Abstract
Given the potential X-ray radiation risk to the patient, low-dose CT has attracted a considerable interest in the medical imaging field. The current main stream low-dose CT methods include vendor-specific sinogram domain filtration and iterative reconstruction, but they need to access original raw data whose formats are not transparent to most users. Due to the difficulty of modeling the statistical characteristics in the image domain, the existing methods for directly processing reconstructed images cannot eliminate image noise very well while keeping structural details. Inspired by the idea of deep learning, here we combine the autoencoder, the deconvolution network, and shortcut connections into the residual encoder-decoder convolutional neural network (RED-CNN) for low-dose CT imaging. After patch-based training, the proposed RED-CNN achieves a competitive performance relative to the-state-of-art methods in both simulated and clinical cases. Especially, our method has been favorably evaluated in terms of noise suppression, structural preservation and lesion detection.
Introduction
Medical professionals face a constant tug-of-war between diagnostic clarity and patient safety. While Computed Tomography (CT) is a vital diagnostic tool, the X-ray radiation it uses carries potential health risks. Following the ALARA (As Low As Reasonably Achievable) principle, clinicians try to use the lowest dose possible. However, reducing the X-ray flux introduces significant noise and "streaking" artifacts into the images, which can obscure small lesions or mimic pathology.
Existing solutions have major hurdles. Sinogram filtering requires access to proprietary raw data. Iterative Reconstruction (IR) is computationally expensive and slow. Standard image processing often results in "waxy" textures or the loss of critical structural details. To address these issues, the Residual Encoder-Decoder Convolutional Neural Network (RED-CNN) is proposed as a deep-learning solution for denoising low-dose CT (LDCT) scans.
Methods
The RED-CNN is a 10-layer architecture designed specifically for the "low-level" task of image denoising. It combines a symmetric architecture of convolutional and deconvolutional layers with shortcut "residual" connections.
Architectural Components
- Symmetric Encoder-Decoder: The network consists of 5 convolutional layers (the encoder) to extract features and 5 deconvolutional layers (the decoder) to reconstruct the image. They follow a "First-In-Last-Out" (FILO) symmetry.
- Residual Learning & Shortcuts: Instead of forcing the network to learn the entire clean image from scratch, shortcut connections allow the network to learn the residual (the noise/difference). This makes the optimization process much more efficient and helps preserve fine structural details.
- No Pooling Layers: Unlike standard CNNs used for image recognition, RED-CNN discards pooling. This ensures that spatial resolution is maintained throughout the network—a necessity for medical imaging where every pixel counts.
- Patch-based Training: The model is trained on small, overlapping image patches, allowing the network to focus on local textures and edges while effectively boosting the dataset size.
Datasets and Evaluation
The study utilized two primary data sources for validation:
- Simulated Data (NBIA): 7,015 normal-dose CT images from 165 patients, with low-dose versions simulated by adding Poisson noise to the sinograms.
- Clinical Data (Mayo Clinic): Data from the "2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge," featuring 2,378 images from 10 patients. This included real-world pairs of full-dose and quarter-dose scans.
Results and Key Insights
RED-CNN was compared against five state-of-the-art methods, including TV-POCS, K-SVD, and BM3D.
- Quantitative Superiority: RED-CNN consistently achieved the highest Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) across both datasets.
- Qualitative Performance: Radiologist evaluations ranked RED-CNN highest for artifact reduction, noise suppression, and lesion discrimination.
- Efficiency: Once trained, the model can process a 512x512 image in roughly 3.68 seconds, significantly faster than traditional iterative reconstruction methods.
Key insights from the development include the finding that deconvolutional layers are superior to simple upsampling for recovering structural details, and that learning the noise (residual) is mathematically easier than reconstructing the underlying anatomy from scratch.
Conclusion
The RED-CNN provides a powerful deep-learning solution for LDCT denoising, significantly outperforming traditional methods in both image quality and processing speed. While the current model operates on 2D slices, future work could extend the architecture to 3D to utilize cross-slice information for even greater performance improvements.
Summary updated