The dataset we used is from Pulmonary Chest X-Ray Abnormalities.

It contains 635 pieces of images and includes normal lungs and abnormal lungs. All of the images are labeled with either “0” or “1”. “0” means that this lung is normal and “1” means abnormal. For each image, its size is 3000 × 3000, which is a large image. So we resize the images into 150 × 145, which saved a lot of time in training the model.

Figure 1 illustrates the different color between the normal and abnormal lungs. Therefore, our model specified different color and shape of the lungs to identify diseases.

To begin with, we randomly separated the CT images of lungs in the database into three different sets—the training set, the validation set and the test set and the division ratio is 6:2:2. The training set and validation set were taken into operation in training phase, and test set was used to check the accuracy (Figure 2). Then, we put the two sets into the CNN model we built.

Convolutional neural networks (CNNs) are a class of deep neural networks. CNN is widely used and the de-facto standard in image recognition. Each CNN model consists of two main parts, one is feature learning and the other is classification. For the feature learning, we can [remove can’s] operate different numbers of layers, and there will be a Convolution filter, Activation function and a pooling. And we used Maxpooling, whose function is to combine small local region into single output with the largest value in the region, which could greatly avoid the massive calculation. Then in the classification part, we flattened and densed the matrix (Figure 3).

In our CNN model, we built two layers in our code. We added two Convolution2D functions. In the first layer, we had 25 filters with size (5, 5), and we used 50 filters in the second layer with the same size. After each Convolution2D, we used Maxpooling to reduce the calculation and then used the Activation function—“relu”. Finally, because our images are labeled with either “0” or “1”, we densed the output into two chann. After we ran our training program, the trained CNN model was able to distinguish whether the lungs in the CT image are healthy or not. Then we visualized the result we got Figure 4 and Figure 5.

In Figure 4, we set the study rate as 0.01, and the validation accuracy we got was 78%. Then in Figure 5, we tried to modify different parameters, we improved validation accuracy to 81% with study rate of 0.001 at the end.

In this method, our idea is to cut out the lungs part, because the TB is mainly diagnosed from the lungs. We extract the lungs from the original image (Figure 6).

Unfortunately, this method does not solve our problem, because we do not remove the spinal cord and rib from the lungs. Additionally, due to different sizes of the lungs, sometimes we could only cut out part of the lungs using fixed coordinates (shown in Figure 7).

In Figure 7, the accuracy result is nearly 82% and increases a little bit due to the fact that the noise of shoulders is avoided.

Secondly, we used CANNY algorithm [2] [3] to extract the contour of the lungs before estimation.

CANNY edge detection operator is a multi-level edge detection algorithm developed by John F. Canny in 1986. Generally, the purpose of edge detection is to significantly reduce the size of image data while preserving the original image attributes. At present, there are many algorithms for edge detection. Although CANNY algorithm has a long history, it can be said that it is a standard algorithm for edge detection, and it is still widely used in research.

CANNY algorithm contains 5 main steps as shown in the flowchart (Figure 8).

The extraction effect is shown in Figure 9.

The result of CANNY algorithm is not as good as that in the first model (Figure 10).

This time the accuracy of the results is about 80%, and still maintained rather low because the contour of the spinal cord and rib was extracted at the same time. Plenty of noise still remained. More importantly, the color feature of the

image disappears after the CANNY algorithm and only the Image contour remains. Hence, the CNN model cannot distinguish the abnormal images without the color feature. We need to employ another method to replace based on color feature.

We used our own mask algorithm [4] [5] [6] to crop out the lungs. To be more detailed, the algorithm is called mask method, and the 3 main steps are shown in the flowchart (Figure 11).

First, we make binarization of the images to ensure every Pixel Point is 0 or 255. To better divide and binarize the image, we draw the grey histogram to choose the suitable turning point of binarization (Figure 12).

We choose the Pixel 190 to be a turning point, and then we sign the position of each point. Next we change the 0 point to 255 in the original image and remain other points sustain. Finally we receive the new images only containing lungs without noise (Figure 13).

We used this new dataset to train the CNN model, and this time the accuracy jumps to 87% and the loss value keeps low condition (Figure 14). This method extracts the main features from the image by color brightness and the result is satisfactory, which implies the color brightness is an important factor during the image recognition.