3D CNN Cone-Beam Artifact Reduction for CBCT Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for reducing cone-beam computed tomography (CBCT) artifacts rely on two-dimensional neural networks or pseudo-3D networks, requiring substantial data sets and are computationally heavy, while there is a need for a deep learning-based method that can be easily trained and generalized across various CBCT cone angles and helical artifacts.
Innovation Solution
A method involving a three-dimensional convolutional neural network (CNN) trained using simulated three-dimensional digital phantoms to correct cone-beam artifacts, utilizing a U-net architecture with half-precision training and positional encoding to efficiently learn artifact reduction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If two-dimensional neural networks or pseudo-3D networks are used for artifact reduction, then artifact correction can be achieved, but substantial data sets and computational resources are required
Solution Approach 1:
The patent transitions from 2D neural networks to 3D convolutional neural networks by adding the depth dimension to process volumetric data directly. This dimensional expansion enables the network to capture spatial relationships in three dimensions, improving artifact correction capability while maintaining computational efficiency through optimized 3D convolution operations.
Solution Approach 2:
The patent uses simulated digital phantoms as synthetic copies of real patient data to generate training datasets. These phantoms are computationally generated three-dimensional volumes that mimic anatomical structures and can be efficiently forward-projected to create realistic CBCT images with artifacts, eliminating the need for substantial clinical datasets.
2Reliability
If iterative reconstruction or forward-back projection methods are used, then artifact reduction can be achieved, but computationally heavy operations are required
Solution Approach 1:
The patent replaces the mechanical iterative reconstruction process with a learned neural network model. Instead of repeatedly performing forward and back projections to iteratively minimize reconstruction error, the 3D CNN directly maps artifact-prone images to corrected images through learned transformations, dramatically reducing computational power requirements while maintaining artifact reduction effectiveness.
Solution Approach 2:
The patent performs preliminary training of the 3D neural network using simulated phantom data before deployment. This preliminary action allows the network to learn artifact patterns and correction strategies in advance, so that during actual clinical use, artifact correction can be performed rapidly without requiring computationally intensive iterative reconstruction operations.
3Measurement precision
If high radiation dose is used to generate noiseless images, then training data quality is improved, but patient exposure increases
Solution Approach 1:
The patent creates synthetic copies of anatomical structures using digital phantoms that can be forward-projected to generate training images. These phantoms allow generation of both artifact-prone and artifact-free images without requiring additional patient scans, eliminating the need for high radiation dose exposure while providing unlimited training data with perfect registration.
Solution Approach 2:
The system uses simulated phantoms that serve their own purpose as both the object being imaged and the ground truth reference. The same phantom volume is forward-projected with different parameters to generate both the corrupted CBCT images and the corresponding clean reference images, making the system self-sufficient for training data generation without external clinical scans.
4Volume of moving object
If cone angle is increased to improve Z-axis coverage, then scan coverage is improved, but artifacts become more pronounced
Solution Approach 1:
The patent trains the 3D neural network using simulated data with varying cone angles and acquisition parameters. By exposing the network to diverse parameter conditions during training, including different Z-axis coverages and cone angles, the model learns to generalize artifact correction across multiple scenarios, making it effective for both small and large cone angle acquisitions.
Solution Approach 2:
The 3D neural network is designed to be universally applicable across different CBCT acquisition configurations. The model processes volumetric data in a cone-angle-agnostic manner, enabling it to correct artifacts regardless of the specific cone angle or Z-axis coverage used during scanning, making a single model sufficient for multiple clinical protocols.
Data Source
AI summary
Systems and methods for training a machine-learning model for artifact reduction are provided. Such methods include retrieving a three-dimensional digital phantom reconstructed from CT imaging data. The method then selects a first Z position along the central axis and simulates a first set of forward projections from the digital phantom taken along an axial trajectory at the first Z position along the central axis. The first set of forward projections has a first simulated collimation in the axial direction. The method then reconstructs a first simulated image from the first set of forward projections and identifies a plurality of secondary Z positions along the central axis other than the first Z position. For each of the secondary Z positions and the first Z position itself, the method then simulates a set of secondary forward projections from the digital phantom taken along corresponding axial trajectories at the corresponding secondary Z position.