGE CMR) images is crucial for quantifying atrial scar burden.
• Previous methods struggle with segmenting unbalanced atrial targets (LA and scars) due to significant volume differences, often requiring multi-phase approaches.

Purpose of the Study:

• To develop an automated and accurate end-to-end segmentation method for simultaneously segmenting LA and atrial scars from LGE CMR images.
• To address the challenge of segmenting unbalanced atrial targets by proposing a novel generative adversarial network architecture.

Main Methods:

• Introduction of an inter-cascade generative adversarial network, named JAS-GAN, for simultaneous segmentation of LA and atrial scars.
• Implementation of an adaptive attention cascade to model the inclusion relationship between LA and atrial scars, using LA segmentation to guide scar segmentation.
• Application of adversarial regularization for consistent optimization and matching of joint distributions between segmented and real images.

Main Results:

• JAS-GAN demonstrated superior segmentation performance on a 3D LGE CMR dataset (192 scans).
• Achieved high average Dice Similarity Coefficient (DSC) values: 0.946 for LA and 0.821 for atrial scars.
• Outperformed existing state-of-the-art methods in segmenting unbalanced atrial targets.

Conclusions:

• JAS-GAN effectively segments unbalanced atrial targets in LGE CMR images automatically and accurately.
• The proposed adaptive attention cascade and adversarial regularization contribute to improved segmentation accuracy.
• This approach holds significant potential for enhancing atrial scar quantification in clinical practice.