Synthetic Data Generation and Deep Learning-Based Object Detection and Segmentation for Interventional Devices in Cardiac and Neurovascular Fluoroscopy – MT Intro Talk by Mersad Shoaei Taklimi

Join us for the introductory talk of a master’s thesis on synthetic data generation and deep learning-based detection and segmentation of interventional devices in cardiac and neurovascular X-ray fluoroscopy. AI approaches for fluoroscopy-guided interventions require large well-annotated datasets. However, access to such data is limited due to privacy constraints, cost of annotation, and the lack of public datasets. This thesis aims to address these challenges by generating simulation-based datasets to support deep learning approaches.

The work focuses on creating diverse and realistic fluoroscopy images of interventional devices and training on different deep learning models for multi-device detection and segmentation. The simulation of X-ray images is based on physics-based digitally reconstructed radiographs, and the placement of the devices is guided by realistic vascular anatomy. The goal is to investigate how effectively models trained on synthetic data generalize to real clinical fluoroscopy images.