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

Background: Early detection and quantification of the irreversibly damaged core and the potentially salvageable penumbra regions are crucial for the diagnosis and treatment of acute ischemic stroke (AIS). However, accurately segmenting different pathological structures from computed tomography perfusion (CTP) images is challenging due to variations in multi-lesion features and the complexity of pathological structures.

Purpose: This study aims to utilize prompt-enhanced multi-task learning to provide multiple prompt information for different pathological structures segmentation, enabling precise quantification of the core and penumbra regions in CTP images, thereby promoting clinical resource allocation and optimizing treatment decisions.

Methods: We propose PEMT-Net, a multi-task learning framework that integrates dual prompt information for multi-class segmentation of core and penumbra regions in CTP images. Category-specific prompts are learned through separate classification networks, while instance-specific prompts, based on class activation maps (CAMs), provide region-level guidance. The segmentation network uses an encoder-decoder structure with an attentional adaptive context extraction (AACE) module to capture multi-lesion features. A dual-decoder structure with prompt parsers refines the relationship between lesion and prompt information, and supervised loss is applied to optimize segmentation. A private CTP dataset for AIS, referred to as GLis, was collected to evaluate the model's performance.

Results: Our method achieved the highest DC values in both categories on the public SPES dataset, with improvements of 0.8% and 0.5% over the second-best method. On the GLis dataset, it showed improvements of 2.6% and 1.1%. Qualitative results and statistical analyses (Pearson correlation and Bland-Altman) demonstrated good consistency.

Conclusions: The proposed PEMT-Net framework validates the effectiveness of category-specific and instance-specific prompts, with the AACE module enhancing lesion feature extraction. Experimental results demonstrate improved segmentation performance of the core and penumbra regions, highlighting its clinical potential for AIS diagnosis and treatment.