BLCov: A novel collaborative-competitive broad learning system for COVID-19 detection from radiology images.

With the global outbreak of COVID-19, there is an urgent need to develop an effective and automated detection approach as a faster diagnostic alternative to avoid the spread of COVID-19. Recently, broad learning system (BLS) has been viewed as an alternative method of deep learning which has been applied to many areas. Nevertheless, the sparse autoencoder in classical BLS just considers the representations to reconstruct the input data but ignores the relationship among the extracted features. In this paper, inspired by the effectiveness of the collaborative-competitive representation (CCR) mechanism, a novel collaborative-competitive representation-based autoencoder (CCRAE) is first proposed, and then collaborative-competitive broad learning system (CCBLS) is proposed based on CCRAE to effectively address the issues mentioned above. Moreover, an automated CCBLS-based approach is proposed for COVID-19 detection from radiology images such as CT scans and chest X-ray images. In the proposed approach, a feature extraction module is utilized to extract features from CT scans or chest X-ray images, then we use these features for COVID-19 detection with CCBLS. The experimental results demonstrated that our proposed approach can achieve superior or comparable performance in comparison with ten other state-of-the-art methods.

Introduction

Coronavirus disease 2019 (COVID-19) is an acute respiratory disease caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The global number of COVID-19 infections has increased significantly since December 2019, and we are still suffering from a huge wave of COVID-19 caused by SARS-CoV-2 variants (Hashimoto et al., 2022). COVID-19 can be detected with radiology images for its symptoms are often associated with the symptoms of pneumonia. And the use of image testing for COVID-19 detection is a fast and efficient method in commercial and large-scale use, which can be used to control the spread of COVID-19. The most commonly used technique to detect COVID-19 is real-time reverse transcription-polymerase chain reaction (RT-PCR) (Zu et al., 2020). Chest radiographs, such as computed tomography (CT) scans and chest X-ray images, also play an important role in the rapid diagnosis and early treatment. RT-PCR is widely recognized as a standard method of diagnosis, but there are still several disadvantages. RT-PCR is low sensitivity and requires sufficient expertise to collect viral RNA from patient’s nasopharyngeal swabs. In addition, testing by RT-PCR is time-consuming because it needs strict laboratory conditions (Panday et al., 2021). From this perspective, it is worthy to note that CT scans and chest X-ray images are alternative techniques for detecting COVID-19, which can be used along with RT-PCR (Kanne et al., 2020, Zu et al., 2020). As noted in Pan et al. (2020), COVID-19 pneumonia started showing changes in CT scans ten days after the appearance of related symptoms. In the early pandemic, CT scans and X-ray images were widely used for COVID-19 detection when under bad testing conditions. Also, there are research works suggesting that a combination of clinical image features and laboratory results may help in the early detection of COVID-19 (Shi et al., 2020, Lee et al., 2020, Li and Xia, 2020). Recent studies showed that the noticed changes in chest X-ray images and CT scans were found before the clinical features of COVID-19 appear (Chan et al., 2020). For example, Zhao et al. (2020) discovered that ground-glass opacity (GGO) or mixed GGO was found in most patients, and vascular dilation in the lesions was found in some patients. Zu et al. (2020) informed that round lung opacities were reported in chest CT scans. Hence, major discoveries can also be disclosed according to the diagnostic results of chest X-ray images and CT scans (Li and Xia, 2020).

