Background

X-rays are one of the most common imaging modalities used in modern medicine. It is a quick and painless imaging procedure used for a wide variety of diagnostic purposes. During an X-ray, electromagnetic waves are sent through the patient and absorbed at rates specific to the physical properties of different tissues (radiodensity). This projects the radiodensity of the patient's volume on the exposed photographic film, creating a 2D image.

Computed tomography (CT) imaging, on the other hand, uses rotating X-ray source and detector to produce a series of cross-sectional images of the body at a defined slice interval. The resulting images can be stacked together to produce a 3D volume of radiodensities.

CT scans can provide detailed three-dimensional information of the human body, eliminating problem of overlapping structures in X-rays. However, this extra dimensionality comes at both a monetary and safety cost. CT imaging is more expensive and exposes patient to an order of magnitude greater level of radiation than X-ray imaging. For these reasons, X-rays are preferred to CTs in cases where the extra dimensionality doesn't provide a large enough clinical benefit to offset the costs.

When radiologists interpret X-rays, they use their internal knowledge of 3D human anatomy to guide their interpretation. They are in essence combining image inputs and a priori knowledge of human anatomy to formulate their diagnosis. Intuition would suggest that providing some kind of 3D understanding of a 2D X-ray film (via the embeddings from a 3D reconstruction process of that X-ray film) could improve accuracy on standard machine learning image tasks like recognition and segmentation on 2D X-rays. While testing this hypothesis is outside the scope of our project, it is a potential avenue for future research.

It is possible that an "accurate enough" 3D reconstruction of X-rays could be used instead of a CT in certain clinical scenarios, which could also allow patient to avoid unnecessary exposure to CT-sourced radiation.

State-of-the-art

Neural network-based 3D reconstruction

To our best knowledge, no literature exists for 2D-to-3D reconstruction of chest X-rays and chest CTs. However, there exist several works that can perform 2D-to-3D reconstruction in somewhat analogous domains.

  • Henzler et al. (2018) has performed 2D-to-3D volume reconstruction of mammalian crania using a Convolutional Neural Network (CNN) with an encoder-decoder structure. However, they focused on optimizing surface structure and the generalizability of the algorithm (reconstruction any angle, wide variety of species).
  • Jackson et al. (2017) has proposed a Volumetric Regression CNN (VRN) built upon paired 2D images and 3D facial models or scans. Their model is able to reconstruct 3D human faces from a single photograph taken at any angle.
  • Karade and Ravi (2015) used a “bone template reconfiguration” algorithm involving Kohonen self-organizing maps to simulate 3D surface geometry of femur from biplane X-ray images.

Synthetic X-ray generation

Since we are interested in 2D-to-3D reconstruction, our training datasets, consisting of 3D CT scans, would also need an equivalent of the input X-ray images.

We have decided to generate synthetic X-rays for each of the 3D CT scans in order to ensure matching pairs of 3D-2D dataset. The following are resources we have used to make this decision:

  • Moturu and Chang (2019) proposed a method to create synthetic frontal chest X-rays using ray-tracing and Beer’s Law from chest CT scans.
    • Their research focused on designing a neural network that can detect lung cancer, thus this method also involved randomized nodule generation.
    • However, the authors eventually decided to focus on generating synthetic images from a parallel perspective instead using the point-source approach, which will be closer to actual 2D projection process in generating X-ray images.
  • Teixeira et al. (2018) also presented a framework that generates synthetic X-ray images from the surface geometry of the thorax and abdomen.
    • The resulting X-ray images from this study were only intended to serve as an approximation of the true internal anatomy.
    • Aside from these, we have not found any other papers that addressed the generation of synthetic X-ray images from CT scans specifically in the chest region.
  • Henzler et al. (2018) also generated synthetic x-rays, but their algorithm simply flattened the CT scans at various angles and did not consider the point-source aspect of X-ray imaging. The only paper we found that included the point-source aspect of X-ray imaging was that of Moturu and Chang.

Dataset

The dataset used in this project is the Lung Image Database Consortium image collection (LIDC-IDRI), available at The Cancer Imaging Archive (TCIA).

The database contains a total of 1010 patients.