Attention convolutional neural network for accurate segmentation and quantification of lesions in ischemic stroke disease
Graphical abstract
Advanced methods for image analysis play a critical role in the diagnosis of the ischemic stroke lesion and white matter hyperintensity (WMH). These conditions are associated with presence of lesions in the MRI images. While radiologists have difficult time to distinguish these two types of lesions, computer programs offer much needed help via high-quality segmentation of these images. The segmentation of these two types of lesions is particularly challenging and was covered in several recently published articles. We propose a novel solution for this important problem with the overarching aim to substantially improve the quality of the segmentation. Our solution is a deep residual attention convolutional neural network (DRANet). This network uses several architectural innovations that result in the ability to provide very accurate segmentation. We include comprehensive empirical tests that demonstrate that DRANet outperforms the state-of-the-art segmentation methods. Our article also provides an intuitive explanation for how our network is capable of providing such accurate segmentation.
Introduction
Ischemic stroke is a major cerebrovasular disease that prevents blood from reaching brain regions by directly blocking blood vessels. White matter hyperintensity (WMH), which is also called the signal of the silent cerebral small vessel diseases (CSVDs), is an abnormality in the cerebral tissue Cloonan et al. (2015). WMH lesions are frequently observed in elderly individuals and ischemic stroke patients. The presence of WMH is a risk of future symptomatic ischemic stroke Lewis H et al. (2004); Cho et al. (2015). WMH is also used to monitor ischemic stroke lesion growth Hakan et al. (2008) and to predict post-stroke outcomes Arsava et al. (2011). Accurate identification and segmentation of these two types of lesions would improve the ability of physicians to correctly diagnose patients. However, ischemic stroke and WMH lesions may coexist in some areas of the brain. In addition, these two lesions have similar signals in the MRI images. For example, both lesions appear as hyperintensities in the T2 or fluid-attenuated inversion recovery (FLAIR) MRI, and as hypointensity in the longitudinal relaxation time (T1) MRI. We show two examples of ischemic stroke lesion and WMH on FLAIR MRIs in Fig. 1. Radiologists often struggle to distinguish these two types of lesions.
Radiologists usually delineate the lesions slice by slice in the FLAIR images. Presence of similar lesions and the ambiguity of their boundaries may impede correct segmentation of the ischemic stroke and WMH lesions. In addition, both lesions may have different sizes and shapes, which further complicates the segmentation process. Manual segmentation is a time-consuming work and the quality of manual segmentation depends on the level of experience of radiologists. Computer programs can learn from large collections of images to rival performance of experienced radiologists and do not tire when applied to process many images in a short amount of time. Thus, the development of accurate, concurrent and automatic segmentation for the two types of lesions is highly desirable.
Many methods have been developed for the brain image analysis tasks in the past few years Liu et al. (2017); Kong et al. (2019); Liu et al. (2020). We review methods that are used for the segmentation of the stroke lesions, WMH, and other CSVD lesions. These methods use medical images of the brain as input to segment the target with the help of several different types of algorithms that include k-nearest neighbor (KNN) Steenwijk et al. (2013); Boer et al. (2009), support vector machine (SVM) Lao et al. (2008); Vamsi et al. (2014), random forest (RF) Dadar et al. (2018); Mitra et al. (2014) and various designs of the convolutional neural network (CNN) methods Griffanti et al. (2016); Maier et al. (2017); Zhang et al. (2018); Li et al. (2018); Liu et al. (2019).
Several studies aim to perform the segmentation of the stroke or WMH lesions. These methods can be broadly divided into two main categories: unsupervised methods and supervised methods. The unsupervised methods do not require labeled images during the training process. For instance, Erin et al. (2010) proposed an automated WMH segmentation method based on fuzzy C-means (FCM) clustering and thresholding. Matus et al. (2010) developed an unsupervised method to assess the diagnosis and treatment decisions for the acute ischemic stroke. Nguyen et al. (2015) published a location-sensitive deep network to accurately locate stroke lesions. However, the accuracy of these unsupervised segmentation methods is relatively low. Recently, supervised methods, which require labeled data to train the predictive model, have gained popularity. Many classical supervised machine learning methods were used for the stroke lesion or WMH segmentation tasks. For instance, Mitraa et al. (2014) proposed a supervised RF model to segment ischemic stroke lesions. The authors first predicted the probable lesion volumes. Next, they used RF to extract lesion regions based on multi-modality MRIs followed by the thresholding that generates segmentation. Similarly, Dadar et al. (2018) developed a segmentation model based on the RF model. Lao et al. (2008) proposed the SVM model that was trained on expert-defined white matter lesions.
