PaperlayerView exact revision on arXiv

Source-linked AI summary

EDCNN: Edge enhancement-based Densely Connected Network with Compound Loss for Low-Dose CT Denoising

Tengfei Liang, Yi Jin, Yidong Li, Tao Wang, Songhe Feng, Congyan Lang

arXiv:2011.00139v1eess.IVcs.CV

TL;DR

Low-dose CT denoising seeks to reduce radiation-related noise while avoiding the oversmoothing and detail loss seen in existing methods. EDCNN combines adaptive trainable-Sobel edge enhancement, dense feature fusion, and a compound MSE plus multi-scales perceptual loss. Experiments report better detail preservation and noise suppression than existing algorithms.

  • Problem

    Low-dose CT reduces radiation exposure but introduces noise, while existing denoisers can oversmooth images and lose edges and details.

  • Method

    EDCNN combines a trainable-Sobel edge enhancement module, dense connections for fusing edge and input information, and a compound MSE plus multi-scales perceptual loss.

  • Results

    EDCNN achieves the best or suboptimal results on every reported criterion and balances per-pixel and structurewise performance.

  • Takeaways & Limitations

    The model reports better preservation of subtle CT details and suppression of noise than previous low-dose CT denoising models.

Abstract

from arXiv · show

In the past few decades, to reduce the risk of X-ray in computed tomography (CT), low-dose CT image denoising has attracted extensive attention from researchers, which has become an important research issue in the field of medical images. In recent years, with the rapid development of deep learning technology, many algorithms have emerged to apply convolutional neural networks to this task, achieving promising results. However, there are still some problems such as low denoising efficiency, over-smoothed result, etc. In this paper, we propose the Edge enhancement based Densely connected Convolutional Neural Network (EDCNN). In our network, we design an edge enhancement module using the proposed novel trainable Sobel convolution. Based on this module, we construct a model with dense connections to fuse the extracted edge information and realize end-to-end image denoising. Besides, when training the model, we introduce a compound loss that combines MSE loss and multi-scales perceptual loss to solve the over-smoothed problem and attain a marked improvement in image quality after denoising. Compared with the existing low-dose CT image denoising algorithms, our proposed model has a better performance in preserving details and suppressing noise.

I. INTRODUCTION

Low-dose CT reduces radiation exposure but introduces noise that complicates diagnosis, while existing denoisers can oversmooth images and lose edges and details. EDCNN addresses these issues with adaptive edge extraction, dense feature fusion, and a compound training loss.

  • Lower X-ray doses reduce safety risks but introduce more CT image noise, complicating later diagnosis.
  • CNN-based low-dose CT denoisers have shown strong potential, but existing models still produce oversmoothed results and lose edge and detail information.
  • EDCNN uses an edge enhancement module based on trainable Sobel convolution to adaptively extract edge features during optimization.
  • Its fully convolutional architecture uses densely connected conveying paths to fuse input information with extracted edge features.
  • A compound loss combines MSE and multi-scales perceptual loss to address over-smoothing during training.

II. RELATED WORK

Related low-dose CT denoising methods mainly use encoder-decoder, fully convolutional, GAN-based, or multi-model architectures. The paper positions EDCNN as a single-model fully convolutional approach.

  • Network structure: Existing low-dose CT denoising networks are broadly categorized into encoder-decoder, fully convolutional, and GAN-based methods.
  • Network structure: Encoder-decoder models use convolutional encoders, deconvolutional decoders, and skip connections to denoise LDCT images.
  • Network structure: Fully convolutional models obtain denoised images through successive convolutional layers and may use dilation, residual, or conveying-paths connections.
  • Network structure: GAN-based methods train a generator and discriminator adversarially, while the generator can serve as the denoiser during testing.
  • Network structure: Some alternatives use cascaded or parallel multi-model architectures, whereas this paper targets efficient denoising with a single model.

1) Per-pixel loss:

Low-dose CT denoising commonly uses pixel-level, perceptual, adversarial, or specialized losses, each emphasizing different image properties and introducing distinct limitations.

  • Per-pixel loss: MSE and L1 losses compare output and target images at the pixel level but do not describe image structural information.
  • Perceptual loss: Perceptual loss compares images in feature space and can preserve details, but may introduce cross-hatch artifacts.
  • Adversarial and composite loss: GAN-based denoising uses adversarial loss to capture structural information and generate more realistic images.
  • Adversarial and composite loss: Specialized composite losses can combine adversarial, MSE, and edge incoherence terms for low-dose CT denoising.

III. METHODOLOGY

EDCNN combines an edge enhancement module with a densely connected convolutional architecture for low-dose CT denoising. Its trainable Sobel operators adaptively extract edge information before combining it with the input.

  • The proposed EDCNN methodology includes an edge enhancement module, an overall model structure, and a training loss.
  • Edge enhancement Module: The edge enhancement module directly acts on the input image before the rest of the network processes it.
  • Edge enhancement Module: Trainable Sobel operators introduce an adjustable Sobel factor that extracts edge information at different intensities.
  • Edge enhancement Module: The module uses four operator directions and applies their feature maps to enrich the input information.
  • Edge enhancement Module: It stacks the extracted edge feature maps with the input low-dose CT image along the channel dimension.