Automatically diagnosing diseases by machine learning techniques has drawn increasing attention because it can save a large number of medical resources. Machine learning has been applied in various medical fields, such as lung segmentation (Gaál et al., 2020) and pneumonia detection using chest X-ray images (Wang et al., 2017). In recent years, more researches focus on developing AI approaches for COVID-19 detection by machine learning-based methods. Various supervised machine learning methods, such as support vector regression (SVR), ridge regression (RIDGE), and random forest (RF) have been used for the predictions of the number of new cases (Ribeiro et al., 2020). Apart from the traditional machine learning methods, deep learning-based methods have also been applied to COVID-19 detection. For instance, Wang et al. (2020a) proposed a COVID-Net implemented on chest X-ray images, which achieved a promising accuracy in COVID-19 detection. According to different architecture and learning strategy, the COVID-Net was redesigned and a joint learning framework is further proposed based on this backbone in Wang et al. (2020b). Javaheri et al. (2021) designed a CovidCTNet to distinguish COVID-19 infection from common pneumonia and other lung diseases using CT scans. In Nasiri and Hasani (2021), they first employed DenseNet169 to extract features and then performed classification with the XGBoost algorithm. To evaluate the suitability of extreme learning machine (ELM) for COVID-19 classification, a COV-ELM classifier is proposed in Rajpal et al. (2020) using the texture and frequency features extracted from chest X-ray images. Besides, TLCoV is proposed in Das et al. (2021) using CNN, VGG-16, and ResNet-50 individually for detecting COVID-19 from chest X-ray images. CoroDet as a CNN-based model is proposed in Hussain et al. (2021) for automatic detection of COVID-19 using chest X-ray images or CT scans, which can achieve superior performances.

Deep neural networks have achieved significant outbreaks in many applications, however, deep neural networks are of complexity and there are numerous parameters to be calibrated. It is time-consuming for training a deep neural network and a large amount of computing resources need to be consumed. And existing COVID-19 automatic diagnosis models often face the problem of overfitting due to the small size of the dataset, resulting in low accuracy. Therefore, inspired by the collaborative–competitive representation (CCR) mechanism in CCRC (Yuan et al., 2018), we try to design a novel collaborative–competitive representation-based broad learning system for COVID-19 detection, expecting to address the issues above. In this paper, an autoencoder termed CCRAE is first proposed based on collaborative–competitive representation mechanism, and then a novel collaborative–competitive representation-based broad learning system (CCBLS) is proposed by incorporating CCRAE into BLS. In CCBLS, the initial weights can be calculated directly concerning the collaborative and competitive relationship between features. The main contributions of this paper are listed as follows.

To the best of our knowledge, this is the first time that BLS has been used to detect COVID-19 from radiological images such as CT scans or chest X-ray images. The simple structure and faster training speed of BLS make it more competitive when compared to the conventional deep learning methods.

By introducing an effective mechanism of collaborative– competitive into BLS, a novel collaborative–competitive representation-based broad learning system (CCBLS) is first proposed for COVID-19 detection, and our proposed method can achieve competitive performances.

The features are visualized by uniform manifold approximation and projection (UMAP) method. The visualization results demonstrated that the feature representations are distinguishing between the different categories, which can perform well in classification tasks.

The rest of this paper is organized as follows. In Section 2, we briefly review the classic broad learning system, sparse autoencoder, and collaborative–competitive representation mechanism. The proposed CCBLS and the details of our proposed COVID-19 detection approach are given in Section 3. In Section 4, we evaluate the performance of our proposed approach with a series of experiments and report the experimental results and analysis. Section 5 is the conclusion of this paper.

Related works

Broad learning system

Broad learning system (Chen and Liu, 2017) is presented as a flat and incremental learning neural network. A classic structure of BLS is shown in Fig. 1. As we can see, in BLS the original inputs are first transformed into the mapping features, then further obtaining the enhancement features from the mapping features. All these feature nodes will be connected to the ground truth matrix through the output weights. The details of BLS are introduced as follows.

Fig. 1

Illustration of classic structure of broad learning system.

Suppose the training data {} has classes with samples in total. There are groups of nodes in the feature node, and the enhancement nodes contain groups. The th group of mapping feature nodes is calculated as where weighted matrix and bias term are randomly generated, and is the th feature mapping function which is usually a nonlinear function. Next, collecting all the groups of mapping features together, i.e. . By feeding the obtained into enhancement feature nodes, the th group of enhancement feature nodes can be expressed as where is also a nonlinear activation function. Similarly, and are also randomly initialized. These enhancement feature nodes are denoted as . Composing the mapping nodes and enhancement nodes together, we can obtain a new broad feature . And finally, BLS can be expressed as Then the key point of BLS is to get the solution of efficiently. Fortunately, by using the pseudoinverse algorithm, the desired output can be calculated as However, it may be too costly to compute by some common methods because of the large dimension of training data. Alternatively, we can solve this problem by minimizing the least square regression function as where is a regularization parameter. So can be formulated as Specially, we have computed as

