Multi-Scale Deep Learning Reconstruction for Sparse-View DBT Artifacts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital breast tomosynthesis (DBT) systems suffer from limited angular artifacts and tissue superposition, leading to ambiguity and reduced physician confidence, while also facing high computational costs for 3D reconstruction.
Innovation Solution
A multi-scale deep learning-based approach using a first low-resolution network for artifact reduction and a second high-resolution network for image detail refinement, with decoupled training to minimize computational burden and enhance in-depth resolution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If DBT uses large sampling interval (≥3°) and limited angular sampling range (≤50°) to save dose and scanning time, then scanning time and radiation dose are reduced, but sparse-view artifacts and tissue superposition become severe
Solution Approach 1:
The patent introduces a deep learning-based reconstruction algorithm as an intermediary between the sparse projection data and the final tomosynthesis image. This intermediary process learns to infer missing information and suppress artifacts, enabling acceptable image quality from limited angular views without requiring increased sampling density
Solution Approach 2:
The patent transforms the reconstruction problem by changing parameters in the optimization process, using learned priors and regularization terms that adapt to the specific characteristics of sparse-view data. This allows the system to achieve artifact reduction through parameter optimization rather than through increased angular sampling
2Use of energy by moving object
If DBT uses large sampling interval (≥3°) and limited angular sampling range (≤50°) to save dose and scanning time, then scanning time and radiation dose are reduced, but tissue superposition increases
Solution Approach 1:
The deep learning reconstruction algorithm serves as an intermediary that processes the limited projection data to produce images with reduced tissue superposition. The learned models can separate overlapping structures more effectively than traditional methods, compensating for the limited angular range without requiring additional radiation dose
Solution Approach 2:
The system performs preliminary actions by pre-training the deep learning models on comprehensive datasets that include various tissue configurations. This preliminary training enables the model to anticipate and resolve tissue superposition issues during actual reconstruction from sparse views, without needing additional scanning angles
3Measurement precision
If 3D DBT is used to achieve high resolution, then in-depth resolution is improved, but training computational cost becomes huge
Solution Approach 1:
The patent segments the training process into multiple stages: pre-training on synthetic data, fine-tuning on real data, and specialized training for different anatomical regions. This segmentation reduces the computational burden of any single training run while achieving high in-depth resolution through cumulative learning
Solution Approach 2:
The system uses synthetic copies of breast tissue data with known ground truth for pre-training the deep learning models. These synthetic datasets can be generated efficiently without expensive acquisitions, allowing the model to learn fundamental reconstruction patterns before being fine-tuned on limited real patient data
Data Source
AI summary
Systems and methods are provided for a multi-scale deep learning-based digital breast tomosynthesis (DBT) image reconstruction that mitigates the superposition of breast tissue along with the limited angular artifacts, and improves in-depth resolution of the resulting images. A multi-scale deep neural network may be used where a first network may focus on a first parameter, such as limited angular artifacts reduction, and a second network may focus on a second parameter, such as image detail refinement. The output from the first neural network may be used as the input for the second neural network. The systems and methods may reduce the sparse-view artifacts in DBT via deep learning without losing image sharpness and contrast. A deep neural network may be trained in a way to reduce training-time computational cost. An ROI loss method may be used for further improvement on the resolution and contrast of the images.