Broad learning solution for rapid diagnosis of COVID-19

2.1. Transfer learning and feature extraction

Transfer learning (TL) is the process of transferring the learned and trained model parameters to a new model [20]. In image processing, TL mainly uses a large general dataset to obtain fixed weights, and uses the weights to extract data features from the dataset of the object under study. In special cases such as epidemic outbreaks, it is difficult to obtain large public datasets, so TL is popular for small-scale dataset tasks [11]. Wang et al. proposed a chest image classification model using pre-trained deep networks for transfer learning and added an attention module [21]. Li et al. proposed objective metaplasia (GIM) recognition model based on transfer learning, which uses different modules to learn feature weights and then perform feature fusion [22]. In the study by Ahuja et al. four different pre-trained CNN models are utilized to detect COVID-19 [23]. By combining 15 pre-trained CNN architectures with an ensemble method based on transfer learning, Gifani et al. further improved the recognition ability of COVID-19 [24].

Although the pre-trained weights on the ImageNet dataset can effectively extract data features, not all features contribute to the classification work, and the attention mechanism can help extract meaningful features from many features [25]. [26] propose a “Residual Attention Network”, which uses the attention-aware function generated by the attention module for image classification. Zhou et al. combined U-net with attention mechanism, used re-weighted features to better-acquired features, and proved the effectiveness of attention mechanism [27]. [28] propose prior-attention strategy to highlight diseased regions in the lung for COVID-19 diagnosis. [29] innovatively fused CT and X-ray images and demonstrated that convolutional attention modules have a positive effect on diagnostic accuracy. In response to the loss of detail, Xia et al. extract semantic information around the target by introducing multi-scale and context attention modules [30]. Inspired by the above work, we add an attention mechanism and transfer learning to the diagnosis network to improve the network performance for the feature extraction problem with limited datasets.

2.2. Broad learning system

Broad learning (BL) [31] is evolved on the basis of random vector functional-link neural network (RVFLNN). BL with a single layer network structure was proposed as an alternative to DCNN, BL can use fewer neurons to achieve higher test recognition accuracy [32]. BLS has been widely utilized in various fields due to its fast learning ability [33], [34], [35], such as image recognition [36], [37], classification and regression [38], and data modeling [39]. Although BL has achieved the most advanced performance in many applications, the original BLS still has many limitations when dealing with complex tasks. It uses arbitrary continuous probability distribution to randomly generate node weights, which cannot effectively extract the deep features of the image and is not suitable for the pathological diagnosis task of chest CT images. Therefore, in this study, we used convolutional networks to capture more features to make up for the shortage of BL, and then BL was utilized to efficiently diagnose COVID-19.

2.3. Artificial intelligence for COVID-19

Artificial intelligence has made considerable strides in image processing over the past few years. In the field of medical image processing, AI can effectively classify different images to assist diagnosis and reduce the rate of misdiagnosis. [40] proposed a lightweight neural network framework called COVID-Net, which improves the transparency and reliability of predictions and proposes an open benchmark X-ray image dataset. Gozes et al. combine deep models of different dimensions with clinical understanding to enable automated diagnosis of COVID-19 [41]. Shan et al. achieve a precise description of COVID-19 condition by quantifying infected area [42]. Kollias D et al. introduced a new large CT scanning database and used CNN-RNN architecture for COVID-19 diagnosis [43]. In [44], visual features were extracted from the image to distinguish COVID-19 from community-acquired pneumonia and healthy individuals. Fan et al. propose a semi-supervised deep network (Inf-net) to segment pulmonary infection by combining parallel decoder and attention mechanism [45]. It turns out that the convolutional layer in DCNN is very suitable for extracting these features from images.

3.1. Framework

The network framework we proposed is shown in Fig. 1. This system includes two parts, the feature fusion module is used for feature extraction and feature fusion, and we add the spatial attention module to strengthen the target area. The extracted features are then sent to the BLS for definitive diagnosis.

3.2. Feature extraction module based on residual attention mechanism

Feature extraction is the elementary work of image extraction. It can be seen from the aforementioned article that the convolutional layer in DCNN is very suitable for extracting these features from images. Therefore, in our network FA-BLS, for the case where the dataset is difficult to obtain, we use ResNet50 for feature extraction. As shown in Fig. 2, the backbone network ResNet50 [46] is pre-trained on ImageNet to obtain weights with generalization ability, which are used to extract COVID-19 features in the form of transfer learning.