Sparse autoencoder and collaborative–competitive representation mechanism

Since BLS generates the mapping features by randomly initializing the connecting weights, in order to overcome the randomness, a sparse autoencoder is adopted to fine-tune the random features and give a sparse representation. As we can see, the random features are generated as equation , where is randomly initialized. And the loss function of SAE can be defined as where is the solution of the sparse autoencoder, is a predefined parameter. To solve the issue of minimization of the loss function Eq. (8), several methods such as fast iterative shrinkage thresholding algorithm (FISTA) (Beck and Teboulle, 2009) and alternating direction method of multipliers (ADMM) (Boyd et al., 2011) are usually adopted. Here, the penalty applied in Eq. (8) can make the autoencoder get more sparse features, which is the same as hierarchical extreme learning machine (H-ELM) (Tang et al., 2015). To extract more dense features, penalty is usually adopted and then we can get another loss function of ELMAE (Kasun et al., 2016), which can be denoted as Eq. (9). It is much easier to compute just by taking the derivative operation, and the solution is termed as where the matrix is generally nonsingular.

The collaborative–competitive representation mechanism is first introduced into CCRC model (Yuan et al., 2018), which makes CCRC model achieve superior performance to CRC model (Zhang et al., 2011). The loss function of the CCRC model is defined as where denotes the label vector, is the training data and notes the number of categories. and are two balancing parameters. and denote the collaborative–competitive mechanism of data. Here, the first term aims to represent the test sample with all training samples collaboratively, while the second term promotes the competitive representation between different classes. Currently, there are several classical methods for feature representation such as sparse representation (SR) (Beck and Teboulle, 2009), collaborative representation (CR) (Zhang et al., 2011), and collaborative–competitive representation (CCR) (Yuan et al., 2018). As we know, sparse representation mainly performs sparse decomposition to obtain the coefficients for representing the features of original data according to the dictionary. When the size of the dictionary is huge, the process of sparse decomposition is usually very slow and the computational complexity of SR is also very large. Collaborative representation usually adopts all the training samples to collaboratively represent the features, nevertheless, it fails to make full use of the localities and discrimination information of data, which may degrade the representation-based classification performance. The collaborative–competitive representation not only considers the collaborative reconstruction of the features but also adds the crucial discrimination among different samples for better feature representation. Based on the advantages of collaborative–competitive representation (Gou et al., 2019, Gou et al., 2021, Li et al., 2020), we consider adopting this mechanism of feature representation and introducing it to our proposed model for better classification performance.

BLS and COVID-19 detection

BLS has been widely used in the fields of image processing and computer vision, such as facial recognition, human action recognition, and image classification. For example, Han et al. (2020) designed a Personalized BLS to recognize personalized expression by taking the emotional information entropy as mapping features. Zhang et al. (2019) introduced the concept of feature blocks to BLS for processing facial data. Dang et al. (2020) proposed a DWnet to recognize human actions, which feeds the extracted spatial–temporal features into BLS. BLS has also achieved good results in other tasks, such as event-based object classification (Gao et al., 2020), Chinese herbal medicine classification (Cai et al., 2019), and robotic material recognition (Wang et al., 2019). BLS is also applied to many other fields, such as medical data analysis, system modeling, and fault detection (Gong et al., 2019, Chu et al., 2020, Liu et al., 2019). From these applications of BLS, we can see that BLS often works along with good feature representations. With good feature representation of sparse autoencoder and the advantages of random vector functional link neural network (RVFLNN), BLS can achieve better performance in classification problem (Chen and Liu, 2017).