B. Overall Network Architecture

EDCNN combines an edge-enhancement module, dense connections, residual reconstruction, and a compound MSE–perceptual loss for low-dose CT denoising.

  • EDCNN comprises an edge enhancement module followed by eight convolution blocks, using 32 trainable Sobel operators.
  • Dense skip connections convey edge-enhanced features to each convolution block to retain image details.
  • The final convolution output is added element-wise to the original low-dose CT image to produce the denoised result.
  • The denoising objective maps an LDCT input toward an NDCT target through a noise-reduction function F.
  • The compound loss combines pixelwise MSE with multi-scales perceptual loss to address the blur and over-smoothing associated with MSE alone.
  • The perceptual component uses frozen ImageNet-trained ResNet-50 convolutional features, while wp balances the two loss components.

IV. EXPERIMENTS AND RESULTS

The experiments evaluate the proposed model’s low-dose CT noise-reduction performance using a defined experimental configuration and reported results.

  • The section describes the dataset, experimental configuration, and evaluation of the model’s noise-reduction performance.

A. Dataset

Experiments use paired quarter-dose and normal-dose CT images from the NIH AAPM-Mayo Clinic dataset, with patient-level splitting and specified training procedures.

  • The dataset contains paired 512x512 synthetic quarter-dose LDCT inputs and normal-dose NDCT targets from 10 patients.
  • Nine patients are used for training and one patient for testing.
  • The implementation initializes convolution layers randomly, initializes Sobel factors to 1, and sets wp to 0.01.
  • Training randomly crops four 64x64 patches per image, uses batches of 128 patches, AdamW, a 0.001 learning rate, and 200 epochs.
  • The fully convolutional model accepts 512x512 LDCT images during testing and directly outputs denoised results.

C. Results

EDCNN is evaluated against REDCNN, WGAN, and CPCE using quantitative metrics, visual comparisons, and a blind reader study. It achieves strong overall balance between noise suppression and structure preservation, while visual results expose limitations of relying on PSNR and RMSE alone.

  • EDCNN achieves the best or suboptimal results on every reported criterion, balancing per-pixel and structurewise performance.
  • REDCNN achieves the highest PSNR and lowest RMSE, but its output is visibly over-smoothed with blurred structures and lost details.
  • EDCNN and CPCE both preserve subtle CT structures, while EDCNN suppresses noise more effectively than CPCE.
  • The blind reader study evaluates noise reduction, structure preservation, and overall quality across 20 groups of denoised results.
  • EDCNN balances noise reduction and structure preservation and receives a high score for overall image quality.

D. Ablation Study

The ablation study progressively adds dense connections and the edge enhancement module to a basic convolutional model. Each addition improves PSNR, while the edge enhancement module also accelerates convergence and the complete EDCNN performs best on the reported metrics.

  • The study compares a basic model, a dense-connection variant, and EDCNN by adding components sequentially under the same MSE-based training strategy.
  • PSNR increases continuously when dense connections and the edge enhancement module are added.
  • The edge enhancement module accelerates the model’s convergence process.
  • The complete EDCNN achieves the best PSNR, SSIM, and RMSE results among the compared model structures.

2) Models of Perceptual Loss:

The paper compares ResNet-50 and VGGNet-19 as perceptual-loss feature extractors. ResNet-50 produces visually less grainy outputs that are closer to the NDCT target, motivating its use in EDCNN.

  • The perceptual-loss experiment compares ResNet-50 and VGGNet-19 using features from each model’s last convolutional layer.
  • ResNet-50-based outputs have smaller noise graininess and appear closer to the NDCT image than VGGNet-19-based outputs.
  • The paper selects ResNet-50 for its perceptual loss because it is described as having stronger feature-extraction ability than VGGNet.

3) Multi-Scales Perceptual Loss:

The paper evaluates multi-scales perceptual-loss configurations and selects four ResNet-50 stages because increasing the stages makes denoised texture closer to the NDCT image. Visual comparisons further show that compound loss produces clearer results than single losses.

  • Four ResNet-50 stages are used to compute the multi-scales perceptual loss after comparing S-4, S-43, S-432, and S-4321 configurations.The loss computes MSE on extracted feature maps and averages the stage-specific losses.
  • Increasing the number of feature-map stages makes denoised texture closer to the NDCT image.
  • The visual experiment compares EDCNN models trained with single MSE, single multi-scales perceptual, and compound losses under otherwise identical training conditions.
  • MSE-based EDCNN is over-smoothed and loses detail, while compound-loss EDCNN is slightly clearer than the perceptual-loss result and avoids its visible artifacts.
  • The conclusion reports better overall performance for EDCNN, including its compound-loss design, compared with previous models on the Mayo dataset.
Loading 2011.00139v1…