More recently, supervised CNN-based methods have become popular for the medical image segmentation Yu et al. (2019). These methods incorporate novel types of network modules, such as U-net Ronneberger et al. (2015), residual module Chen et al. (2018), dense module Zhang et al. (2018), dilated module Lessmann et al. (2018), and attention module Jin et al. (2018), to improve predictive performance when compared with classical CNNs and the other supervised machine learning algorithms. For instance, a deep uResNet was used to segment the WMH Guerrero et al. (2017). In uResNet, residual blocks were embedded into the U-net network to improve the speed of convergence and to solve the vanishing gradient problem in the back-propagation flow. Similarly, Chen et al. (2018) proposed a novel deep network which with 25 layers and voxel-wise residual module. This network integrates multi-modal and multi-level MRIs for brain segmentation. Zhang et al. (2018) proposed a deep CNN for the segmentation of the acute ischemic stroke lesion from DWI MRIs. They used weakly labeled subjects for training the predictive model, which limited the predictive performance of the model. Similarly, Li et al. (2018) proposed an ensemble of FCNs for the prediction of WMHs. The development of accurate advanced designs of CNNs for real-world clinical use, including the detection of small-volume lesions in MRI images, remains a challenging problem. Several studies changed the model structure to extract more abundant features. Lessmann et al. (2018) used two consecutive dilated convolutional neural networks to automatic detection of coronary artery, thoracic aorta and cardiac valve calcifications in low-dose chest CT. In addition, Vesal et al. (2018) have developed a deep CNN for the left atrial segmentation. Jin et al. (2018) published a novel design of CNN named RA-UNet to annotate liver and tumor in medical images. Moeskops et al. (2018) released a multi-scale convolutional neural network for brain tissue segmentation. Just this year, Nakarmi et al. (2020) proposed a novel multi-scale deep learning framework for medical image segmentation that combined the deep model and prior knowledge. While these models have achieved promising results for the lesion segmentation task, most of them focus on a single lesion type, and those that can segment multiple lesion types struggle to do it well.
We propose a state-of-the-art deep network DRANet to simultaneously segment the ischemic stroke and WMH lesions from MRIs. We introduce a novel attention module to improve the predictive quality of the proposed network. Our contributions include:
- 1.
Development of an advanced deep CNN architecture that aims to improve predictive performance and that allows for accurate and simultaneous prediction of both lesion types. We extend the popular U-net architecture by inclusion of novel modules, layers and the attention mechanism. Details are explained in Section 2.2.
- 2.
Inclusion of comprehensive empirical tests that compare our novel design with the current solutions using multi-modality MRIs on a modern benchmark data set.
- 3.
Analysis and visualization of the results produced by our network with the goal to explain the benefits of our novel design features.
The rest of this paper is organized as follows. Section 2 introduces the architecture of the proposed deep network and provides a quantitative evaluation of our segmentation network. Section 3 introduces the data used in our study. Section 4 conducts empirical comparative experiments and ablation experiments that investigate the contributions of individual components of our proposed network. Finally, we discuss and summarize this study in Sections 5 and 6, respectively.
Section snippets
Overview of our proposed architecture
The architecture of DRANet is shown in Fig. 2. Fig. 2(b) shows the high-level topology which is inspired by the U-Net architecture Ronneberger et al. (2015). The U-Net has been shown to perform well for samples number limited segmentation tasks. In DRANet, we embed residual blocks into the original topology the U-Net. Residual blocks help to alleviate the vanishing gradient problem. The vanishing gradient problem describes a situation where a deep neural network is unable to propagate useful
Data
The sub-acute ischemic stroke lesion segmentation (SISS) challenge is a sub-task of the ISLES challenge 20151. The segmentation results of SISS were shown to be lacking in quality Maier et al. (2017). The main reason is the similarity of the ischemic stroke lesion to other lesions, especially the WMH lesion. This motivated our work to develop an accurate method to simultaneously segment the ischemic stroke and WMH lesions from MRIs. There are 28 publicly available
Experimental setup
We split the samples into three parts (training, validation and testing) to perform robust training and assessment of our predictive model. In particular, we use the samples in training and validation parts to train and adjust the network (to optimize the quality of the mapping of input images to the outcomes), and the samples in test part to perform comparative assessment with the other algorithms. We emphasize that the test part was used after we completed the parametrization of our model by
Residual and attention mechanisms
DRANet model is composed of the backbone network (the sequence of the convolutional and residual blocks, see Fig. 2(b)) and the attention module. The residual blocks, which are the main component in the backbone network, increase the depth of the model to improve the predictive performance of the segmentation while also combating the vanishing gradient problem. Moreover, we utilize a self-gated soft-attention module to implement the contacting path (Section 2.2 provides details). The attention
Conclusions
We propose a novel deep learning architecture with the attention module, called DRANet. DRANet was designed to accurately and concurrently segment ischemic stroke and WMH lesions in the multi-modality MRIs. The key architectural features of our solution include the use of the residual blocks and the Dice loss function to facilitate effective training of the network, as well as the application of the attention modules to generate high-quality representation of the input images inside the
Credit author statement
Liangliang Liu and Jianxin Wang conceived and designed the study.