In addition to sparse autoencoder, there are other kinds of autoencoders for better feature representation. For instance, collaborative–competitive representation (CCR) can effectively represent the features of data by collaborative–competitive mechanism, and it has been widely applied to image classification. Gou et al. (2019) propose a weighted discriminative collaborative–competitive representation-based (WDCCR) classifier, which not only considers the competitive representation of each class but also enhances the inter-class discrimination to promote the competitive representation of the ground truth. Besides, they designed a locality-constrained weighted collaborative–competitive representation-based classification (LWCCRC) (Gou et al., 2021) to make full use of the localities and discrimination information of data. Li et al. (2020) proposed a sparse and collaborative–competitive representation-based classification (SCCRC), which fuses SRC and CCRC by multiplying their coefficients. The use of CCR results in a noticeable improvement in image classification tasks. Therefore, in BLS, it is reasonable to achieve better classification performance by replacing the sparse autoencoder with CCRAE. In most cases, COVID-19 detection from radiology images can be recognized as a classification problem. For achieving better performance of classification, an effective and efficient automated detection model is urgently proposed for COVID-19 detection. Fortunately, as a flat neural network, BLS can achieve outstanding performance with low time consumption in classification and regression tasks. Moreover, BLS has been applied to the forecast of the trend of COVID-19 infection. Zhan et al. (2021) proposed a hybrid model named RF-Bagging-BLS to predict COVID-19, and the results indicated that the RF-Bagging-BLS model can get a promising performance in timely short-term forecasts. From the analysis, both the advantages of BLS and CCRAE give the potential possibility for using BLS with CCRAE for COVID-19 detection from radiology images.

Proposed method

In this section, we first propose a CCRAE algorithm to better generate the feature representation. Then we propose the CCBLS algorithm based on CCRAE for COVID-19 detection.

Collaborative–competitive Representation-based Autoencoder (CCRAE)

Inspired by the effectiveness of the collaborative–competitive representation (CCR) mechanism, we propose a CCR-based autoencoder to explore the relationship among features and expect a better feature representation in this paper. Assumed that the original data has a form of and denotes the randomly generated feature, where is the number of samples, is the number of original features and is the number of transformed features. The loss function of CCRAE is designed as follows. where represents the Frobenius norm, and . The location of in is the same as . The first term uses all features to represent the original data collaboratively, and the second term is to promote each type of feature to competitively reconstruct the original data. Here, is a regularization parameter for balancing collaborative and competitive representation. When equals 0, CCRAE will reduce to ELMAE, which is also called collaborative representation. As parameter increases, the effect of competitive representation is gradually enhanced. What is more, we can obtain a closed form solution of Eq. (12) by calculating the partial derivative about , which is formulated as where . is defined as

After obtaining , the refined feature representation is decided as

Note that the proposed CCRAE is different from ELMAE. ELMAE just utilizes the collaborative ability of all features. CCRAE not only utilizes the collaborative ability of all features but also the competitive ability of each feature. So that CCRAE can promote features to competitively reconstruct the original data. The proposed CCRAE algorithm is summarized in Algorithm 1.

Collaborative–competitive Representation-based Broad Learning System (CCBLS)

Based on CCRAE, we proposed a novel broad neural network termed CCBLS. The structure of CCBLS is given in Fig. 2. As shown in Fig. 2, we first adopt CCRAE to generate the mapping features. The mapping features in CCBLS are rewritten as follows. where is obtained by CCREAE and is a nonlinear function. The mapping features can be denoted as . Similar to the classic BLS, each group of enhancement nodes is calculated as where is generated randomly. We collect all these enhancement nodes as , where is recognized as the high-level features of . We cannot ensure the quality of feature representation of because the connecting weight is generated randomly. Therefore, we also adopt competitive representation mechanism to fine tune . The transformed features are denoted as , where is randomly initialized, and is the number of transformed features. The fine-tuned weight can be calculated as Eq. (18). where . The location of in is the same as . The above equation can be solved by taking the derivation operation and the solution can be written as where has a similar form as Eq. (14)

Fig. 2

The network structure of CCBLS: CCRAE is on the left, and the connecting weight of enhancement nodes is fine turned with competitive representation (CR) on the right side.

Collecting the mapping feature nodes and the transformed enhancement feature nodes together, CCBLS could be expressed as Let , just as the classical BLS, the weights of CCBLS model can be computed as where is a hyperparameter. The algorithm of proposed CCBLS is summarized in Algorithm 2.

Materials and experiments

In this section, we will describe our proposed method termed BLCov, which consists of a feature extraction module and CCBLS for the COVID-19 detection task. Then we evaluate the effectiveness of the proposed BLCov with extensive experiments. Our proposed model is illustrated in Fig. 3.

Fig. 3

BLCov: Preprocessing, feature extraction, CCBLS-based classification model.

