Source-linked AI summary
Low-Dose CT with a Residual Encoder-Decoder Convolutional Neural Network (RED-CNN)
Hu Chen, Yi Zhang, Mannudeep K. Kalra, Feng Lin, Yang Chen, Peixi Liao, Jiliu Zhou, Ge Wang
TL;DR
Low-dose CT denoising needs methods that better suppress noise while preserving image details, but existing image-denoising CNNs generally use limited depth. RED-CNN combines an autoencoder, deconvolutional layers, and shortcut connections, achieving favorable results on simulated and clinical data for noise suppression, structural preservation, and lesion detection.
Problem
Existing image-denoising CNNs generally use limited depth, leaving the potential of deep CNNs for low-dose CT insufficiently realized.
Method
RED-CNN combines a symmetrical convolutional-deconvolutional network with shortcut connections, using deconvolution to recover structural details after feature extraction.
Results
Simulated and clinical evaluations found RED-CNN favorable for noise suppression, structural preservation, and lesion detection, with the highest PSNR/SSIM reported for all regions of interest.
Takeaways & Limitations
RED-CNN shows potential for low-dose CT reconstruction across simulated and clinical cases while operating at high computational speed.
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
Low-dose CT reduces radiation exposure but increases image noise, motivating image-domain methods that preserve diagnostic details. RED-CNN addresses this gap by combining deep CNN processing with deconvolution and shortcut connections.
- Lower X-ray flux reduces patient radiation exposure but produces noisier reconstructed CT images that can compromise diagnostic performance.
- Sinogram filtering and iterative reconstruction use raw-data or statistical models, but they can lose spatial resolution, reconstruct slowly, or alter image appearance.
- Post-processing reconstructed images avoids raw-data access, but image-domain noise statistics are difficult to model while preserving structure.
- Prior low-dose CT denoising CNNs generally used shallow models, leaving the potential of deeper networks insufficiently realized.
- RED-CNN incorporates a deconvolution network and shortcut connections into a CNN model for low-dose CT imaging.
A. Noise Reduction Model
The method treats low-dose CT denoising as image-domain recovery from a degraded image to its corresponding normal-dose image. Deep learning estimates an inverse mapping by minimizing reconstruction error.
- The workflow begins with filtered back-projection of a low-dose scan and restricts denoising to the image domain.
- A low-dose image X is modeled as the degradation σ(Y) of a corresponding normal-dose image Y.
- The degradation process includes quantum noise and other factors, so the model seeks a function f approximating the inverse degradation.
- The inverse mapping is estimated with deep learning by minimizing the squared difference between the predicted and normal-dose images.
B. Residual Autoencoder Network
RED-CNN combines symmetric convolutional and deconvolutional layers with shortcut connections to suppress noise while recovering and preserving image details. Residual compensation also addresses optimization difficulties in deeper networks.
- Residual Autoencoder Network: The network uses symmetric convolutional and deconvolutional layers rather than fully connected encoding and decoding layers.
- Residual Autoencoder Network: RED-CNN contains 10 layers—5 convolutional and 5 deconvolutional—with shortcuts connecting matching layers and ReLU activations after each layer.
- Residual Autoencoder Network: Convolutional encoders transform image patches into feature representations, while deconvolutional decoders reconstruct patches from those features.
- Residual Autoencoder Network: Because convolution can remove image details and deeper networks can accumulate reconstruction loss and gradient diffusion, residual compensation is introduced.
- Residual Autoencoder Network: Residual mapping facilitates deeper-network training and preserves structural and contrast details by processing only the residual through convolutional and deconvolutional layers.
- Residual Autoencoder Network: Shortcut connections preserve structural details and facilitate deeper training, while paired deconvolution layers suppress noise and artifacts without pooling or down-sampling.
A. Data Sources
The study used both simulated and clinical low-dose CT data to assess RED-CNN across diverse imaging conditions. Simulated data were generated from normal-dose images with controllable Poisson noise, while clinical data came from the NIH-AAPM-Mayo Clinic challenge.
- Simulated data: 7,015 normal-dose 256×256 CT images from 165 NBIA patients formed the basis for simulated data.The dataset included different body parts for greater diversity.
- Simulated data: Poisson noise was added to sinograms simulated from normal-dose images to produce corresponding low-dose images.The blank scan factor controlled the simulated noise level.
- Simulated data: The simulation used Siddon’s ray-driven method in fan-beam geometry with 512 detector elements and 1,024 uniformly distributed viewing angles.The source-to-rotation-center and detector-to-rotation-center distances were each 40 cm.
- Training data: RED-CNN processed image patches because patch extraction represented local denoising details and increased the number of training samples.This approach addressed the difficulty of training deep learning models with limited medical images.
- Training and clinical data: 200 normal-dose/simulated low-dose image pairs were used for training and 100 pairs for testing, with no patient overlap between sets.Clinical validation used the authorized 2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge database.