Liangliang Liu performed the experiments and wrote the paper.
Fang-Xiang Wu, Lukasz Kurgan and Jianxin Wang reviewed and edited the manuscript.
All authors read and approved the manuscript.
Declaration of Competing Interest
The authors declare that they do not have any financial or nonfinancial conflict of interests
Acknowledgment
The work described in this paper was supported by the National Natural Science Foundation of China under Grants (no. 61828205, no. 61732009, no. 61802442); the 111 Project (no. B18059); and the Hunan Provincial Science and Technology Program (2018WK4001).
References (52)
- et al.
Voxresnet: Deep voxelwise residual networks for brain segmentation from 3d mr images
Neuroimage
(2018) - et al.
Metabolic determinants of white matter hyperintensity burden in patients with ischemic stroke
Atherosclerosis
(2015) - et al.
Bianca (brain intensity abnormality classification algorithm): A new tool for automated segmentation of white matter hyperintensities
Neuroimage
(2016) - et al.
Classification of autism spectrum disorder by combining brain connectivity and deep neural network classifier
Neurocomputing
(2019) - et al.
Computer-assisted segmentation of white matter lesions in 3d mr images using support vector machine
Acad. Radiol.
(2008) - et al.
Fully convolutional network ensembles for white matter hyperintensities segmentation in mr images
Neuroimage
(2018) - et al.
Deep convolutional neural network for accurate segmentation and quantification of white matter hyperintensities
Neurocomputing
(2020) - et al.
Isles 2015 - a public evaluation benchmark for ischemic stroke lesion segmentation from multispectral mri
Med. Image Analy.
(2017) - et al.
Lesion segmentation from multimodal mri using random forest following ischemic stroke
Neuroimage
(2014) - et al.
Evaluation of a deep learning approach for the segmentation of brain tissues and white matter hyperintensities of presumed vascular origin in mri
Neuroimage Clin.
(2018)
A large margin algorithm for automated segmentation of white matter hyperintensity
Patt. Recogn.
Accurate white matter lesion segmentation by k nearest neighbor classification with tissue type priors (knn-ttps)
Neuroimage Clin.
Neuroimaging standards for research into small vessel disease and its contribution to ageing and neurodegeneration
Lancet Neurol.
Tractseg - fast and accurate white matter tract segmentation
NeuroImage
Deep convolutional neural networks for multi-modality isointense infant brain image segmentation
NeuroImage
Severity of leukoaraiosis determines clinical phenotype after brain infarction
Neurology
White matter lesion extension to automatic brain tissue segmentation on mri
Neuroimage
White matter hyperintensity in ischemic stroke patients: It may regress over time
Stroke
Attention-based models for speech recognition
Advan. Neur. Inform. Process. Syst.
Validation of t1w-based segmentations of white matter hyperintensity volumes in large-scale datasets of aging
Hum. Brain Mapp.
Learning normalized inputs for iterative estimation in medical image segmentation
Med. Image Analy.
Automatic segmentation of white matter hyperintensities in the elderly using flair images at 3t
J. Magn. Reson. Imag.
Joint optic disc and cup segmentation based on multi-label deep network and polar transformation
IEEE Trans. Med. Imag.
White matter hyperintensity and stroke lesion segmentation and differentiation using convolutional neural networks
Neuroimage Clin.
Cited by (69)
A feature-enhanced network for stroke lesion segmentation from brain MRI images
2024, Computers in Biology and MedicineDeep learning-based grading of white matter hyperintensities enables identification of potential markers in multi-sequence MRI data
2024, Computer Methods and Programs in BiomedicineA few-shot learning-based ischemic stroke segmentation system using weighted MRI fusion
2023, Image and Vision ComputingMulti-class disease detection using deep learning and human brain medical imaging
2023, Biomedical Signal Processing and Control