A computer-aided diagnosis system for brain magnetic resonance imaging images using a novel differential feature neural network

https://doi.org/10.1016/j.compbiomed.2020.103818Get rights and content

Highlights

  • More than 10,000 brain MRI images are employed in the proposed system.

  • A novel DFM block is proposed to detect the structural features of brain images.

  • SE blocks are combined with DFM blocks to form a DFNN.

  • The proposed system can achieve state-of-the-art performance.

Abstract

To improve the performance of brain tumor diagnosis, numerous automatic brain tumor diagnosis systems that use machine learning technologies have been proposed. However, most current systems ignore the structural symmetry of brain magnetic resonance imaging (MRI) images and regard brain tumor diagnosis as a simple pattern recognition task. As a result, the performance of the current systems is not ideal. To improve the performance of the brain tumor screening process, an innovative differential feature map (DFM) block is proposed to magnify tumor regions, and DFM blocks are further combined with squeeze-and-excitation (SE) blocks to form a differential feature neural network (DFNN). First, an automatic image rectification method is applied so that the symmetry axes of brain MRI images are approximately parallel to the perpendicular axis. Moreover, a DFNN is constructed to classify the brain MRI images into two categories: “abnormal” and “normal”. The experimental results show that the average accuracy of the proposed system on two databases can reach 99.2% and 98%, and the introduction of the proposed DFM block can improve the average accuracy on these two databases by 1.8% and 1.3%, respectively, which indicates that the proposed DFM block can improve the performance of the brain tumor screening process.

Introduction

In general, a brain tumor is a collection of abnormal brain cells [1]. The uncontrollable division of abnormal cells can increase the intracranial pressure and thus damage the normal brain tissues [1]. Since the brain is one of the most vital organs [1], brain tumors that can damage normal brain tissues are among the most dangerous diseases. It has been reported that brain tumors result in 3.24 deaths per 100,000 persons on average [2]. To improve the survival rate of patients, brain tumors need to be accurately detected at an early stage. Generally, brain MRI images are analyzed manually in most hospitals. However, radiologists must analyze thousands of images per day, and such tedious reading work can lead to misdiagnoses [3]. As a result, manual diagnosis is inefficient and inaccurate [4]. In fact, normal MRI slices occupy the majority of the brain images. Thus, an automatic system that labels the 2D MRI slices with “normal” and “abnormal” can improve the efficiency and accuracy of the brain tumor diagnosis process since radiologists need only to be concerned with abnormal slices.

To improve the performance of the brain tumor diagnosis process, numerous automatic systems based on machine learning methods have been proposed [[5], [6], [7], [8]]. In 2016, Subashini et al. developed an automatic brain tumor grading system. Shape, intensity and texture features were firstly extracted from 200 brain MRI images; then, a support vector machine (SVM) was adopted as the classification model, and this study reported an accuracy of 91% [1]. In the same year, Yang et al. developed an automatic brain tumor screening system. Wavelet-energy features were firstly extracted from 90 T2-weighted brain MRI images; and then biogeography-based optimization was combined with an SVM to train the classifier; this method can detect abnormal brain MRI images with excellent performance [5]. In 2017, Gupta et al. proposed a non-invasive computer-aided diagnosis system for brain tumors. Textural and shape features were extracted from 1100 brain MRI images, and an SVM was applied to construct the classifier. More than 98% accuracy was reported in this research [6]. In 2018, Mohsen et al. proposed an automatic system to classify 66 brain MRI images into 4 categories, namely, normal, glioblastoma, sarcoma and metastatic bronchogenic carcinoma tumors. Fuzzy C-means (FCM) was firstly utilized to segment the brain regions [7], and then a discrete wavelet transformation (DWT) and principal component analysis (PCA) were implemented for feature extraction and reduction, respectively. Finally, a deep neural network was applied for classification; the accuracy of this system reached 96.97% [8]. Although automatic brain tumor diagnosis systems based on traditional machine learning methods can achieve excellent performance on test samples, these systems rely on manual feature extraction, and the generalization of these systems is not ideal [8]. Over the past decade, deep learning technologies have been widely researched. Compared with traditional machine learning methods, deep learning technologies are more efficient because deep learning technologies can automatically extract features. Moreover, numerous studies indicate that the generalization of deep learning technologies is more excellent than machine learning methods [9]. To further improve the performance of the brain tumor diagnosis process, numerous automatic systems based on deep learning technologies have been proposed for efficient brain tumor detection. In 2019, Talo et al. proposed an automatic system to detect brain abnormalities. First, data augmentation techniques, including flipping and rotation, were applied to enlarge the dataset; then, a ResNet-34 was utilized to train the classifier. The recognition rate of this model reached 100% on 613 brain MRI images [3]. In the same year, Anaraki et al. developed an automatic system for glioma grading. 1600 original brain images were employed in this system, and a convolutional neural network combined with a generic algorithm was applied to construct a classifier; the accuracy of the proposed system reached 94.2% [10]. In addition, Han et al. proposed a conditional PGGAN-based data augmentation in 2019 to improve the performance of the brain metastases detection task, and the results indicated that this data augmentation technique can boost the diagnostic sensitivity by 10% [11].

To some extent, existing systems based on deep learning technologies can contribute to improving the efficiency of the brain tumor diagnosis process. However, some limitations remain. (1) Due to the expensive labelling cost and lack of brain MRI images, small databases are employed in most current systems. Generally, no more than 5000 original MRI images are employed in most current systems; thus, the generalization of these systems is not ideal. (2) Most current systems ignore the structural characteristics of the brain and regard the task of brain tumor diagnosis as just a simple pattern recognition application. In fact, a normal brain image is approximately bilaterally symmetric, and this characteristic of the brain can contribute to detecting brain abnormalities.