- Dataset rationale: Both simulated and clinical datasets were used because they differed in body-part diversity and low-dose image-generation conditions.The NBIA data covered more body parts, whereas the Mayo data included full-dose images and corresponding quarter-dose counterparts.
B. Parameter selection
The study selected patch-extraction and image-quality evaluation settings, then compared RED-CNN with five established low-dose CT methods. Assessment used RMSE, PSNR, and SSIM.
- Parameter selection: 55×55 patches with a 4-pixel sliding interval produced 10^6 training patch pairs after extraction.Rotation, flipping, and scaling augmentation were also applied.
- Evaluation metrics: RMSE, PSNR, and SSIM were selected as quantitative image-quality metrics.These metrics assessed reconstruction quality from complementary error, signal-to-noise, and structural-similarity perspectives.
- Comparison methods: RED-CNN was compared with TV-POCS, K-SVD, BM3D, CNN10, and KAIST-Net.CNN10 omitted RED-CNN’s shortcut connections and deconvolutional layers, while the other methods represented image-based, iterative, or CNN-based alternatives.
C. Experimental Results
Across simulated and clinical cases, RED-CNN generally provided the strongest balance of noise suppression, structural preservation, quantitative accuracy, and lesion visibility among the compared methods.
- RED-CNN suppressed most noise and artifacts while preserving structural features better than the other methods in chest-image comparisons.It also discriminated low-contrast regions best and produced the smallest absolute difference from the normal-dose image.
- The RED-CNN had the lowest RMSE and highest PSNR/SSIM for all evaluated chest-image ROIs.
- In abdominal images, RED-CNN best retained a linear high-attenuation liver structure, avoided pseudo-lesions, and better delineated tissue margins.The thin right adrenal gland was also best appreciated on the RED-CNN image.
- RED-CNN obtained the best scores on all indexes for the abdominal image and outperformed state-of-the-art methods on every metric across the 100-image testing dataset.
- In clinical scans, RED-CNN delivered the best noise suppression and structure preservation, with the strongest image quality for low-contrast liver lesions.It preserved lesion edges better than KAIST-Net and made two tiny focal lesions noticeable.
- Reader-study scores improved significantly for all methods, while RED-CNN performed slightly better than KAIST-Net and ran significantly faster in training and testing.The study assessed artifact reduction, noise suppression, contrast retention, lesion discrimination, and overall quality.
D. Model and Performance Trade-Offs
The study examines RED-CNN’s architectural and training trade-offs, robustness to noise-level variation, and computational cost, alongside clinical image comparisons.
- D. Model and Performance Trade-Offs: The evaluation examines the deconvolutional decoder, shortcut connections, network depth, patch size, dataset robustness, and computational costs.These factors were identified as critical components of RED-CNN performance.
- D. Model and Performance Trade-Offs: The abdominal comparisons include metastasis and focal fatty sparing cases, with zoomed views identifying a lesion and a contrast-enhanced blood vessel.The figures compare NDCT, LDCT, TV-POCS, K-SVD, BM3D, CNN10, KAIST-Net, and RED-CNN.
- D. Model and Performance Trade-Offs: A 55×55 patch size was selected because increasing the size from 55 to 100 produced no significant difference while requiring more training time.The chosen size was treated as a trade-off between training time and imaging performance.
- D. Model and Performance Trade-Offs: RED-CNN+ achieved the best performance in most inconsistent-noise situations when trained with mixed data across possible noise levels.RED-CNN remained competitive even when trained at a single noise level.
- D. Model and Performance Trade-Offs: Table III and Table IV report quantitative results for different algorithms, including full cross-validation results in Table IV.The supplied passages provide the table headings but not their numerical entries.
- D. Model and Performance Trade-Offs: After offline training, RED-CNN required 3.68 seconds for execution, compared with 21.36 seconds for ASD-POCS and 38.45 seconds for K-SVD.The reported execution times also included BM3D at 4.22 seconds, CNN10 at 3.22 seconds, and KAIST-Net at 30.22 seconds.
IV. CONCLUSION
The conclusion presents RED-CNN as a symmetrical convolutional-deconvolutional network with shortcut connections and reports favorable simulated and clinical results. These results indicate potential for noise suppression, structural preservation, lesion detection, and high computational speed.
- IV. CONCLUSION: RED-CNN uses a symmetrical convolutional and deconvolutional neural network aided by shortcut connections.The model was evaluated and validated using two well-known databases against state-of-the-art methods.
- IV. CONCLUSION: Simulated and clinical results demonstrated potential for noise suppression, structural preservation, lesion detection, and high computational speed.The conclusion reports these outcomes across the evaluation settings.
- IV. CONCLUSION: The authors plan to optimize RED-CNN, extend it to 3D, dynamic or spectral CT reconstruction, and adapt the ideas to other imaging tasks or modalities.These directions are presented as future work rather than established results.