Generating X-ray Images from Point Clouds Using Conditional Generative Adversarial Networks.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Simulating medical images like X-rays is crucial for reducing radiation exposure during non-diagnostic visualization.
• Existing methods, such as ray tracing, require 3D models, limiting their application with point cloud data from depth cameras and sensors.

Purpose of the Study:

• To develop and evaluate a method for generating X-ray images directly from generic 3D point clouds.
• To assess the impact of point cloud density on the performance of the proposed image translation method.

Main Methods:

• A conditional generative adversarial network (CGAN) pix2pix model was employed for image translation.
• A custom synthetic data generator was used to create a dataset of point clouds and corresponding X-ray images for training.
• Experiments were conducted using point clouds of varying densities to analyze their effect on translation quality.

Main Results:

• The CGAN successfully predicted X-ray images from point cloud data, demonstrating the feasibility of the approach.
• Higher point cloud densities generally yielded better results compared to lower densities.
• No significant performance difference was observed between networks trained with high-density and medium-density point clouds.

Conclusions:

• Conditional generative adversarial networks (CGANs) are applicable to medical image translation tasks, particularly when 3D models are unavailable.
• The proposed method offers a viable alternative for simulating X-ray images using point cloud data.
• Future research should focus on addressing limitations such as occlusion and image quality, and exploring CGANs for other medical image translation applications.