ResNet50 stands for 50-layer residual network, which uses the residual structure of skip connection and implements identity mapping between layers, enabling the network to achieve better performance with an increased number of layers of the convolutional network. Compared with CNN stacked by other convolutional layers, it addresses the problems of insufficient training data and deep network degradation to a certain extent. The ResNet50 convolution module, pre-trained based on a large dataset Imagenet, despite efficient extraction of generalization features, has degraded the performance of the model in the target domain due to the particularities of medical images. The introduction of attention mechanism can provide performance improvements to models by assigning more weight parameters to important features, and using this mechanism feature extraction module could retain more important information and reduce the interference caused by redundant content to the network.

The spatial attention module (SAM) utilizes two pooling operations to generate two 2-dimensional features for channel information of feature maps. SAM is based on the global average pooling channel as well as the global biggest pooling operation, produced two represent different characteristics of figure, the merged by another larger receptive field 7 × 7 convolution of feature fusion, and then the weight graph is generated by operation and combined with the features extracted in the previous step so that the target area can be strengthened. The formula for calculating SAM is:

3.3. Broad learning system

The biggest difference between broad learning and deep learning lies in the number of network layers and the way of data processing. Fig. 3 shows the differences between the two image processing methods. CNN uses a convolution kernel of appropriate size to extract features by calculating pixels of each image one by one. While BL stretches each image pixel into a row, all pixels collectively make up a large matrix for operation, which makes BL’s model complexity greatly reduced compared to DL, and operation speed improved significantly. But single-layer network which is concise and efficient also brings some disadvantages, it cannot extract image deep features like a convolutional neural network, so this research combined CNN with BL to synthesize the advantages of both for better performance of detection tasks.

The structure of the system is shown in Fig. 4. The input, feature node, enhancement node, and output make up the entire network. In this section, the output of the feature extraction layer will be used as the input matrix X of BLS, then $X={[x_1,\dots ,x_n]}^T\in R^{N\times B}$,output $Y={[y_1,\dots ,y_n]}^T\in R^{N\times C}$, N represents the total number of samples, B represent the dimension of samples, C represents the total number of classes.

For n feature maps, each map generates n nodes, then the $i$th feature node can be expressed as follows:

Where $W_{e_i}$ and ${\beta }_{e_i}$ are both randomly generated weights and offsets,Since $\phi$ is a linear activation function, the mapping feature is linear. The sparse autocoding method is used by BLS to optimize the input weights in order to get over the unpredictable nature of random initialization.

$Z^n\equiv [Z_1,\dots ,Z_i]$ is used to symbolize each feature node, and for m enhanced nodes, the $j$th enhanced node can be expressed as:

$\xi$ is a nonlinear activation function where Typically, it can be configured as a hyperbolic tangent function:

All enhancement nodes are represented as$H^m\equiv [H_1,\dots ,H_j]$, therefore, the BLS model can be expressed as the following equation:

Where $W={[Z^n\mid H^m]}^{+}$ is the connection weight between the hidden layer of the broad structure and the output layer. $W_{e_i}$, ${\beta }_{e_i}$, $W_{h_j}$ and ${\beta }_{h_j}$ are all functional nodes randomly generated and fixed throughout the whole process. By solving the ridge regression approximation of ${[Z^n\mid H^m]}^{+}$, $W$ can be easily computed.

We choose L2-norm regularization to solve this optimal problem, which has convexity and good generalization performance. The objective function of BLS is:

Where, $\lambda$ is the regularization coefficient, and $\hat{Y}$ is the prediction result, it is easy to obtain:

The identity matrix is denoted by I in the equation, while $A^T$ represents A permutation and A pseudo-inversion is denoted by $A^{+}$ $={lim}_{x\to 0}{(A^{T}A+\lambda I)}^{-1}{A}^T$, formula (7) can be written as $W=A^{+}Y$.

For the label Y, $y_i$ represents the probability that the system predicts each category, which can be defined as

Finally, the probability output is

4.1. Data description

In this paper, we select three publicly available COVID-19 datasets to obtain more reliable evaluation results (as shown in Table 1). Fig. 5 indicates the CT images of COVID-19 patients and non-infected cases.

• $•$

  CC-CCII [47] is obtained from the China Consortium of Chest CT Image Investigation. All images were divided into novel coronavirus pneumonia (NCP) caused by SARS-CoV-2 virus infection, common pneumonia and a normal control group, 750 with labels are selected for study in this paper.

• $•$

  COVID-CT-Dataset [48] is collected from COVID19-related papers from medRxiv, bioRxiv, NEJM, JAMA, Lancet, etc. It includes 349 CT scans with COVID-19 and 463 CT scans without COVID-19, and a senior radiologist confirmed the validity of the dataset.

• $•$

  SARS-CoV-2 dataset [49] is a public CT scanning dataset constructed by Soares and others, including 1252 CT images infected with COVID-19 and 1230 CT images uninfected with COVID-19, with a total of 2482 CT images. These datas was collected from actual patients in hospitals.

