Аңдатпа
Accurate segmentation of brain stroke lesions in magnetic resonance imaging (MRI) is essential for guiding clinical decision-making yet remains challenging due to heterogeneous lesion appearance and limited annotated data. We propose PA-Swin-AMH, a novel framework that integrates pseudo-annotation guidance with a hierarchical Swin Transformer backbone and an auxiliary mask head for multi-scale supervision. Coarse pseudo-labels, generated from bounding-box predictions, are incorporated during training to expand the supervisory signal without additional manual delineation. The auxiliary mask head enforces pixel-level refinement at intermediate feature maps, combining Dice and binary cross-entropy losses to sharpen lesion boundaries. On a publicly available stroke MRI dataset comprising 9 600 annotated slices, PA-Swin-AMH achieved a mean Dice Similarity Coefficient of 0.87 and a 95th percentile Hausdorff Distance of 5.2 mm, outperforming both CNN-based and standard Swin Transformer baselines by 6 % and 12 % on these metrics, respectively. Classification precision and recall exceeded 0.91, and localization metrics (IoU and MAE) demonstrated tight alignment with expert annotations. Qualitative results confirm robust delineation across varied lesion sizes and contrasts. These findings indicate that pseudo-annotation guidance and auxiliary supervision synergistically enhance transformer-based segmentation, offering a scalable solution for clinical deployment and reducing annotation burden in neuroimaging studies.