Automatic 3D Whole Heart Registration-Based Segmentation Using Mutual Information and B-Splines

Ken Cai, Rongqian Yang, Lihua Li, Xiao-Ming WuInternational Journal of Advancements in Computing Technology 2011 · DOI

TL;DR

This paper presents a novel, automated 3D method for segmenting the whole heart from CT scans. By combining a coarse affine registration for initial localization with a fine-grained B-spline non-rigid registration guided by Normalized Mutual Information, the approach aims to overcome the limitations of manual segmentation, offering a more consistent and efficient solution for cardiac analysis.

What Problem Does This Paper Address?

Accurate segmentation of the heart from medical images, particularly CT scans, is crucial for understanding its anatomy and function, especially in diagnosing cardiovascular diseases. Traditional segmentation methods often rely on manual or semi-automatic approaches, which are time-consuming, prone to user variability, and can be inconsistent. This paper addresses the need for an automatic, robust, and anatomically driven segmentation technique that can reliably delineate the entire heart structure in 3D CT volumes.

Methods

The proposed method employs a two-step registration-based segmentation strategy:

  1. Coarse Global Affine Registration: This initial step uses an affine transformation to roughly align a pre-segmented heart atlas (template) with the target CT image. The goal is to quickly localize the general area of the heart within the scan, providing a starting point for more detailed registration. This transformation accounts for translation, rotation, scaling, reflection, and shear.
  2. Accurate Non-Rigid B-spline Registration: Following the coarse alignment, a more sophisticated non-rigid registration is performed. This step utilizes cubic B-splines to model complex, local deformations of the heart. The B-spline Free Form Deformation (FFD) framework allows for smooth and continuous transformations. The similarity between the deformed atlas and the target image is measured using Normalized Mutual Information (NMI), a robust metric that is less sensitive to intensity variations and partial overlaps. The optimization process uses an adaptive stochastic gradient descent method.

The novelty lies in the automated and anatomically driven nature of the registration process, where the segmentation problem is framed as finding the optimal transformation to align anatomical structures.

Datasets

The paper states that the method was evaluated on cardiac CT volumes. A "heart atlas" with manually segmented substructures (chambers, ventricles, great vessels) was used as the template. The manual segmentation by a clinician served as the "gold standard" for evaluation. No specific public datasets are mentioned, and no new benchmark datasets were introduced.

Key Insights and Ideas

  • Registration as Segmentation: The core idea is reframing the segmentation problem as a registration task. By aligning a detailed anatomical atlas to a target image, the atlas's segmentation labels can be transferred to the target.
  • Two-Stage Registration for Efficiency and Accuracy: Combining a fast affine registration for global localization with a flexible B-spline registration for local refinement offers a balance between computational efficiency and the ability to capture complex anatomical shapes.
  • NMI's Robustness: The use of Normalized Mutual Information is highlighted as a key factor in achieving robust segmentation, particularly in the presence of noise, artifacts, and intensity variations common in CT images.
  • Automation: The method's primary advantage is its automation, eliminating the need for manual interaction and thus reducing variability and time.

Key Results

The paper demonstrates the method's performance on cardiac CT volumes. Figure 2 shows example segmentation results, visually indicating successful delineation of the whole heart. The authors state that their method provides "good segmentation for whole heart" and is "valid and is robust to noise." Quantitative results (e.g., Dice scores, accuracy metrics) are not explicitly detailed in the provided text.

Limitations and Future Work

  • FFD Mesh Limitations: The authors acknowledge that the FFD registration, while effective, might lead to diminishing modeling ability of the FFD meshes.
  • Future Work:
    • Evaluating the method on more clinical scans with different imaging modalities and a wider range of anatomical structures.
  • Testing a "free-form deformations registration using adaptive control point status" to potentially improve the modeling capabilities of the deformation meshes.

Summary updated

Abstract

Abstract Extracting the anatomical information of whole heart to computed tomography (CT) images is important in both anatomic and functional studies of human heart. The most recent advances in image segmentation adopt an interactive approach. That approach is prone inconsistency and variability, such as, laborious, time-consuming and subjective process. To avoid this, an automatic technique for whole heart segmentation of cardiac CT volumes is highly desired. In this paper, we propose a novel three-dimensional (3D) automatic anatomy segmentation method for the whole heart, in which, process of the registration divides into two steps, coarse registration and accurate registration. The coarse registration using affine transformation is first used to localize the heart and the accurate registration uses cubic B-splines to parameterize the deformation field, and Normalized Mutual Information (NMI) as the similarity measure. The novelty of our method is that the registration procedure is automated and anatomically driven. Exploiting similarity structures of different individuals, the segmentation problem is converted to a registration problem. Test results show that our method is valid and is robust to noise by providing segmentation results on real heart CT images.
BibTeX
@article{cai2011automatic,
  author = {Ken Cai and Rongqian Yang and Lihua Li and Xiao-Ming Wu},
  title = {{Automatic 3D Whole Heart Registration-Based Segmentation Using Mutual Information and B-Splines}},
  journal = {International Journal of Advancements in Computing Technology},
  year = {2011},
  doi = {10.4156/ijact.vol3.issue11.1},
}
Figure