A Novel 3D Reconstruction Algorithm of Motion-Blurred CT Image

Jing Zhang, Guo Qiang, Han Fang, Li Zhan-Li, Li Hong-An, Sun YuComput. Math. Methods Medicine 2020 · DOI

TL;DR

This paper introduces a two-pronged approach to improve 3D CT reconstructions: first, a Generative Adversarial Network (GAN) for deblurring motion-induced artifacts in CT images, and second, a novel Marching Cubes (MC) algorithm (GI-MC) that fuses golden section search and isosurface direction smoothing for more accurate and efficient 3D model generation. The combined method significantly enhances image clarity and reconstruction accuracy, crucial for better medical diagnosis and surgical planning.

What Problem Does This Paper Address?

Computed Tomography (CT) is a vital medical imaging technology, but its utility is hampered by motion artifacts. Patient movement, whether voluntary (e.g., children, elderly) or involuntary (e.g., Parkinson's disease), causes blurring in CT images. This blurring degrades image quality, making it difficult for clinicians to accurately assess lesion size and shape from 2D slices. While 3D visualization can offer a more comprehensive view, motion blur directly impacts the fidelity of these reconstructions. Existing motion compensation techniques often rely on device tracking or multiframe merging, which can be impractical or fail to preserve fine details. This paper aims to address the challenge of motion-induced blurring in CT images and improve the accuracy and efficiency of subsequent 3D reconstructions.

Methods

The paper proposes a two-stage solution:

  1. GAN Image Translation Network for Deblurring: This approach treats deblurring as an image-to-image translation task. A U-Net architecture is employed for the generator, incorporating skip connections to preserve spatial information. The network is trained using a combination of reconstruction loss (L2 norm for high-frequency details) and adversarial loss to ensure generated images are both accurate and realistic. This method aims to remove motion blur without requiring time-sequential CT images, thus preserving texture details.
  2. GI-MC Algorithm for 3D Reconstruction: This is a novel Marching Cubes (MC) algorithm designed to overcome the limitations of traditional MC, such as low computational efficiency and nonsmooth reconstructed surfaces.
  3. Golden Section Algorithm: This is used to calculate equivalent points and normal vectors. By employing golden section points, the number of calculations for these parameters is reduced from four to one, significantly speeding up the process.
  4. Isosurface Direction Smooth Algorithm: This component addresses the "squamous effect" often seen in traditional MC reconstructions due to discontinuous normal vectors. It smooths the direction of all triangular patches constituting the isosurface by computing the mean value of neighboring normal vectors, resulting in a smoother and more visually appealing 3D model.

Datasets

The performance evaluation utilizes the 3D-IRCADb-01 database, which contains 3D CT scans of 20 patients (10 men, 10 women) with liver tumors. The database comprises approximately 100 images per patient in DICOM format. The authors use examples from this database, specifically focusing on a liver tumor patient and a fatty liver patient, for both deblurring and 3D reconstruction experiments.

Key Insights and Ideas

  • Deblurring as Image Translation: Framing motion blur removal as an image-to-image translation problem solvable by GANs is a powerful approach, allowing for effective deblurring without needing sequential data.
  • Golden Section for Efficiency: Applying the golden section search to calculate intersection points and normal vectors in the MC algorithm is a clever way to reduce computational load and improve reconstruction speed.
  • Isosurface Smoothing for Realism: The isosurface direction smoothing technique directly tackles the visual artifacts of traditional MC, leading to more aesthetically pleasing and potentially more diagnostically useful 3D models.
  • Synergistic Approach: Combining advanced deblurring with an optimized 3D reconstruction algorithm offers a comprehensive solution to the challenges posed by motion-blurred CT data.

Key Results

  • Deblurring Performance: The GAN deblurring algorithm demonstrated superior restoration capabilities compared to other algorithms (Sun et al., Zhang et al., Wang et al., Hernandez et al.) across various blurred angles and amplitudes, as evidenced by Shannon entropy ratio and PSNR metrics.
  • 3D Reconstruction Accuracy: The GI-MC algorithm achieved higher reconstruction accuracy for liver CT images compared to traditional MC, Li's algorithm, and Pratomo's algorithm. Specifically, it showed an average improvement of 9.9%, 7.7%, and 3.9% respectively.
  • 3D Reconstruction Speed: The GI-MC algorithm significantly reduced the time required for 3D reconstruction, offering a substantial advantage over traditional MC, Li's algorithm, and Pratomo's algorithm. For instance, for liver tumor patients, GI-MC took 65.93s compared to 266.97s for traditional MC.

Limitations and Future Work

  • Manual Segmentation Dependency: The paper notes that manual segmentation of the liver region, a prerequisite for 3D reconstruction, relies on the doctor's experience and can lead to variations. The authors suggest that image processing and 3D reconstruction technologies could improve the accuracy and consistency of this segmentation.
  • Volume Calculation: While 3D reconstruction is discussed, the paper highlights that accurately determining organ volume from 2D CT images is currently a rough estimation. Future work could focus on integrating accurate volume calculation into the 3D reconstruction process, which would be highly beneficial for clinical diagnosis and treatment planning.
  • Generalizability of Deblurring: While the GAN deblurring shows promise, further exploration of its performance on a wider range of motion artifacts and CT imaging scenarios could be beneficial.

Summary updated

BibTeX
@article{zhang2020novel,
  author = {Jing Zhang and Guo Qiang and Han Fang and Li Zhan-Li and Li Hong-An and Sun Yu},
  title = {{A Novel 3D Reconstruction Algorithm of Motion-Blurred CT Image}},
  journal = {Comput. Math. Methods Medicine},
  year = {2020},
  doi = {10.1155/2020/9324689},
}
Figure