Source-linked AI summary
Sharpness-aware Low dose CT denoising using conditional generative adversarial network
Xin Yi, Paul Babyn
TL;DR
LDCT reduces radiation exposure but increases noise, while pixel-wise denoising can blur reconstructed images. SAGAN combines conditional adversarial training with sharpness guidance, achieving improved quantitative and visual performance with small resolution loss. The authors report substantial dose-reduction correspondence, while clinical diagnosis performance remains unevaluated.
Problem
LDCT noise can impede diagnosis, and pixel-wise denoising methods can produce blur, especially at high noise levels.
Method
SAGAN combines a conditional adversarial network with a sharpness detection network to guide LDCT denoising.
Results
SAGAN achieves improved quantitative and visual performance, with very small spatial-resolution loss relative to tested competitors.
Takeaways & Limitations
At the lowest dose, SAGAN corresponded to 90% dose reduction by PSNR and SSIM and 95% by mean standard deviation of attenuation.
Takeaways & Limitations
Clinical diagnosis performance remains to be evaluated, and sharpness estimation may limit final sharpness in small low-contrast regions.
Abstract
from arXiv · showhide
Low Dose Computed Tomography (LDCT) has offered tremendous benefits in radiation restricted applications, but the quantum noise as resulted by the insufficient number of photons could potentially harm the diagnostic performance. Current image-based denoising methods tend to produce a blur effect on the final reconstructed results especially in high noise levels. In this paper, a deep learning based approach was proposed to mitigate this problem. An adversarially trained network and a sharpness detection network were trained to guide the training process. Experiments on both simulated and real dataset shows that the results of the proposed method have very small resolution loss and achieves better performance relative to the-state-of-art methods both quantitatively and visually.
1 Introduction
Reducing tube current lowers radiation dose but increases LDCT noise, potentially impeding diagnosis. The paper explores GAN-based LDCT denoising to address blur associated with pixel-wise reconstruction losses.
- Lowering tube current reduces radiation dose but inevitably introduces more noise into LDCT than convCT.
- Increased LDCT noise may impede subsequent diagnosis or require more advanced reconstruction algorithms.
- Deep learning has shown strong performance across image tasks and is being applied in medical imaging.
- The paper investigates generative adversarial networks for LDCT denoising.
- Pixel-wise losses can produce blur, motivating adversarial loss and a sharpness network focused on low-contrast regions.
2 Related Works
LDCT denoising methods operate in sinogram, image, or iterative reconstruction domains, each with practical or structural trade-offs. The paper positions SAGAN among cGAN approaches through its architecture, patch discriminator, and broader evaluation.
- LDCT denoising methods broadly operate in sinogram, image, or iterative reconstruction domains.
- Sinogram-based methods require raw projection data, and edge-preserving smoothing can remove small edges and reduce reconstructed spatial resolution.
- MBIR iterates between image and sinogram domains while modeling scanner geometry and photon-counting or source-spectrum properties.
- Image-based methods avoid raw-data requirements and include redundancy-based approaches such as Non-Local means and BM3D.
- CNN-based CT denoising has used patch regression, encoder-decoder networks, wavelet-domain training, perceptual loss, and MTANNs.
- SAGAN differs from prior cGAN denoisers through a deeper residual U-Net generator, PatchGAN discriminator, and evaluation across dose levels and anatomical regions.
3 Methods
SAGAN denoises LDCT with a conditional generator trained against adversarial, pixel-wise, and sharpness objectives. Its U-Net and residual design preserves detail while the discriminator evaluates conditioned image patches.
- SAGAN maps LDCT x to a denoised virtual convCT ˆy using generator G, discriminator D, and sharpness network S.
- The discriminator compares conditioned pairs (x, ˆy) and (x, y), while G learns to make the virtual pair appear real.
- The generator combines adversarial and pixel-wise L1 objectives to balance realistic appearance with data fidelity.
- Sharpness loss compares sharpness maps S(ˆy) and S(y), explicitly encouraging generated images to match convCT sharpness.
- The residual block in the generator is repeated K times, with K chosen as 9 for the experiment.
- The generator uses a U-Net with long skip connections and residual components to recover subtle CT structure and ease training.
- PatchGAN classifies overlapping patches rather than whole images, enabling arbitrary image sizes with fewer discriminator parameters.
4 Experiment Setup
The experiments evaluate SAGAN on simulated, phantom, piglet, and clinical LDCT data using image-quality, noise, spatial-resolution, and generality assessments. The setup combines dose-controlled datasets, multiple baselines and generator architectures, and a separately trained sharpness detection network.
- Evaluation and training: PSNR, SSIM, and homogeneous-region noise measurements were used across simulated and real evaluations, while generator architectures and sharpness-network training were separately assessed.The evaluated generators included unet256, res9, and Derain; the sharpness network was trained from 704 defocused images using a UNet256 structure.