Datasets

The proposed approach is applied to detect COVID-19 from radiology images such as CT scans or chest X-ray images. We adopt three publicly available datasets including ChestX-ray8 (Wang et al., 2017), COVID-CT (Yang et al., 2020), and SARS-CoV-2 (Angelov and Almeida Soares, 2020) to evaluate our proposed approach, in which ChestX-ray8 is a collection of chest X-ray images, both COVID-CT and SARS-Cov-2 are CT scans. The brief descriptions of these datasets are given as follows.

ChestX-ray8 This dataset contains 1125 chest X-ray images, including 500 images with no findings, 500 images with pneumonia, and 125 images with COVID-19. Some sample images of ChestX-ray8 are shown in Fig. 4.

Fig. 4

Samples in the ChestX-ray8 dataset: (a) labeled as COVID-19; (b) labeled as No findings; (c) labeled as Pneumonia.

COVID-CT This dataset consists of 397 CT scans from 171 patients without COVID-19 and 349 CT scans from 216 patients containing clinical findings of COVID-19.

SARS-CoV-2 This dataset includes 2482 CT scans from 120 patients in total, among which 1252 CT scans are labeled as COVID-19 and the other 1230 are non COVID-19. Fig. 5 displays the sample images in COVID-CT and SARS-CoV-2 datasets.

Fig. 5

A chunk of lung CT scans: (a) COVID-19 (COVID-CT dataset); (b) Non COVID-19 (COVID-CT dataset); (c) COVID-19 (SARS-CoV-2 dataset); (d) Non COVID-19 (SARS-CoV-2 dataset).

Table 1 summarizes the datasets used in our experiments. For the preprocessing of these three datasets, we resize the images to the same size 512 × 512, and apply min–max normalization to them to ensure the uniformity of these images.

Table 1

A summary of the adopted datasets.

Datasets	Type	COVID-19	Non COVID-19	Patients
ChestX-ray8	chest X-ray images	125	1000	–
COVID-CT	CT scans	349	397	387
SARS-CoV-2	CT scans	1252	1230	120

Training process

The ChestX-ray8 dataset is used in two phases. The first phase is a two-class classification problem, including classes of COVID-19 and no findings (625 images in total), the other phase is a three-class classification problem, including COVID-19, no findings, and pneumonia (1125 images in total). For the two-class classification task, we apply five-fold cross-validation to obtain the average evaluation metrics, and for the three-class problem, we randomly sampled 80% of the dataset for training, and the remaining 20% are used for testing. COVID-CT dataset is also used in two phases. In the first phase, we split the dataset into two parts: 746 CT scans for training and the other 204 for testing, which refers to the experimental setting in Yang et al. (2020). In the second phase, we conduct five-fold cross-validation on it in our experiments. For the SARS-Cov-2 dataset, we adopt cross-validation just similar to the second phase of the COVID-CT dataset.

We implemented the proposed method using NumPy and scikit-learn in Python 3.9. And the experiments were carried out on Google Colaboratory. As for the parameters of BLS and CCBLS, we adopt the grid search method to obtain the best values. is searched in the range of and is searched in the range of . And in both BLS and CCBLS, is fixed to 2−30, are chosen from the set . The evaluation metrics used to assess the performance of models are: Sensitivity(%), Accuracy(%), Specificity(%), Precision(%), F1-score(%) and AUC(%).

Feature extraction

As noted in Haralick et al. (1973), texture features play an important role in the tasks of image classification. In the stage of feature extraction, two types of features: texture and frequency-based features, are taken into account. What is more, there are three groups of texture features. The first group is the first-order features which are directly extracted from the original image. The first-order features include energy, total energy, entropy, standard deviation, skewness, kurtosis, and so on. Another two groups of texture features are obtained from the gray-level co-occurrence matrix (GLCM) (Zare et al., 2013), and gray-level difference matrix (GLDM) (Kim and Park, 1999) accordingly. Apart from texture features, frequency-based features also play a significant role in the classification of medical images (Varuna Shree and Kumar, 2018). In this work, we apply discrete wavelet transform (DWT) to the images to obtain the frequency features. Finally, we concatenate the texture feature vector of length 54 and the frequency feature vector of length 216 to get a feature vector of size 270 for each image.

