Assessment of Transformer-Based Models for Cardiac Systolic Abnormality Segmentation in Cardiac Catheterization X-Ray Images

This study aims to evaluate the performance of transformer-based models learned on segmenting cardiac systolic abnormalities using a dataset comprising 1717 cardiac catheterization X-ray images. This study evaluated the segmentation performance of seven transformer architectures, including Swin-UNet, MaskFormer, AgileFormer, Swin Transformer, DETR, MaxViT, and SETR. Swin-UNet and MaskFormer had the best dice coefficients and accuracy among all models tested. Swin-UNet achieved an accuracy of $\mathbf{9 9. 3 1 \%}$ and a dice coefficient of $\mathbf{9 8. 5 4 \%}$, and MaskFormer achieved an accuracy of $\mathbf{9 8. 4 9 \%}$ and a dice coefficient of $99.78 \%$, indicating their high segmentation performance. The results thus show the substantial benefits transformer-based models offer in medical imaging, with improvements in diagnostic accuracy and potential for clinical practice. The research allows future studies to develop these models and examine their wider cardiological use.