To overcome the drawbacks of current systems, an automatic computer-aided diagnosis system that can classify the brain images into two categories, namely, normal and abnormal, is proposed. The main contributions of the proposed system are as follows:

  • (1)

    A large database consisting of more than 10,000 brain MRI images is employed to construct a classifier for the first time. Therefore, the experimental results of the proposed system are convincing, and the generalization of the proposed system is excellent.

  • (2)

    Inspired by the symmetrical structure of brain MRI images, a DFM block is proposed to magnify the tumor characteristics and thus improve the performance of deep neural networks.

  • (3)

    To generate an adaptive response for each feature map, the DFM blocks are combined with SE blocks to form a DFNN that can further improve the performance of the brain tumor diagnosis process.

The rest of this paper is organized as follows: Section 2 introduces an automatic image rectification method and the proposed DFNN in detail. Section 3 describes the experimental results. Finally, the conclusions and a discussion of future work are presented in Section 4.

Section snippets

Automatic image rectification and standardization

The construction of the DFNN is based on the assumption that brain images are approximately bilaterally symmetrical. However, there are angles between the symmetry axes and the vertical line in original brain MRI images; thus, it is essential to apply image rectification before constructing DFNN so that the symmetry axes are parallel to the vertical line. The main procedures of image rectification are as follows:

  • Step 1

    Apply image binaryzation to the original brain images by Otsu's thresholding [6]

Databases and evaluation metrics

To verify the performance and generalizability of the DFNN, two databases are employed in this paper. Database I is provided by Shengjing Hospital of China Medical University and the First Hospital of Qiqihaer City, and Database II is obtained from a public database (TCGA-LGG) [22]. Database I consists of 10,678 T2-weighted brain MRI images from 650 patients and 120 normal persons, including 6382 normal images and 4296 abnormal brain images. In addition, 2056 brain images, including 1198 normal

Conclusion and future work

An automatic brain tumor screening system based on deep learning method is proposed in this paper. The main contributions of this system are as follows:

  • (1)

    Compared with most current systems, larger databases are employed, as more than 10,000 original images are utilized to train and test the classifier for the first time. Therefore, the generalization of our proposed system is better than that of previous systems.

  • (2)

    A DFM block is proposed to detect the structural features of brain MRI images, and

Acknowledgements

This work was sponsored by the National Key R&D f of China [grant number 2017YFB1302802]; the National Natural Science Foundation of China [grant number 61703394]; the Special Fund for High-level Talents (Shizhen Zhong Team) of the People's Government of Luzhou-Southwestern Medical University and the National Natural Science Foundation of China [grant number 61821005].

References (28)

  • G. Song et al.

    A noninvasive system for the automatic detection of gliomas based on hybrid features and PSO-KSVM

    IEEE Access

    (2019)
  • G. Yang et al.

    Automated classification of brain images using wavelet-energy and biogeography-based optimization, Multimedia Tools and Applications

    Multimed. Tool. Appl.

    (2016)
  • C. Militello et al.

    Gamma Knife treatment planning: MR brain tumor segmentation and volume measurement based on unsupervised Fuzzy C-Means clustering

    Int. J. Imag. Syst. Technol.

    (2015)
  • A.K. Anaraki et al.

    Magnetic resonance imaging-based brain tumor grades classification and grading via convolutional neural networks and genetic algorithms

    Biocybern. Biomed. Eng.

    (2018)
  • Cited by (28)

    • CSAUNet: A cascade self-attention u-shaped network for precise fundus vessel segmentation

      2022, Biomedical Signal Processing and Control
      Citation Excerpt :

      Next, the SAM, coarse and fine segmentation models will be introduced in detail. Recently, numerous researches have verified that attention mechanisms can improve the performance of image segmentation tasks [19]. In fact, the main idea of attention mechanisms is to obtain the spatial attention maps and channel weights by introducing self-learning neural networks.

    • AMF-Net: An adaptive multisequence fusing neural network for multi-modality brain tumor diagnosis

      2022, Biomedical Signal Processing and Control
      Citation Excerpt :

      In fact, the samples in Databases I and II are stored as RGB images, and the T1-C, T2-W and T1-W sequences form the red, green and blue channels of these RGB images, respectively. In fact, as shown in Fig. 2, the normal T2-W brain images are approximately horizontally symmetrical, and the existence of tumors can damage the symmetry of brain images [12]. Therefore, the horizontal differential images can detect the structural changes in the brain and roughly locate the brain tumor regions.

    • Automatic surface defect segmentation for hot-rolled steel strip using depth-wise separable U-shape network

      2021, Materials Letters
      Citation Excerpt :

      Therefore, an efficient surface defect segmentation method is required. Since the fully convolutional neural network (FCNN) was proposed in 2015, more and more researches have verified that the FCNN can achieve state-of-the-art performance on image segmentation tasks [4]. Compared with the conventional machine learning methods and patch-based convolutional neural networks, FCNN is more efficient and accurate, as FCNN is an end-to-end model that can directly model the image context [5].

    • GammaNet: An intensity-invariance deep neural network for computer-aided brain tumor segmentation

      2021, Optik
      Citation Excerpt :

      After data augmentation, the brain slices are reshaped to the standard size 224 × 224. In fact, the feature maps in deep convolutional neural networks can be regarded as images, which can be mapped to their corresponding input images [1]. It indicates that the gamma correction can be applied to adjust the intensity of feature maps.

    View all citing articles on Scopus
    View full text