Experimental results

For the ChestX-ray8 dataset, the average test accuracy of the two-class problem is 99.68% and the accuracy of the three-class problem is 91.56%. What is more, the confusion matrices of each fold in two-class problem are shown in Fig. 6, the confusion matrix and the receiver operating characteristic curve (ROC) for three-class problem is shown in Fig. 7. A comparison of the proposed approach with the other previous methods for the three-class classification problem can be seen in Table 3, where DarkCovidNet (Chen et al., 2020) and DenseNet169＋ XGBoost (Nasiri and Hasani, 2021) are two deep learning-based methods used in the detection of COVID-19, which directly take the images as inputs. SRC (Beck and Teboulle, 2009), CRC (Zhang et al., 2011), and CCRC (Yuan et al., 2018) are three classical representation-based classification methods, which also use the similar feature representation method as the collaborative–competitive mechanism. As a classical machine learning method, SVM (Hearst et al., 1998) is also used as a compared method. What is more, the above methods take the extracted features as input data which is similar to BLS and CCBLS. As shown in Table 3, CCBLS achieves 1% improvement in Sensitivity, 1.24% in Precision, and 1.34% in Accuracy compared to BLS. Table 2 summarizes the compared methods in our experiments.

Fig. 6

Confusion matrices of testing set in each fold of two-class problem of the ChestX-ray8 dataset.

Fig. 7

The confusion matrix and the receiver operating characteristic curve of three-class problem of the ChestX-ray8 dataset.

Table 3

A comparison of the proposed CCBLS with other methods on the ChestX-ray8 dataset (three-class problem).

Methods	Sensitivity	Specificity	Precision	F1-score	Accuracy
DarkCovidNet (Chen et al., 2020)	85.35	92.18	89.96	87.37	87.02
DenseNet169＋XGBoost (Nasiri and Hasani, 2021)	90.05	93.77	92.59	91.21	89.70
SRC	68.94	83.03	83.00	71.15	70.67
CRC	85.45	90.25	86.93	86.13	84.00
CCRC	85.31	88.81	86.80	86.02	81.78
SVM	90.33	93.07	91.74	90.80	88.44
BLS	92.67	94.13	93.05	92.64	90.22
CCBLS (Proposed)	93.67	94.93	94.29	93.63	91.56

Table 2

A brief description of the compared methods.

Methods	Category	Characteristic
DarkCovidNet	Deep learning-based	X-ray images; Binary accuracy: 98.08%, multiclass accuracy: 87.02%
DenseNet169+XGBoost	Deep learning-based	X-ray images; Binary accuracy: 98.23%, multiclass accuracy: 89.70%
DenseNet169	Deep learning-based	CT scans; Accuracy: 79.5%, F1-score: 76.0%, AUC: 90.1%
ResNet-50	Deep learning-based	CT scans; Accuracy: 77.4%, F1-score: 74.6%, AUC: 86.4%
COVID-Net	Deep learning-based	X-ray images; Sensitivity: 91.0%, PPV: 98.9%
Redesigned COVID-Net	Deep learning-based	CT scans; Accuracy: 77.07% Sensitivity: 74.69%, AUC: 84.13%
SRC	Representation-based	Extracted features; sparse representation
CRC	Representation-based	Extracted features; collaborative representation
CCRC	Representation-based	Extracted features; collaborative–competitive representation
SVM	Machine learning	Extracted features; classical machine learning method
BLS	Broad learning system	Extracted features; classical broad learning system

As for the COVID-CT dataset, Table 4, Table 5 show the results of the two phases respectively. In Table 4, CCBLS achieves 1.02% improvement in Sensitivity, 1.07% in Precision and 0.99% in Accuracy compared to BLS. And in Table 5, we can see that the results of CCBLS are higher than BLS on average and lower on standard deviation.

Table 4

Results of the first phase of the COVID-CT dataset.

Methods	Sensitivity	Precision	F1-score	Accuracy
DenseNet-169 (Yang et al., 2020)	–	–	76.00	79.50
ResNet-50 (Yang et al., 2020)	–	–	74.60	77.4
SRC	77.55	62.30	69.09	66.50
CRC	79.59	68.42	73.58	72.41
CCRC	82.65	72.32	77.14	76.35
BLS	80.61	84.95	82.72	83.74
CCBLS (Proposed)	81.63	86.02	83.77	84.73

Table 5

Results of the second phase of the COVID-CT dataset (mean sd).

