AWSnet

AWSnet performs segmentation of myocardial scar and edema in multi-sequence cardiac magnetic resonance (CMR) images using an auto-weighted supervision attention network to improve pathological delineation.

Key Features:

• Auto-Weighted Supervision Framework: Employs reinforcement learning to manage and optimize interactions among multiple supervised layers for improved learning across supervision signals.
• Coarse-to-Fine Segmentation Strategy: Uses a coarse model that identifies the left ventricle myocardial structure as a shape prior followed by a fine model to refine small pathology regions.
• Pixel-wise Attention Integration: Integrates pixel-wise attention mechanisms within the fine segmentation model to precisely extract salient pathological structures.
• Task-Specific Objective: Utilizes task-specific objectives to emphasize relevant features for segmenting myocardial scar and edema in CMR images.

Scientific Applications:

• Myocardial infarction assessment: Segments scar and edema in multi-sequence CMR to support diagnosis and assessment of myocardial pathology, with validation on the Myocardial Pathology Segmentation (MyoPS 2020) dataset reporting superior performance versus other state-of-the-art methods.

Methodology:

Integration of reinforcement learning with a coarse-to-fine segmentation approach using left ventricle shape prior knowledge, auto-weighted supervision across layers, and pixel-wise attention mechanisms.