Abstract

Cardiac catheterization is a procedure in minimally invasive cardiovascular interventions, where
safe and effective performance relies on accurate force awareness, catheter visualization, and navigation
within complex vascular geometries. This doctoral research investigates learning-based,
sensor-free frameworks using standard clinical imaging modalities, with the objective of advancing
toward autonomous catheter-based systems. The work is conducted as a multidisciplinary study
at the intersection of artificial intelligence, deep learning–based computer vision, medical robotics,
mechatronics, and autonomous systems, based on the premise that catheter deflections observed
from multiple imaging viewpoints encode sufficient information to infer three-dimensional contact
forces and support goal-directed navigation. A lightweight fusion-based convolutional neural network
is first proposed to estimate three-dimensional contact forces directly from stereo catheter
images without reliance on physical sensors or explicit mechanical modeling. This is extended
to a multitask encoder–decoder architecture that simultaneously performs catheter segmentation
and force estimation from biplane fluoroscopic images within a single end-to-end framework, supported
by a synthetic X-ray image generator designed to resemble clinical fluoroscopy. The framework
is further extended through a Vision Transformer–based architecture with cross-attention for
joint stereo segmentation and force estimation. Finally, autonomous catheter navigation is explored
through a goal-conditioned, multimodal vision-to-action model trained using imitation learning for
perception-driven catheter steering in a physical robotic setup. Overall, this dissertation demonstrates
that force estimation, catheter perception, and navigation can be learned within sensor-free
frameworks, supporting future autonomous robotic catheterization systems.