CycleGAN Deformable Layers for CBCT Artefact Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods lack a simple and efficient way to suppress and remove artefacts and noise in Cone-Beam Computed Tomography (CBCT) images, which can impact adaptive treatment planning and diagnosis, and existing techniques do not effectively decouple the effects of adversarial losses from preserving original structures in enhanced imaging.
Innovation Solution
The use of generative adversarial networks (GANs) and cycle generative adversarial networks (CycleGANs) with deformable offset layers to generate synthetic CT images from CBCT images, where the networks are trained to minimize pixel-based loss terms and apply adversarial losses only to deformed images, preserving original anatomy structures and reducing artefacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional CBCT imaging is used, then the imaging process is simple and fast, but the image quality is low with many artefacts and noise
Solution Approach 1:
A generative adversarial network (GAN) is introduced as an intermediary processing system between the CBCT imaging device and the final image output. The GAN comprises a generator that creates enhanced images and a discriminator that evaluates them, working together to transform low-quality CBCT images into high-quality images that resemble real CT scans while preserving anatomical accuracy
Solution Approach 2:
The patent replaces traditional image processing algorithms with a deep learning-based GAN system. Instead of using conventional filtering or reconstruction methods, the system uses trained neural networks with deformable offset layers to automatically learn and apply the transformations needed to enhance image quality while removing artefacts
2Manufacturing precision
If adversarial losses are applied during GAN training, then image quality improves, but original anatomy structures may be distorted or lost
Solution Approach 1:
The GAN training process is segmented into distinct loss components: adversarial loss for image quality enhancement and a separate deformation loss for preserving anatomical structures. The deformation loss specifically monitors and constrains the deformable offset layers to ensure they do not distort critical anatomical features, while the adversarial loss focuses on overall image quality improvement
Solution Approach 2:
The patent applies different quality requirements to different regions of the image. The deformable offset layers are trained to preserve local anatomical structures with high fidelity while allowing global image quality enhancement. This is achieved by weighting the loss function to prioritize preservation of anatomically critical regions while still improving overall image appearance
3Manufacturing precision
If multiple artefact suppression methods are applied, then different artefacts can be reduced, but the processing becomes complex and inefficient
Solution Approach 1:
The GAN system is designed as a universal artefact removal solution that handles multiple types of artefacts simultaneously through a single integrated model. The deformable offset layers and adversarial training mechanism work together to remove various artefacts including motion artefacts, beam hardening artefacts, and noise, eliminating the need for separate processing steps for each artefact type
Data Source
AI summary
Techniques for generating a synthetic computed tomography (sCT) image from a cone-beam computed tomography (CBCT) image are provided. The techniques include receiving a CBCT image of a subject; generating, using a generative model, a sCT image corresponding to the CBCT image, the generative model trained based on one or more deformable offset layers in a generative adversarial network (GAN) to process the CBCT image as an input and provide the sCT image as an output; and generating a display of the sCT image for medical analysis of the subject.