Approximate Volumetric Reconstruction from Projected Images

Gábor Fichtinger, Sheng Xu, Attila Tanács, Kieran Murphy, L. Myers, Jeffery A. Williams, Mengyan ZhangLecture Notes in Computer Science 2001 · DOI · OpenAlex

Introduction

In medical procedures such as radiosurgery, accurately determining the 3D volume and shape of targets—like arteriovenous malformations (AVMs)—is critical for effective treatment planning. However, clinicians often rely on 2D X-ray angiograms, which present significant challenges for spatial reconstruction. This paper proposes a robust method to estimate 3D volumes from a limited set of 2D projections by combining back-projection and a voxel-based "carving" technique.

The Volumetric Reconstruction Challenge

Determining a precise 3D shape from a small number of 2D projections is mathematically underdetermined. Furthermore, clinical practice introduces subjectivity; surgeons must manually outline targets on 2D images. These silhouettes often fail to align perfectly in 3D space, and existing tools frequently fail to produce a solid volume suitable for dose planning.

A lateral projection of a skull showing a manually outlined target (AVM) used as a 2D silhouette for reconstruction.

Methodology: The "Carving" Approach

The authors address these issues through a two-stage process:

  1. Initial Enclosure: Silhouettes drawn on multiple 2D images are back-projected into 3D space to create a bounding volume guaranteed to contain the target.
  2. The Carving Phase: The algorithm evaluates every voxel within this initial shape. If a voxel's projection falls outside the silhouette in any of the 2D images, it is "carved away." This continues until only the volume consistent with all silhouettes remains.
  3. Consistency Check: To mitigate human error, the final 3D object is re-projected back onto the original 2D planes. Discrepancies between this "shadow" and the original drawings provide a confidence measure, alerting clinicians to potential inaccuracies in their outlines.

Experimental Validation

The method was validated using synthetic datasets, including basic geometric shapes and a complex "tree-branch" phantom designed to mimic the intricate branching of blood vessels.

A complex 3D tree-branch phantom used to validate the algorithm's ability to reconstruct intricate, non-convex structures.

Key Results and Insights

The study revealed several critical findings regarding reconstruction accuracy:

  • High Precision: For the complex tree-branch phantom, the relative error was as low as 0.8% when using seven projections from optimal angles. Even with only three projections, the error remained a respectable 4.3%.
  • Strategic Angles: Accuracy is highly dependent on the direction of X-ray shots. Objects with high aspect ratios are best reconstructed when projections are taken perpendicular to their long axis. Randomly selected angles increased the mean relative error to 11.3%.
  • Resolution: The method can resolve fine details as small as 1 mm using approximately six views with table rotation.

Conclusion and Future Directions

By starting with a bounding block and removing inconsistent voxels, this "carving" method provides a fast, efficient way to generate 3D volumes for clinical use. While the current study relies on synthetic data, it demonstrates the potential for real-time feedback loops that quantify clinical judgment. Future work will focus on clinical validation with real-world angiograms and integrating the method into other X-ray guided treatments like radio-frequency ablation.

Summary updated

BibTeX
@inproceedings{fichtinger2001approximate,
  author = {Gábor Fichtinger and Sheng Xu and Attila Tanács and Kieran Murphy and L. Myers and Jeffery A. Williams and Mengyan Zhang},
  title = {{Approximate Volumetric Reconstruction from Projected Images}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2001},
  doi = {10.1007/3-540-45468-3\_235},
}

Loading related content...

Figure