Computational Intelligence and Machine Learning Approaches in Biomedical Engineering and Health Care Systems

Convolutional Neural Network for Denoising Left Ventricle Magnetic Resonance Images

Zakarya Farea Shaaf¹, Muhammad Mahadi Abdul Jamil¹, *, Radzi Ambar¹, Mohd Helmy Abd Wahab¹

¹ Biomedical Engineering Modelling and Simulation Research Group, Department Of Electronic Engineering, Faculty of Electrical And Electronic Engineering, Universiti Tun Hussein Onn Malaysia, 86400 Parit Raja, Batu Pahat, Johor, Malaysia

Abstract

Medical image processing is critical in disease detection and prediction. For example, they locate lesions and measure an organ's morphological structures. Currently, cardiac magnetic resonance imaging (CMRI) plays an essential role in cardiac motion tracking and analyzing regional and global heart functions with high accuracy and reproducibility. Cardiac MRI datasets are images taken during the heart's cardiac cycles. These datasets require expert labeling to accurately recognize features and train neural networks to predict cardiac disease. Any erroneous prediction caused by image impairment will impact patients' diagnostic decisions. As a result, image preprocessing is used, including enhancement tools such as filtering and denoising. This paper introduces a denoising algorithm that uses a convolution neural network (CNN) to delineate left ventricle (LV) contours (endocardium and epicardium borders) from MRI images. With only a small amount of training data from the EMIDEC database, this network performs well for MRI image denoising.

Keywords: Deep learning, Image denoising, Image processing, Left ventricle, Neural network.

* Corresponding author Muhammad Mahadi Abdul Jamil: Biomedical Engineering Modelling and Simulation Research Group, Department Of Electronic Engineering, Faculty of Electrical And Electronic Engineering, Universiti Tun Hussein Onn Malaysia, 86400 Parit Raja, Batu Pahat, Johor, Malaysia; E-mail: [email protected]

INTRODUCTION

Medical image analysis in radiology is critical in the healthcare system for detecting and diagnosing the disease at an early stage. Computed tomography (CT), ultrasound (US), positron emission tomography (PET), and Magnetic resonance imaging (MRI) are the most commonly used medical imaging tools. Because of its advantages over other imaging techniques, the MRI tool is widely used in clinical imaging. The MRI imaging technique uses contrast to create

diagnostic images by combining many pulse sequences. MRI also has distinct parameters such as a strong magnetic field, imaging planes, and dimensions [1]. Furthermore, cardiac MRI is one of the most effective techniques for estimating clinical parameters such as myocardium mass, ventricular volumes, stroke volume, and ejection fraction [2].

Medical image quality is still noised due to the imaging condition process and various patients, resulting in a lower resolution. As a result, improving image quality is critical for disease detection and prediction, particularly in the early stages of cardiovascular disease. There are two important factors. Image reconstruction, which is based on an algorithm that creates 2D and 3D images of an object, and image processing, which uses algorithms to improve image quality, remove noise, and detect regions of interest (ROIs), are both used in medical imaging [3, 4].

Many methods for denoising MRI images have been proposed in the literature, including spatial domain approaches, statistical techniques, transform domain methods, and filtering techniques [5]. Currently, a conventional denoising method, such as the block matching 3D (BM3D) filter, is being introduced [6]. The BM4D technique was developed by Foi et al. [7], who extended the BM3D filter to volumetric data. However, neither the BM3D nor the BM4D filters can be applied to varying image contents [8]. Several novel learning methods, such as neural network-based techniques [9-11], have recently been proposed to over come this limitation. With the current development in deep learning architecture, several models, including convolutional denoising autoencoder (CNN-DAE) [12], residual learning (RL) of deep convolutional neural network (DnCNN) [13], and generative adversarial network (GAN), have shown promising results for medical image denoising.

DENOISING MRI IMAGES USING CONVOLUTIONAL NEURAL NETWORKS

A significant amount of work has been done in the medical image analysis field to denoise images. Denoising is essential in image processing to improve segmentation and classification accuracy. Rician and Gaussian noise is the most common types of noise in MRI images [14]. An efficient algorithm for denoising MR images aims to minimize noise while retaining the image's useful features. The most important metric in processing a diagnostic image is edge preservation. As a result, the denoising algorithms must be robust to reduce noise effectively.

For denoising MRI images, various filters are proposed. A median filter [15] is a non-linear low pass filter that reduces unpredictable noise. Weiner filters [16] are used to reduce the gap between filtered and preferred output. The Gaussian filter [17] is used to remove blur noise. In contrast, the mean filter [18] replaces each pixel with pixels' calculated mean to reduce the intensity variation between two pixels using a convolution process. The wavelet filter uses an energy compaction feature to denoise the image. Also, filtering MRI images can be done through noise reduction, re-sampling, and interpolation. The selection of the applied filter is based on the type and amount of noise.

Linear and non-linear image filters are the two types of image filters. Linear filters have been used to eliminate spatial noise but without retaining image textures [19]. Mean filters have been proposed for reducing Gaussian noise, but they over-smooth images with high noise. A Wiener filter was used to address this issue, but it still blurs sharp edges [20]. Non-linear filters such as median and weighted median filters are used to eliminate noise.

