Q347 : Segmentation of thoracic lesions using different types of transformers
Thesis > Central Library of Shahrood University > Computer Engineering > MSc > 2026
Authors:
[Author], Mansoor Fateh[Supervisor]
Abstarct: Abstract Segmenting affected regions in chest radiographs of COVID-19 patients plays a crucial role in developing AI-baxsed medical diagnosis and decision-support systems. Despite the capabilities of U-Net-baxsed networks, these architectures face limitations regarding the simultaneous capture of local and global features, the modeling of long-range dependencies, and the precise delineation of lesion boundaries. This study proposes an enhanced U-shaped network architecture designed to improve the accuracy of pulmonary lesion segmentation in X-ray images; the model incorporates a transformer bottleneck for global dependencies, a dual-branch structure for local and global features, multi-scale processing, a cross-attention mechanism, and guided gating. Gradient clipping is also employed to enhance training stability. The proposed architecture was evaluated on the standard QaTa-COV19 dataset, achieving a Dice similarity coefficient of 88.22%, an IoU of 78.92%, a mean accuracy of 95.45%, and an area under the curve AUC) of 99.22%. Comparisons with existing methods demonstrate the superior performance of the proposed architecture in pulmonary lesion segmentation and its ability to effectively integrate local and global information.
Keywords:
#_Medical image segmentation #COVID-19 #U-shaped neural network #Transformer #cross-attention #uncertainty #deep learning_ Keeping place: Central Library of Shahrood University
Visitor: