2.1 CNN MODEL

LeCun $\textit{et al.}$ proposed the fundamental structure of the convolutional neural network [6], which is an algorithm used widely to classify images, sounds, text, and videos. The CNN has improved performance in recognizing image objects and finding patterns by extracting the features directly from the data [7].

Unlike the general neural network, the CNN can encode the characteristics (horizontal, vertical, and color channels) of image data assuming that the input data is an image and consists of learning weights and bias [8].

Each neuron has an internal operation and a non-linear operation according to its choice. The entire network has a one score function and a loss function on the last layer. The same techniques can be applied to general neural network learning.

The CONV(Convolutional) layer is a key component of a CNN. Receiving input from the input image, this connected area and its weight are calculated, and the output of the CONV layer is interpreted as neurons arranged in three dimensions.

ReLU(Rectified Linear Unit. Activation function) is a layer that applies the following function (1) to the output of the CONV layer.

The reason for applying RELU as an activation function is that it solves the problem of a vanishing gradient better than the sigmoid function.

POOL(Pooling) is a layer that reduces the size to reduce the number of parameters or computations in the network periodically in the middle of the CONV layer in the CNN structure. This layer performs downsampling for each depth and controls the overfit.

The neurons in the Fully-Connected (FC) layer are linked to all the activations in the previous layer. Hence, the FC layer activations multiply the metrics, and add the bias.

Fig. 1 presents the CNN model.

Fig. 1. Unilateral CNN Model Structure.

2.2 Bayesian Optimize and Hyperband

The BOHB is a combination of the Bayesian optimization method and a hyperband. The algorithm produces hyperparameter sets of Bayesian optimization, selects one of them, and progresses the hyperband to find the hyperparameters.

The performance of the deep learning algorithm depends greatly on the configuration of the hyperparameter. Bayesian optimization is used to reduce the unnecessary hyperparameter repetitive navigation to find the optimal hyperparameter faster.

Hyperband algorithms first extract the arbitrary sets within a given range for hyperparameter navigation. The predictive performance after learning was compared by applying the corresponding hyperparameters. The low-performance hyperparameter is removed, and the remaining values are used. The performance is then compared. This process is repeated to obtain the optimum until the last hyperparameter remains.

The estimation of hyperparameters can be made faster with the BOHB than Bayesian optimization because it combines the hyperband and Bayesian optimization methods using stochastic estimation information [9].

In this experiment, the three parameters were explored: learning rate (learning rate), number of learning (epochs), and weight decay (weight attenuation) using the BOHB. This is a preferable method to estimate the hyperparameters in a deep learning algorithm because there are more parameters to decide than traditional machine learning algorithms.

2.3 Influenza Kit

Influenza A and B antigens can be detected qualitatively and rapidly in nasal swab specimens using an ulti med Influenza A/B Antigen Test, a chromatographic immunoassay. The test can provide a rapid differential diagnosis of influenza A and B viral infections [10].

Influenza (commonly known as ‘flu’) is a highly contagious, acute viral infection of the respiratory tract that can be transmitted easily by coughing and the sneezing of aerosolized droplets containing live virus. Influenza outbreaks occur each year, generally during the fall and winter months. Type A viruses are typically more prevalent and are associated with most serious influenza epidemics compared to type B viruses. In contrast, type B infections are usually milder. The Influenza A/B Antigen Test detects the presence of the Influenza A or B antigen or both qualitatively in nasal swab specimens, providing results within 15 minutes. The test uses antibodies specific to Influenza A and B to detect the Influenza A and B antigen selectively in nasal swab specimens [10].

The Influenza A/B Antigen Test is a qualitative, lateral flow immunoassay that detects Influenza A and B nucleoproteins in nasal swab specimens. In this test, an antibody specific to the Influenza A and B nucleoproteins are coated separately on the test line regions of the test cassette. During testing, the extracted specimen reacts with the antibodies to Influenza A or B or both. The mixture migrates along the membrane and reacts with the antibodies to Influenza A or B or both. A positive outcome is illustrated by one or two colored lines in the test regions. As a procedural control, a colored line always appears in the control region to show that the test had been performed properly [11].

3.1 Data Acquisition

In this experiment, the designed deep learning model trains the patterns of the influenza detection kit images. One hundred images of six class influenza types were used, which included ‘None’ and low, medium, and high concentrations of ‘Type A’ and ‘Type B’ influenza.

Types A and B influenza image data were the sample data produced by diluting them with a mixture of water, borate-based tween, and bovine serum albumin at the same rate, as listed in Table 2. The data were extracted from an influenza detection kit, and the deep learning model was trained using the corresponding images.

3.2 Data Augmentation

The performance of the designed CNN model was affected strongly by the quality of the image data. Even if the number of image data is sufficient, overfitting can occur when a model is taught with a set of data with a high or low percentage of data in a particular label, i.e., an imbalanced dataset [11].

The dataset in the present study was also unbalanced between the number of ‘None’ and other classes. Hence, data were generated using image aggregation in the remaining classes except for the ‘None’ type class [5].

Geometric variations were used for data augmentation [12]. To set a balance between the images of ‘None’ and the other types, the data were generated by randomly adjusting the image rotation (between 0$^{\circ}$ and 30$^{\circ}$) and brightness level to the remaining data. Data aggregation was carried out by randomly setting the rotation and bright levels of the images, as shown in Table 3.

3.3 Image Classification

Image classification of the influenza detection kits was conducted using an optimized deep learning algorithm, which is explained in Section 2. The model used in this experiment was constructed, as shown in Fig. 4.

Approximately 70% of the total data (427) were used as training data, with the remaining 30% (213) used as test data. The input images were resized to 128${\times}$128. The RGB values were separated, and the convolution operation was conducted. The SGD optimizer was used to update the weights of the networks. The decay rate prevents overfitting. As mentioned earlier, the deep learning model attempted to find better hyperparameters using the BOHB. The BOHB algorithm was used to search for the optimal parameters (learning rate, number of epochs, and weight decay) to fine-tune the image classification model, as shown in Table 4.

Fig. 3. Influenza image data example.

Fig. 4. Architecture of the 2D CNN.