Methods	Sensitivity	Precision	F1-score	Accuracy
COVID-Net (Wang et al., 2020a)	57.73 ± 2.94	64.03 ± 3.91	61.09 ± 1.28	63.12 ± 2.09
Redesigned COVID-Net (Wang et al., 2020b)	74.69 ± 3.91	79.48 ± 0.96	77.04 ± 2.17	77.07 ± 1.92
SRC	76.82 ± 5.37	66.36 ± 7.30	71.13 ± 6.29	70.63 ± 7.04
CRC	85.72 ± 3.91	83.38 ± 7.63	84.25 ± 3.85	85.00 ± 3.67
CCRC	86.52 ± 1.74	86.23 ± 3.79	86.30 ± 1.57	87.13 ± 1.64
BLS	87.69 ± 3.99	89.30 ± 2.39	88.41 ± 2.04	89.26 ± 1.75
CCBLS (Proposed)	87.97 ± 1.69	89.82 ± 1.78	88.86 ± 0.80	89.68 ± 0.77

Table 6 shows the results on the SARS-CoV-2 dataset. COVID-Net (Wang et al., 2020a) is a classic deep learning-based framework for COVID-19 detection using chest X-ray images, which has achieved superior performance over other networks pretrained on ImageNet. The redesigned COVID-Net was redesigned based on COVID-Net in Wang et al. (2020b). From Table 6 we can see that both BLS and CCBLS achieve great performance, and CCBLS behaves more stably.

Table 6

Results of five-cross-validation on the SARS-CoV-2 dataset (mean  sd).

Methods	Sensitivity	Precision	F1-score	Accuracy
COVID-Net (Wang et al., 2020a)	70.97 ± 2.37	80.04 ± 2.87	76.03 ± 1.13	77.12 ± 0.98
Redesigned COVID-Net (Wang et al., 2020b)	83.78 ± 0.62	94.58 ± 2.07	88.97 ± 0.91	89.09 ± 1.08
SRC	74.37 ± 1.03	70.11 ± 2.33	72.15 ± 1.24	71.06 ± 0.83
CRC	93.76 ± 2.05	96.56 ± 1.24	95.12 ± 0.81	95.17 ± 0.66
CCRC	93.78 ± 1.93	96.64 ± 1.19	95.17 ± 0.56	95.20 ± 0.48
BLS	97.52 ± 1.27	98.39 ± 0.39	97.95 ± 0.75	97.94 ± 0.74
CCBLS (Proposed)	98.00 ± 1.12	97.93 ± 0.84	97.96 ± 0.69	97.94 ± 0.69

To obtain the overall performance of our proposed method, we have also computed the average class-wise sensitivity, precision, and F1-score as shown in Tables 7, 8, and 9. We also do the evaluation based on time and resource efficiency compared to a baseline BLS as well as deep learning methods on the COVID-CT dataset. The results are demonstrated in Table 10. From Table 10 we can see that the proposed CCBLS has a faster speed and cost less computational resources for it does not need to train any DNN. What is more, from Fig. 8 and Table 10 we can see that the proposed CCBLS has a faster speed than the classical BLS.

Table 7

Class-wise sensitivity, precision and F1-score of the ChestX-ray8 dataset (three-class problem).

Class	Sensitivity	Precision	F1-score
COVID-19	100.0	100.0	100.0
No findings(Non COVID-19)	98.00	85.22	91.16
Pneumonia	83.00	97.65	89.73

Table 8

Class-wise sensitivity, precision and F1-score of the first phase of the COVID-CT dataset.

Class	Sensitivity	Precision	F1-score
COVID-19	81.63	86.02	83.77
Non COVID-19	87.62	83.64	85.59

Table 9

Class-wise sensitivity, precision and F1-score (mean sd) of the COVID-CT and SARS-CoV-2 datasets.