- Simulated noise dataset: 239 normal-dose CT images from NBIA were converted into sinograms with 937 detectors and 1200 views for simulated-noise experiments.The resulting 256 × 256 sub-images were split into 700 training samples, while 64 full images formed the test set.
- Simulated noise dataset: Simulated dose levels used blank flux values N0 of 1 × 10^5, 5 × 10^4, 3 × 10^4, and 1 × 10^4, with electrical noise discarded.N0 denotes the X-ray source intensity or blank flux.
- Real datasets: Real piglet scans covered 300 to 15 mAs, corresponding to full dose and 50%, 25%, 10%, and 5% tube-current reductions.The real training set was enlarged to 2832 sub-images by dividing each image into four 256 × 256 regions.
- Real datasets: The Catphan 600 phantom compared 300 mAs convCT with 10 mAs LDCT to evaluate reconstructed spatial resolution.Both scans used 120 kVp and 0.625 mm slice thickness.
- Experimental design: Four experiments assess generator and sharpness-loss effects, spatial resolution, real quantum-noise denoising, and generalization to clinical patient scans.The clinical evaluation applies a model trained on piglet data to two Data Science Bowl 2017 scans.
5 Results
Across simulated and real-data experiments, SAGAN preserved fine spatial detail while reducing noise and outperforming comparator methods visually and quantitatively.
- Sharpness loss: Removing the sharpness loss caused only slight PSNR and SSIM differences, while the loss better suppressed noise on small structures without excessive blurring.The sharpness-aware results were visually clearer despite comparable conventional fidelity metrics.
- Sharpness loss: SAGAN achieved better PSNR and SSIM than BM3D and K-SVD at every tested noise level, while recovering sharper details and low-contrast vessels.BM3D produced streak artifacts at high quantum noise, whereas K-SVD sometimes left noise and blocky structure.
- Catphan 600: SAGAN preserved the 4- and 5-line pairs in Catphan images with very small spatial-resolution loss and achieved the highest spatial resolution among comparators.The 6-line pairs remained distinguishable in convCT but not in LDCT or any reconstruction method.
- Piglet dataset: As dose increased, PSNR generally improved for all methods except VEO, while SAGAN reached an average SSIM of 0.95 at the lowest dose.The reported SSIM was slightly higher than FBP at its second-highest dose level.
- Piglet dataset: At the highest noise level, SAGAN reduced the mean standard deviation from 25.35 for FBP to 8.80, a noise reduction factor of 2.88.Its noise measure stayed close to convCT across dose levels, unlike BM3D and K-SVD, which achieved smaller values through over-smoothing.
- Clinical patient data: On clinical patient images with unknown dose levels, SAGAN performed reasonably well and increased contrast-to-noise ratio.These dose levels were unlikely to coincide with those used for training.
6 Discussion
SAGAN combines adversarial and sharpness-aware components to preserve structure, texture, and resolution in LDCT denoising across dose levels and anatomies. It also shows strong noise reduction and visual quality, while clinical diagnosis and broader comparisons remain limited.
- Image quality: SAGAN recovers underlying structures, produces convCT-like texture and sharpness, and outperforms BM3D at lower doses.BM3D performs better at the highest dose, while BM3D and K-SVD show streak artifacts at low doses.
- Image quality: SAGAN mitigates streak artifacts, likely because patch-based discrimination treats artifact-containing patches as outliers.The proposed extension to limited-view CT reconstruction is stated as future work.
- Architecture and design: The generator uses a residual Unet, the discriminator classifies patches, and the sharpness-aware loss targets visual sharpness rather than perceptual similarity.Patch discrimination reduces parameter count and supports arbitrary image sizes.
- Scope and limitations: SAGAN was evaluated across dose levels and anatomic regions, but clinical diagnosis performance remains unevaluated and comparisons with other deep-learning methods lack dose-reduction quantification.The study includes clinical patient data with unknown dose levels and visual assessment.
- Quantitative performance: 90% dose reduction is indicated by PSNR and SSIM, while 95% is indicated by mean standard deviation of attenuation.The 90% comparison uses 0.71 mSv versus 7.07 mSv; the 95% comparison uses 0.71 mSv versus 14.14 mSv.
7 Conclusion
The paper proposes a sharpness-aware network for LDCT denoising that combines adversarial and sharpness losses to address blur under high noise. It reports improved quantitative performance and more appealing visual results, but identifies limitations in sharpness sensitivity and clinical evaluation.
- SAGAN combines adversarial loss and sharpness loss to mitigate blur in image-based LDCT denoising, especially at high noise levels.
- The sharpness detector may miss just-noticeable blur, limiting final sharpness in small low-contrast regions.
- Clinical diagnosis performance remains unevaluated, and deep-learning methods require training against a specific dosage level.