Source-linked AI summary
Medical image denoising using convolutional denoising autoencoders
Lovedeep Gondara
TL;DR
Medical image denoising is important, but deep-learning approaches are limited by large training-data and computational requirements. The paper evaluates convolutional denoising autoencoders and finds that small datasets can support effective denoising, including at very high noise levels. Combining heterogeneous images can increase sample size and marginally improve performance.
Problem
Medical-image noise can impair analysis, while deep-learning denoising methods require large training samples and high computational costs.
Method
The paper evaluates convolutional denoising autoencoders using resized medical images and corrupted datasets, including experiments combining heterogeneous data sources.
Results
CNN DAE outperforms median filtering and NL means across tested noise levels and types, with performance advantages increasing as noise increases.
Takeaways & Limitations
Good denoising performance can be achieved with training datasets as small as 300 samples, and simple networks can partially reconstruct images invisible to humans.
Abstract
from arXiv · showhide
Image denoising is an important pre-processing step in medical image analysis. Different algorithms have been proposed in past three decades with varying denoising performances. More recently, having outperformed all conventional methods, deep learning based models have shown a great promise. These methods are however limited for requirement of large training sample size and high computational costs. In this paper we show that using small sample size, denoising autoencoders constructed using convolutional layers can be used for efficient denoising of medical images. Heterogeneous images can be combined to boost sample size for increased denoising performance. Simplest of networks can reconstruct images with corruption levels so high that noise and signal are not differentiable to human eye.
I. INTRODUCTION
Medical image denoising is needed because acquisition choices and reduced radiation exposure increase noise, while deep models promise strong performance but are often associated with large-data requirements. The paper presents empirical evidence that convolutional stacked denoising autoencoders work with small medical-image datasets and recover signal at very high noise levels.
- Medical images from X-rays, MRI, CT, and ultrasound are susceptible to noise, which can hinder analysis by humans and machines.
- Classical denoising methods include PDE, transform-domain, nonlocal, hybrid, and sparse-coding approaches that estimate the original image from a noisy observation.
- Deep architectures and autoencoders have shown promising denoising results, while convolutional layers exploit spatial correlations and are less restrictive about noise-generation assumptions.
- The paper presents empirical evidence that convolutional stacked denoising autoencoders work well with small sample sizes typical of medical-image databases.
- The paper reports that these methods can recover signal at noise levels where most other denoising methods would fail.
A. Autoencoders
An autoencoder learns to reconstruct its input through an encoded latent representation and a decoding stage. Its capacity and constraints shape whether it learns compressed or sparse representations.
- An autoencoder learns an approximation to the identity function from unlabeled training inputs using backpropagation.
- The encoder maps input x to a hidden representation y, which the decoder maps back to a reconstruction z with the same shape as x.
- Model parameters are optimized to minimize reconstruction error using losses such as squared error or cross-entropy.
- The basic autoencoder architecture encodes the input through a latent layer and reconstructs it at a later layer.
- Fewer hidden units force a compressed approximation, whereas larger hidden layers can still be useful when sparsity constraints are imposed.
1) Denoising Autoencoders:
Denoising autoencoders learn to reconstruct clean inputs from corrupted versions. They can be stacked into deep networks trained layer by layer.
- A denoising autoencoder is trained to reconstruct an input from a noisy version rather than from the original input directly.
- A stochastic corruption process randomly sets inputs to zero, requiring the model to predict the corrupted values.
- Denoising autoencoders can be stacked into deep networks by feeding each lower layer’s output into the current layer.
2) Convolutional autoencoder:
Convolutional autoencoders replace standard autoencoder encoding and decoding with convolutional layers suited to image structure. Shared weights preserve local spatial information during representation and reconstruction.
- Convolutional autoencoders use convolutional encoding and decoding layers and are suited to image processing because they exploit image structure.
- Weights are shared across input locations, helping preserve local spatiality in feature-map representations.
- The convolutional decoder reconstructs inputs from latent feature maps using flipped convolutional weights and channel-specific biases.
- Backpropagation computes the error gradient with respect to the convolutional autoencoder parameters.
A. Data
The study uses two heterogeneous medical-image datasets: mammograms from MMM and cephalometric X-rays from DX, with random examples shown in Fig. 4.
- MMM contains 322 mammograms at 1024 × 1024 resolution, while DX contains 400 cephalometric X-rays at 1935 × 2400 resolution.
- Fig. 4 shows DX X-ray images in rows 1 and 2 and MMM mammograms in row 3.
B. Experimental setup
The experiments resize and corrupt medical images, train a relatively simple CNN denoising autoencoder, and evaluate reconstruction with mean SSIM under fixed basic settings.
- All images were resized to 64 × 64 before modelling, and corruption parameters were varied according to Table I.
- A flattened dataset was corrupted simultaneously across images before modelling with a relatively simple convolutional denoising autoencoder.
- SSIM compares original and coded images through luminance, contrast, and structural components, using means, standard deviations, and covariance.
- Training used 100 epochs and batch size 10 without fine-tuning, and mean SSIM over test images was reported.
C. Empirical evaluation
The empirical evaluation tests CNN DAEs on small and combined medical-image datasets across Gaussian and Poisson corruption. CNN DAEs outperform conventional filters, while higher noise makes reconstruction and convergence more difficult.
- Small-dataset evaluation: 300 images from each dataset were used for baseline comparison, with 22 MMM and 100 DX images reserved for testing.
- Small-dataset evaluation: CNN DAE improved denoising performance over the median filter on small datasets.
- Small-dataset evaluation: The model converged for the evaluated noise levels and sample size, with 50 epochs producing similar results to 100 epochs.
- Combined-dataset evaluation: Combining heterogeneous datasets produced 721 training images and 100 testing images, marginally enhancing CNN DAE performance.
- Noise-level evaluation: At very high Gaussian noise, the network partially generated real images even when the originals were not visible to the human eye, but reconstruction quality declined as noise increased.
- Noise-level evaluation: CNN DAE outperformed NL means and median filtering across tested noise levels and types, with the margin increasing at higher noise.
- Noise-level evaluation: For Gaussian noise with µ = 0, p = 0.2, and σ = 10, the model had not converged after 100 epochs.
V. CONCLUSION
The paper concludes that convolutional denoising autoencoders can efficiently denoise medical images using small training datasets. It also identifies architectural optimization, higher-resolution images, preprocessing, and larger external datasets as future directions.
- Convolutional denoising autoencoders can efficiently denoise medical images using training samples as few as 300.
- Future work will seek an optimal architecture for small-sample denoising and investigate the approach on high-resolution images.
- The authors propose testing preprocessing with SVD or median filters and combining few images with external datasets to increase training sample size.