[Paper Review] 3D Path Planning from a Single 2D Fluoroscopic Image for Robot Assisted Fenestrated Endovascular Aortic Repair

Jian-Qing Zheng, Xiao-Yun Zhou+2 more|arXiv (Cornell University)|Sep 16, 2018

Aortic aneurysm repair treatments11 references3 citations

TL;DR

This paper proposes a real-time 3D robotic path planning framework for robot-assisted fenestrated endovascular aortic repair (FEVAR) using a single 2D fluoroscopic image. It employs graph matching to align preoperative 3D aortic aneurysm (AAA) skeletons with intraoperative 2D projections, followed by non-rigid skeleton deformation via thin-plate splines with length and smoothness regularization, achieving 2 mm 3D distance error in phantom validation and outperforming ICP-TPS in accuracy, robustness, and speed.

ABSTRACT

The current standard of intra-operative navigation during Fenestrated Endovascular Aortic Repair (FEVAR) calls for need of 3D alignments between inserted devices and aortic branches. The navigation commonly via 2D fluoroscopic images, lacks anatomical information, resulting in longer operation hours and radiation exposure. In this paper, a framework for real-time 3D robotic path planning from a single 2D fluoroscopic image of Abdominal Aortic Aneurysm (AAA) is introduced. A graph matching method is proposed to establish the correspondence between the 3D preoperative and 2D intra-operative AAA skeletons, and then the two skeletons are registered by skeleton deformation and regularization in respect to skeleton length and smoothness. Furthermore, deep learning was used to segment 3D pre-operative AAA from Computed Tomography (CT) scans to facilitate the framework automation. Simulation, phantom and patient AAA data sets have been used to validate the proposed framework. 3D distance error of 2mm was achieved in the phantom setup. Performance advantages were also achieved in terms of accuracy, robustness and time-efficiency. All the code will be open source.

Motivation & Objective

• To address the lack of 3D anatomical context in 2D fluoroscopic navigation during FEVAR, which leads to prolonged radiation exposure and longer operation times.
• To enable real-time 3D robotic path planning using only a single intraoperative 2D fluoroscopic image.
• To improve accuracy and efficiency of 3D AAA shape reconstruction by integrating deep learning for preoperative segmentation and graph-based skeleton correspondence.
• To reduce reliance on multiple fluoroscopic images or invasive landmarks by enabling 3D path reconstruction from a single 2D projection.

Proposed method

• A 6-block U-Net with cross-entropy loss and stochastic gradient descent is used to automatically segment the main aortic aneurysm from preoperative CT scans.
• Preoperative 3D and intraoperative 2D AAA skeletons are extracted using regionprops and bwareopen for 3D, and MATLAB skeletonization for 2D.
• Graph matching is applied to establish topological correspondence between 3D preoperative and 2D intraoperative AAA skeletons based on structural similarity.
• Non-rigid registration is performed via thin-plate spline (TPS) deformation of the 3D skeleton, regularized by length and smoothness constraints to match the 2D projection.
• The framework integrates deep learning for automated preoperative segmentation, reducing manual intervention and enhancing pipeline automation.
• Validation is performed on simulation, phantom, and patient datasets using 2D/3D distance errors and computational time as performance metrics.

Experimental results

Research questions

• RQ1 Can a single 2D fluoroscopic image be used to reconstruct a 3D path for robotic navigation in FEVAR with sufficient accuracy and speed?
• RQ2 How can topological correspondence between 3D preoperative and 2D intraoperative AAA skeletons be established robustly despite projection ambiguity?
• RQ3 What is the performance gain of graph matching and TPS-based deformation over traditional ICP-TPS in 2D/3D non-rigid registration for AAA?
• RQ4 To what extent can deep learning automate preoperative AAA segmentation to support real-time 3D path planning?
• RQ5 Can the proposed framework achieve sub-second runtime for real-time clinical deployment?

Key findings

• The proposed method achieved a 3D distance error of 2 mm in the phantom setup, significantly outperforming the ICP-TPS baseline.
• In patient data, the proposed method reduced mean 2D distance error to 0.1222 pixels (std: 0.1455), compared to 0.7061 pixels (std: 3.0812) for ICP-TPS.
• The framework achieved a total runtime of approximately 0.9 seconds for 2D skeletonization and 0.8 seconds for non-rigid registration, compared to over 10 minutes for ICP-TPS.
• The method demonstrated robustness to frontal-plane translations and rotations, with consistent performance across patient cases.
• The use of deep learning for preoperative AAA segmentation enabled automated pipeline operation, though manual correction was still needed for renal and iliac branches.
• The framework outperformed ICP-TPS in both accuracy and time-efficiency, with lower 2D and 3D distance errors and standard deviations across all datasets.