4.2. Evaluation metrics

The comparison of diagnostic test results with the gold standard can reflect the pros and cons of diagnostic methods. Some evaluation indicators are commonly used in medical diagnosis, this paper selects the following popular indicators to evaluate our method: Accuracy(Acc), Sensitivity(Sen) and Specificity(Spec). They are defined as follows. A confusion matrix is represented as in Table 2.

4.3. Implementation details

We did our experiments in Python 3.6, the process runs on a 2.2 GHz Core i5 processor with 4 GB of RAM. All datasets are divided into training and testing datasets according to 8:2, and the final evaluation results are obtained through five-fold cross-validation. Bayesian optimization is employed to optimize the hyperparameters during our technique. The hyperparameters of the BLS principally embrace the quantity of feature windows (N1 = 100), the number of nodes within each feature window (N2 = 8) and the number of enhancement nodes (N3 = 2084).

4.4. Quantitative analysis

In this section, FA-BLS is compared with the classical DCNN method to prove its rapidity. To quantify the improved effect of FA-BLS, the same hyperparameters are used to train the ordinary BLS. Some classic deep CNN, such as VGG16 [50], ResNet50, Xception [51] and Efficientnet [52] also use the same hyperparameters as ResNet50 in feature fusion layer(Batch size = 16, Optimizer = Adam, Activation function = Softmax, Loss function = Binary crossentrop). Meantime, the weights of all methods are obtained by pre-training ImageNet, and the training of deep learning is performed only around the last layer. In Table 3, the classification results of each method under CC-CCII dataset are given. The accuracy of ordinary BLS identification is only 65.3%, far lower than that of classical DCNN, but the accuracy of FA-BLS proposed by us reaches 93.9%, 28.6% higher than that of ordinary BLS in the case of small-scale datasets. DCNN runs a long time period thanks to additional parameters, but our FS-BLS introduces a broad structure to reach high identification accuracy with few parameters for limited datasets.

For ordinary BLS and different deep convolutional neural networks, setting the same hyperparameters as our model cannot show the best performance of the network, and may also produce an overfitting phenomenon. We selected appropriate hyperparameters for each network to more objectively evaluate the performance results of our model as shown in table Table 4, Table 5, Table 6. The deep network has achieved better performance improvement at the cost of training time, but it can be observed from the table that the network improvement of DCNN under the small-scale datasets is still not ideal, just achieving the same level of performance as FA-BLS. A key challenge for deep networks is that the total amount of data in this task is too small to meet the needs of network learning, so even using the most appropriate hyperparameter will not achieve the excellent performance of deep models under large dataset tasks. In contrast to this, the training time of FA-BLS is significantly reduced compared with that of the deep network. It is 130 times faster than the Xception network and 66 times faster than ResNet50 while processing the same data. In the deep model, VGG16 provides better results, as it trains fewer parameters and takes less time to train samples. The proposed model has similar performance compared to the competing model, but when we applied the proposed model to larger populations, the algorithmic speed increase could keep more people from getting sick.

4.5. Effectiveness analysis

To prove the effectiveness of the planned framework, the analysis results of the planned framework and other advanced classifiers used to observe COVID-19 were compared in Table 7, where C-19 denotes “COVID-19”, NC denotes “Non COVID-19”. The reported results for the comparator methods were taken from their papers as we were unable to reproduce their codes. The performance of each model was estimated according to three different variables: accuracy, specificity and sensitivity. At present, most classification researchers seek to improve the performance, but seldom report the model training time in the paper. Therefore, this section does not discuss speed, focusing on the performance comparison of our model with other advanced methods.

The results show that our study can distinguish between COVID-19 positive and healthy patients who are not infected, a potential reason being that our feature fusion module provides robust feature representation. In terms of ACC, the use of multiple datasets also further ensures the credibility of our proposed COVID-19 diagnosis model. It can be seen from the Table 7 that when the total amount of data is similar, the classification accuracy of our method has significantly improved compared with the network using ResNet50 only [9]. In addition, our strategy of combining ResNet50 with BLS and the method of combining two deep models [55] have achieved approximate performance. However, from the discussion of training time in the previous section, it is known that our model has considerable advantages in terms of training speed. The evaluation results of FA-BLS performance with three datasets show that the size of input data has little impact on the performance of this method, as an auxiliary diagnostic tool, our proposed method may serve as a reliable clinical adjunct to COVID-19 diagnosis.

4.6. Broad learning parameter analysis

For additional study the classification performance of FA-BLS, we have a tendency to explore the influence of various variety of functional nodes on the experimental results, as shown in Table 8. It is obvious that with the rise of the amount of nodes, the training time progressively becomes longer, however, it is not the model with more nodes that has higher classification accuracy. The selection of nodes is the determinant of the experimental accuracy.