Convolutional neural networks (CNNs) have recently demonstrated remarkable performance in image processing tasks such as image denoising [21] and image super-resolution [22]. Zhang et al. [21] created a noise removal model called fast flexible denoising CNN (FFDNet) that can handle white Gaussian noise. Recently, Jiang et al. [23] used the VGG [24] network with ten layers of CNN for MRI denoising. Tripathi and Bag [25] developed a CNN for MRI denoising, with the network employing an encoder-decoder structure to retain important image features while excluding unwanted ones. Furthermore, several methods for denoising medical images using CNN have recently been developed, as shown in Table 1. Also, the review paper [26] summarized numerous methods for using CNN in image filtering.

Table 1 Previous related works for denoising medical images using CNNs.

CONVOLUTIONAL NEURAL NETWORK FOR DENOISING

This section introduces a wide network regarding the number of convolutional layers and filters. Furthermore, the network demonstrates a sequence of MR images using the CNN model; the design and implementation of this network are thoroughly explained in the following subsections.

Building Blocks of Convolutional Neural Network

CNN is the prominent architecture of deep learning [36, 37] that is used for medical image processing. The input for the network is organized in a grid structure and then processed across layers that maintain the spatial relationship in the data. A CNN consists of several layers such as convolutions, activation, and pooling layers with a fully connected layer to compute the final outputs at the end. Fig. (1) depicts the architecture diagram for the designed model.

Fig. (1))

A diagram of network layers architecture.

1. Convolutional layers: The previous layer's operations are convolved with filters that have small parameters, usually of size 3 × 3, stored in a tensor W(j,i) Where j is the number of filters and i is the number of layers.

2. Activation layer: the convolutional layer's feature maps are fed into non-linear activation function, frequently rectified linear units (ReLUs), to produce new tensors called feature maps.

3. Pooling layers: the feature maps are pooled in a pooling layer that takes small grid regions as input and provides single numbers for each area.

4. Batch normalization (BN) layer: this layer is typically stated after the activation layer, providing normalized activation maps to regularize the network and speed up the training.

Network Architecture

As shown in Fig. (2), the network composition consists of an input layer, one convolutional layer with a rectified linear unit (ReLU), 17 convolutional layers with batch normalization (BN), and ReLU, one convolutional layer, and a regression output layer.

Fig. (2))

Network architecture for denoising MR images.

To denoise the left ventricle (LV) MR image from Gaussian noise, the network architecture consists of 59 layers of depth. As shown in Fig. (3), 64 filters of size 3 × 3× 1 are applied to a noisy image in layer 1 to produce 64 feature maps, where 3 × 3 is the height by the width of convolution applied to the input image, and 1 is the channel for image gray-scale. The layer is followed by a ReLU activation function to obtain non-linearity. Then, for each hidden layer (layers 2-57), 64 filters of size 3 × 3 × 64 are applied to create 64 feature maps, which are then activated using the BN and ReLU functions. The final convolution layer (layer 58) has 64, 3 × 3 filters, followed by a final regression layer that generates the denoised image.

Fig. (3))

Model of convolutional neural network for denoising MR images.

Implementation Platform

The experiments in this work are carried out in MATLAB version R2020a to process and denoise a sequence of MRI images. After reading all images, zero arrays are created to store the image sequence and apply gaussian noise before applying DnCNN to filter images. Fig. (4) depicts the experiment's implementation procedures.

Fig. (4))

Experimental procedures.

Data Description

The automatic evaluation of myocardial infarction from the delayed-enhancement cardiac MRI (EMIDEC) dataset is publicly available with clinical information for 150 patients, with the corresponding ground truth. The MRI images are composed of a sequence of DE-MRI scans in short-axis orientation that covers the heart's left ventricle (LV). For both input images covering the LV and the associated ground truth, all input images are stored in separate folders for each case in neuroimaging informatics technology initiative (NIfTI) format. Also, a text file containing clinical information for each case is included in the same folder [38]. In this experiment, only four patients' datasets were used.

RESULT AND DISCUSSION

Simulated data were taken from the EMIDEC web (http://emidec.com/). The experiments were conducted in MatLab (version R2020a) to denoise the MR image of the left ventricle (LV). The noise applied was a Gaussian noise with a range (0.001-0.01). Fig. (5) shows the processing of images denoising with their histograms, and Fig. (6) shows the same processing to denoise the contours of these images.

Fig. (5))

Processed images of the LV denoising MR images, original image (left), noisy image (middle), and denoised image (right).

Fig. (6))

Processed contours of the LV denoising MR images, original image (left), noisy image (middle), and denoised image (right).

Quantitative Measurements

To assess model performance, quantitative measures such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were used. The following are the definitions of the PNSR and SSIM:

where RMSE is the root-mean-square error between the denoised image and free-noise image,

where μx and μy are the data means for x and y , respectively, C1 and C2 are constants, σxy are x and y covariance, σx and σy are the