tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Multi-View 2.5D Attention U-Net with 3D Fusion for Efficient Stroke Lesion Segmentation from T1-Weighted MRI

Authors :

Fatemeh Salahshourinejad¹ Kamran Kazemi² Negar Noorizadeh³ Mohammad Sadegh Helfroush⁴ Ardalan Aarabi⁵

1- دانشگاه صنعتی شیراز 2- دانشگاه صنعتی شیراز 3- University of Tennessee Health Science Center 4- دانشگاه صنعتی شیراز 5- University of Picardy Jules Verne

Keywords :

Stroke Lesion،MRI،2.5D Segmentation،Deep Learning،Attention mechanism،Multi-view

Abstract :

Accurate segmentation of stroke lesions from Magnetic Resonance Imaging (MRI) scans is critical for clinical decision-making and patient prognosis. However, stroke lesion segmentation from mono-spectral MRI such as T1-weighted (T1w) images suffers from similar gray level characteristics of brain tissues and heterogeneity of lesion properties (e.g., shape and size). Deep learning has become the standard approach for medical image segmentation; however, 2D models lose inter-slice context and 3D models face computational complexity. In this paper, we proposed a multi-view 2.5D model that employed three 2D U-Nets for intra-slice lesion segmentation across axial, coronal, and sagittal views, followed by a 3D convolutional neural network (CNN) to integrate the outputs. The 2D U-Nets incorporated residual blocks in the encoder–decoder and attention blocks in the skip connections, while the 3D CNN with attention mechanisms produced the final segmentation. The proposed model was evaluated on the ATLAS V2.0 dataset for stroke lesion segmentation, achieving a mean Dice score of 0.64±0.27outperforming 2D approaches and comparable with 3D models, while requiring fewer parameters, making it practical for resource-constrained settings.