AGMR-Net: Attention-guided multiscale recovery framework for stroke segmentation.

Area of Science:

• Medical imaging
• Artificial intelligence
• Neurology

Background:

• Accurate stroke lesion segmentation is crucial for clinical assessment and diagnostic systems.
• Existing methods struggle with intraclass inconsistency and interclass indistinction in brain tissue segmentation.
• These challenges limit the clinical adoption of automated stroke segmentation tools.

Purpose of the Study:

• To propose a novel attention-guided multiscale recovery framework (AGMR-Net) for stroke lesion segmentation.
• To overcome limitations of existing methods in handling lesion variability and distinguishing lesions from normal tissue.
• To improve the accuracy and reliability of automated stroke diagnosis systems.

Main Methods:

• Developed AGMR-Net featuring a coarse-grained patch attention (CPA) module to mitigate intraclass inconsistency.
• Introduced a cross-dimensional feature fusion (CFF) module to enhance boundary delineation by integrating 2D and 3D features.
• Employed a multiscale deconvolution upsampling (MDU) module for improved recovery of spatial and boundary information during decoding.

Main Results:

• AGMR-Net achieved a Dice similarity coefficient of 0.594 on the Anatomical Tracings of Lesions After Stroke dataset.
• The method demonstrated a Hausdorff distance of 27.005 mm and an average symmetry surface distance of 7.137 mm.
• Performance metrics indicate superior results compared to state-of-the-art methods in stroke segmentation.

Conclusions:

• AGMR-Net effectively addresses key challenges in stroke lesion segmentation, namely intraclass inconsistency and interclass indistinction.
• The proposed framework shows significant potential for advancing automated stroke diagnosis.
• The method's superior performance suggests its clinical utility in estimating lesion status in stroke patients.