Abstract
Cone-beam breast computed tomography (CT) provides true 3D breast images with isotropic resolution and high-contrast information, detecting calcifications as small as a few hundred microns and revealing subtle tissue differences. However, breast is highly sensitive to x-ray radiation. It is critically important for healthcare to reduce radiation dose. Few-view cone-beam CT only uses a fraction of x-ray projection data acquired by standard cone-beam breast CT, enabling significant reduction of the radiation dose. However, insufficient sampling data would cause severe streak artifacts in CT images reconstructed using conventional methods. In this study, we propose a deep-learning-based method to establish a residual neural network model for the image reconstruction, which is applied for few-view breast CT to produce high quality breast CT images. We respectively evaluate the deep-learning-based image reconstruction using one third and one quarter of x-ray projection views of the standard cone-beam breast CT. Based on clinical breast imaging dataset, we perform a supervised learning to train the neural network from few-view CT images to corresponding full-view CT images. Experimental results show that the deep learning-based image reconstruction method allows few-view breast CT to achieve a radiation dose <6 mGy per cone-beam CT scan, which is a threshold set by FDA for mammographic screening.
Introduction
While traditional 2D mammography remains the standard for breast cancer screening, it suffers from limitations inherent in projecting 3D structures onto a 2D plane. This often results in "false negatives" due to tissue masking or "false positives" from overlapping structures. Cone-beam breast CT (CBBCT) addresses these issues by providing isotropic 3D resolution. However, the high radiation sensitivity of breast tissue necessitates dose reduction for routine screening.
The primary challenge in dose reduction via "few-view" imaging—reducing the number of X-ray projections—is the introduction of severe streak artifacts and noise when using conventional reconstruction algorithms like Filtered Back-Projection (FBP). This study presents a deep-learning-based approach using a residual neural network (ResNet) to reconstruct high-quality images from just 25–33% of standard X-ray data, successfully lowering radiation levels below the FDA’s 6 mGy threshold for mammographic screening.
Methods
The proposed method utilizes a Deep Residual Neural Network (ResNet) as a post-processing step to bridge the gap between low-dose (few-view) and routine-dose (full-view) images. An initial noisy image is generated from few-view data using standard FBP. This image is then processed by the ResNet, which is specifically designed to identify and remove artifacts and noise introduced by undersampling.
Network Architecture
The ResNet architecture consists of an initial convolutional layer with 64 filters, followed by four "Residual Blocks." Each block contains two convolutional layers with shortcut connections to facilitate gradient flow. Final layers consolidate features into a single high-quality output image.
Dataset and Training
The study used a clinical dataset from Koning Corporation. Standard scans with 300 projection views were downsampled to 100 views (1/3 dose) and 75 views (1/4 dose). For training, nearly 2 million pairs of 64x64 image patches were extracted from 2,067 CT images, with an additional 800,000 patches used for validation.
Results
The deep learning approach demonstrated significant improvements over standard FBP for both 1/3 and 1/4 dose scenarios.
- One-Third Dose (100 views): Achieved a Peak Signal-to-Noise Ratio (PSNR) of 34.90 and a Structural Similarity Index (SSIM) of 0.9699.
- One-Quarter Dose (75 views): Maintained high quality with a PSNR of 32.96 and SSIM of 0.9494.
Visually, the ResNet effectively eliminated the heavy streaking artifacts present in the raw few-view FBP reconstructions, producing images nearly indistinguishable from full-dose scans.
Discussion and Conclusion
This research demonstrates that deep learning can push CBBCT into the "safe" radiation zone (< 6 mGy) required for routine screening. By focusing on artifact removal rather than full image generation, the ResNet approach proves more efficient than traditional iterative methods and better preserves subtle tissue details.
Future work should focus on validating the model across diverse hardware and patient populations. Additionally, clinical blind studies with radiologists are necessary to confirm that the AI-reconstructed images maintain diagnostic integrity without introducing "hallucinated" structures or over-smoothing critical micro-calcifications.
Summary updated