Projects Last Projects
Optimal segmentation of devices and vessels on 2D fluoroscopy PCI images
Noam Bitton and Mayan Rousso
Supervised by Daniel Katz and Asi Elad
Coronary heart disease remains a leading cause of mortality worldwide, highlighting the need for accurate cardiovascular diagnosis and prevention. This study focuses on benchmarking the performance of nnUNet neural networks for blood vessel segmentation in coronary angiography videos using a leave-one-out (LOO) methodology. By evaluating the models' generalization across different patients, scenes, and frames, we identify instances of suboptimal performance and propose database enhancements to address these gaps. Root causes of lower segmentation accuracy are analyzed, and targeted solutions are introduced to improve model performance. The proposed method achieves performance comparable to the original model while potentially enhancing accuracy in critical ROIs, offering valuable insights to refine cardiovascular image segmentation. These advancements support safer and more effective catheterization procedures, aligning with CathAlert's mission to prevent patient harm.
Please, see project report.
Please, see final presentation.