TL;DR
This paper introduces DEER, a novel deep learning network for reconstructing high-quality breast CT images from a limited number of X-ray views. DEER achieves this with significantly fewer parameters and computational resources than existing methods, enabling faster and more efficient low-dose breast cancer screening.
What Problem Does This Paper Address?
Breast cancer is a leading cause of cancer death, and early detection through imaging is crucial. While CT scans offer excellent image quality for detecting small calcifications, the ionizing radiation involved poses a health risk. Reducing this radiation dose is paramount, and few-view scanning (acquiring images from fewer angles) is a promising approach. However, reconstructing accurate CT images from limited views is challenging due to severe streak artifacts caused by incomplete projection data. Traditional analytical methods like Filtered Back-Projection (FBP) struggle with this, and existing deep learning methods often require substantial computational resources and parameters, making them impractical for high-resolution 3D reconstruction on standard hardware. This paper aims to bridge this gap by developing an efficient and effective deep learning solution for few-view breast CT reconstruction.
Methods
The core of this work is the Deep Efficient End-to-End Reconstruction (DEER) network. DEER is designed to learn the entire reconstruction process from raw projection data to a final CT image, inspired by the filtered back-projection (FBP) mechanism. Key aspects of the DEER architecture include:
- Generator Network (G): This network takes few-view sinograms as input and outputs a reconstructed image. It's composed of two main parts:
- Back-projection: This learned component aims to reconstruct images from projection data by considering that each sinogram point relates to pixels along a specific X-ray path. The authors propose a "point-wise fully-connected layer" that, in theory, allows for a reconstruction process with O(N) parameters (where N is image dimension), a significant reduction from prior methods. For better performance, they initially use O(N² × N₂) parameters.
- Refinement: A U-net architecture is used to further refine the image, removing residual artifacts and enhancing details.
- Discriminator Network (D): This network, part of a Wasserstein Generative Adversarial Network (WGAN) framework, helps the generator produce more realistic and faithful images by distinguishing between real (ground-truth) and generated images. The WGAN framework is chosen for its stability compared to traditional GANs.
- Objective Function: The generator is optimized using a combination of Mean Absolute Error (MAE) for robust error handling, Structural Similarity Index (SSIM) for visually appealing results, and an adversarial loss from the WGAN.
- Data Preprocessing: For 3D cone-beam data, an initial FDK reconstruction is performed on intermediate slices to generate parallel-beam projections, which are then fed into the DEER network. This strategy helps manage memory constraints for high-resolution 3D data.
Datasets
The study utilizes a clinical female breast dataset provided by Koning Corporation. This dataset consists of 19,575 breast CT images reconstructed from 300 cone-beam projections (ground-truth images). For training and evaluation, 75-view cone-beam projections were used as input. The dataset was split into training (13,970 images from 32 patients), validation (1,197 images from 3 patients), and testing (4,408 images from 7 patients) sets.
Key Insights and Ideas
- Learned Back-Projection for Efficiency: The core innovation is learning the back-projection process itself, rather than relying on fixed analytical methods. This allows for a more efficient parameterization (potentially O(N)) and better adaptation to the specific data.
- WGAN for Stability and Quality: Employing WGAN instead of a standard GAN leads to more stable training and, importantly, better recovery of subtle details, even if it slightly impacts some quantitative metrics like PSNR.
- End-to-End Learning from Sinograms: DEER learns the mapping directly from raw sinogram data to the reconstructed image, bypassing intermediate steps and potentially capturing more information than traditional methods.
- Memory Management for High Resolution: The strategy of using intermediate FDK reconstructions to generate parallel-beam projections for 3D data is a practical solution to overcome GPU memory limitations for high-resolution reconstructions.
Key Results
- Image Quality: DEER produced visually superior reconstructions compared to FDK and other deep learning methods like Residual-CNN and FBPConvNet, particularly in preserving subtle details in challenging regions (ROIs).
- Quantitative Performance: DEER achieved competitive or superior results in SSIM and MAE compared to other methods, indicating good structural similarity and low error. While PSNR was slightly lower than FBPConvNet, the authors argue that PSNR alone doesn't guarantee denoising performance and visual quality is more critical.
- Computational Efficiency: The DEER network requires significantly fewer parameters (potentially O(N)) compared to prior deep learning methods, leading to reduced memory burden and higher computational efficiency.
- Reconstruction Time: DEER reconstructs a single 2D slice (1024x1024) in approximately 0.1422 seconds on an NVIDIA Titan RTX GPU.
Limitations and Future Work
- 3D Reconstruction Memory: While the method is designed for 3D data, processing full 3D cone-beam datasets directly on current hardware is still a memory challenge. The current approach uses slice-based reconstruction.
- View-Dependent vs. View-Independent: The proposed DEER is view-dependent, meaning parameters are learned for specific angles. While this improves quality, a view-independent version (DEER-Lite) was explored, showing promise for generalization to different view numbers without retraining, though it might introduce artifacts in some cases.
- Future Directions: The authors plan to extend DEER for direct 3D cone-beam CT reconstruction, explore its application to other imaging modalities (SPECT, PET), and further optimize it for clinical applications. They also aim to address the divergent cone-beam geometry and ray-specific varying weights for more direct 3D reconstruction.
Summary updated