Class	COVID-CT			SARS-CoV-2
	Sensitivity	Precision	F1-score	Sensitivity	Precision	F1-score
COVID-19	91.40 ± 4.40	87.03 ± 1.95	89.11 ± 2.70	97.60 ± 0.84	98.00 ± 0.62	97.80 ± 0.55
Non COVID-19	88.02 ± 1.82	92.22 ± 3.65	90.03 ± 2.14	97.97 ± 0.64	97.57 ± 0.84	97.77 ± 0.55

Table 10

A comparison of time and resource efficiency: the results are averaged on 50 epochs for deep learning methods, for BLS and CCBLS, they refer to the average results of 50 runs.

Methods	Training time (s)	Testing time (s)	CPU memory (GB)
DenseNet169	26.69	9.131	5.003
ResNet-50	21.06	9.104	4.994
BLS	0.4009	0.0053	0.3522
CCBLS	0.2351	0.0064	0.3521

Fig. 8

Speed comparison between BLS and CCBLS under the same parameters on the SARS-CoV-2 dataset: mapping feature nodes , and enhancement feature nodes increases from to .

Visualization

In order to evaluate the effectiveness of CCBLS further, we utilize the uniform manifold approximation and projection (UMAP) (McInnes et al., 2018) method to visualize the features in the whole process of the classification. UMAP is a novel dimension reduction method based on manifold learning, which has strong mathematical foundations. The algorithm implementing this technique is competitive with the state-of-the-art dimension reduction technique t-SNE for visualization quality. What is more, UMAP can preserve the global structure and run faster. UMAP has been widely used in the fields of biological information, materials science, and machine learning.

We train the BLS and CCBLS model on the SARS-CoV-2 dataset, and we also train a BLS model termed ELMAE-BLS using ELMAE. Then we perform dimensionality reduction on the original extracted features and the processed features obtained from BLS and CCBLS with UMAP. The visualization results are illustrated in Fig. 9. In Fig. 9(a), although the original features are in a mixed structure, we can see the blurry clusters, which show the effectiveness of the extracted features. From Fig. 9(b), after being processed in the standard BLS model, the original features become more separable, but there are still many points mixing with the opposite class. As shown in Fig. 9(c) and (d), the data from the same class transformed by CCBLS and ELMAE-BLS both show distinct clustering, and CCBLS produces a better representation of the original input data, which yields a better performance.

Fig. 9

Visualization results using UMAP: (a) Original extracted features; (b) Features transformed by standard BLS; (c) Features transformed by ELMAE-BLS; (d) Features transformed by CCBLS.

Conclusion

In this paper, a collaborative–competitive broad learning system (CCBLS) based approach is proposed for COVID-19 detection from Radiology images such as CT scans or chest X-ray images. In the proposed approach, the features are first extracted from CT scans or chest X-ray images. As an enhanced model of BLS, there are three parts in the learning strategy of CCBLS. In the first part, a collaborative–competitive representation-based autoencoder (CCRAE) approach is proposed to refine the mapping features. Then in the second part, we obtain the enhancement nodes based on the competitive mechanism after getting the mapping features in the first part. In the last part, the transformed features are linked to the ground truths with the output weight. The ChestX-ray8 dataset, COVID-CT dataset, and SARS-CoV-2 dataset are utilized for the experimentation. And the experimental results indicate that the accuracy of our proposed CCBLS can achieve 99.69%, 89.68%, and 97.94% for Chest X-ray8 dataset, COVID-19 CT dataset, and the SARS-CoV-2 CT dataset respectively in detecting COVID-19 and non COVID-19. Besides, our proposed approach can produce an accuracy of 91.56% for Chest X-ray8 dataset in detecting COVID-19, pneumonia, and no findings. Although CCBLS has achieved promising performance in a preliminary study with COVID-19 data, there still exist some limitations in our proposed method. The feature extraction module can be better designed in future work with the professional domain knowledge from experts. It still takes time to help doctors and radiologists in robust COVID-19 detection and assist them to treat severe cases. In future works, a modified feature extraction module will be designed to further explore the effective features in COVID-19 detection from radiology images. Additionally, our proposed method will be extended to more image data with different environmental conditions to validate its generalized performance.

CRediT authorship contribution statement

Guangheng Wu: Conceptualization, Methodology, Validation, Visualization, Writing – original draft. Junwei Duan: Conceptualization, Methodology, Validation, Writing – review & editing, Supervision, Project administration, Funding acquisition.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.