Prompt-enhanced multi-task learning network for acute ischemic stroke lesion and penumbra segmentation using computed...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Accurate segmentation of core and penumbra in computed tomography perfusion (CTP) images is vital for acute ischemic stroke (AIS) diagnosis and treatment.
• Challenges in CTP image segmentation arise from complex pathological structures and varied multi-lesion features.

Purpose of the Study:

• To develop a prompt-enhanced multi-task learning framework for precise segmentation of core and penumbra regions in CTP images.
• To improve quantification of ischemic stroke regions for better clinical resource allocation and treatment optimization.

Main Methods:

• Proposed PEMT-Net, a multi-task learning framework integrating dual prompt information (category-specific and instance-specific) for multi-class segmentation.
• Utilized an encoder-decoder structure with an attentional adaptive context extraction (AACE) module and a dual-decoder structure with prompt parsers.
• Evaluated on a private CTP dataset (GLis) for acute ischemic stroke.

Main Results:

• Achieved superior Dice coefficient (DC) values on both public (SPES) and private (GLis) datasets compared to existing methods.
• Demonstrated significant performance improvements on the GLis dataset (2.6% and 1.1% increases).
• Qualitative and statistical analyses confirmed good consistency and segmentation accuracy.

Conclusions:

• PEMT-Net effectively leverages category-specific and instance-specific prompts for improved segmentation of AIS core and penumbra regions.
• The AACE module enhances lesion feature extraction, contributing to superior performance.
• The framework shows significant clinical potential for enhancing AIS diagnosis and guiding treatment strategies.