Novel Computer-Aided Diagnosis of Lung Disorders Using Deep Learning Approach

Lung-related diseases, including coronavirus, flu, pneumonia, and lung cancer, are health concerns worldwide. Effective therapy and desirable patient outcomes both require timely diagnosis [1]. As AI develops and medical imaging continues to improve, algorithms are now available to detect and stage these diseases.

In Wuhan, China, in year of 2019, a brand-new coronavirus (COVID-19) was found. This can be a brand-new coronavirus that has never been found in people [2]. When this study was released, the infectious disease was still having an impact on the world, and past tendencies indicated that it might not be the last [3]. Artificial intelligence has been utilized to computerize the determination of various illnesses. Through different machine learning strategies, AI has illustrated its adequacy and high performance in programmed picture classification challenges [4]. Moreover, machine learning portrays models that will learn from and choose based on many cases within the input information. Deep learning (DL) is a widely used method for building systems that can precisely anticipate higher-order frameworks and perform as well as people. For the segmentation of medical images using deep learning (DL), tumors have been primary targets [5].

The work introduces a sophisticated CNN model with assigned prior and self-scan for the algorithmic analysis of the coronavirus. Especially, ResNet50 and ResVGG19 architectures were used in the model since the prediction of results was done from datasets. The study's main contribution is:

i) When applied to lung CT images, the widely used pre-processing methods, such as the median filter, Gaussian filter, and Gabor filter distort edges and boundaries and are incredibly laborious. Consequently, the proposed study employs Data Augmentation with Adaptive Histogram Equalization (DAAHE), which eliminates the previously described problems and enhances performance.

ii) Many segmentation methods found in literature need to determine the number of statistical predicates, or clusters, which adds to their processing time. Without the use of any statistical data, the suggested segmentation technique finds the lung disorder with efficiency. The research employs Region Proposal Network (RPN) and Bounding Box regression methods to isolate the lung disorder location.

iii) When applied to CT images of the lungs, classification techniques like CNN, Resnet50, U-Net and VGG are found to be more computationally intensive and do not yield the desired level of accuracy. Hence, a brand-new hybrid ResVGG19 learning technique is proposed in this research study to immensely enhance the execution time and decreases the false positives while increasing the classification distinction.

iv) The techniques proposed ResVGG19 classifier to differentiate the lung disorder as benign or malignant to get better classification compared to other classification methods.

v) To demonstrate the effectiveness of the suggested method, a detailed evaluation is conducted that compares the segmentation and classification cases with a variety of performance metrics.

The organization of the paper is as follows: In Section 2, a number of research publications pertaining to the various segmentation and classification techniques applied to lung CT images are specifically analyzed. Section 3 formulates the problem description and provides the key information for the suggested system. In the fourth portion of this study, the simulation results are analyzed and discussed. portion 5 presents the last portion, in which the author summarizes the entire book.

The widespread COVID-19 has provoked critical investigation into mechanized strategies for diagnosing the infection utilizing diagnostic imaging, especially chest X-rays and CT scans. These modalities give valuable bits of knowledge into the movement and severity of lung diseases caused by the SARS-CoV-2 infection.

Collect an expansive, large expansive huge dataset of lung CT filter images commented on with the stages of lung cancer, pneumonia, and COVID-19. These images should come from reliable medical sources and should include various stages of the diseases. Ensure the images are labeled accurately, indicating the disease (lung cancer, pneumonia, COVID-19) and its stage. Normalize the images to ensure consistent pixel intensity values. Fragment the lung locale from the CT checks to center the investigation on important regions. Select a ResVGG-19 demonstrate design and Part the information into preparing, approval, and test sets. Prepare the demonstrate on the preparing set and approve its execution on the approval set. Design the output layer of the demonstration to have different classes comparing to diverse stages of each malady (e.g., benign lung cancer, malignant lung cancer, benign pneumonia, malignant pneumonia, benign COVID-19, malignant COVID-19). Utilize a reasonable misfortune work like categorical cross-entropy for multi-class classification.

Analyze the confusion matrix to see how well the demonstration recognizes between distinctive stages and illnesses in Figure 1.

1.png

Figure 1. Experiment pipeline for proposed system flow structure

3.1 Bilinear interpolation

Bilinear insertion could be a resampling strategy utilized in image preparation to gauge the value of an unused pixel based on the values of surrounding pixels [20].

3.1.1 Algorithm: Bilinear interpolation

Step 1: Identify the Surrounding Pixels:

Identify the four surrounding pixels in the original image (a, b). These pixels form a rectangle around the point (a, b). Let the coordinates are $\left(a_1, b_1\right),\left(a_2, b_1\right),\left(a, b_2\right),\left(a_2, b_2\right)$.

Step 2: Compute the Weights:

Compute the distances from the point (x, y) to each of the four surrounding pixels. These distances are used to calculate the weights for the interpolation [21].

$w_1=\left(a_2-a\right)\left(b_2-b\right)$          (1)

$w_2=\left(a-a_2\right)\left(b_2-b\right)$          (2)

$w_3=\left(a_2-a\right)\left(b-b_1\right)$          (3)

$w_4=\left(a-a_2\right)\left(b-b_1\right)$          (4)

Step 3: Interpolate in One Direction (Horizontal Interpolation):

Interpolate along the top and bottom boundaries of the rectangle formed by $\left(a_1, b_1\right),\left(a_2, b_1\right),\left(a, b_2\right),\left(a_2, b_2\right)$.

$v_1=f\left(a_1, b_1\right) w_1+f\left(a_2, b_1\right) w_2$          (5)

$v_2=f\left(a_1, b_2\right) w_3+f\left(a_2, b_2\right) w_4$          (6)

Step 4: Interpolate in the Other Direction (Vertical Interpolation):

Finally, interpolate between $v_1$  and $v_2$ along the vertical direction,

$f(a, b) \approx v_1\left(b_2-b\right)+v_2\left(b-b_1\right)$          (7)

For COVID-19 image processing jobs where a trade-off between accuracy and computing economy is required, bilinear interpolation is a preferred option since it achieves a balance between simplicity and efficacy [21]. After pre-processing, the lung dataset is given to the Data Augmentation section.

3.2 Data augmentation

Data augmentation can be considered an essential process in machine learning, mainly for tasks including lung image data like Healthy, COVID, Non-COVID, Pneumonia, and Lung cancer image classification [22]. It also assists with model generalization by synthesizing various transformation forms of the training dataset in Figure 2. The improved output pictures which are provided to the procedure of adaptive histogram equalization [23]. Accuracy model according to the degree of augmentation can be demonstrated in Table 1.

2.png

Figure 2. Some of the data augmentation examples it is worth considering as follows

Table 1. Comparison of the models with and without information expansion (DA)

3.3 AHE

In contrast to standard Histogram Equalization, which applies histogram equalization to the entire picture, AHE operates when the image is separated into distinct sub-regions or tiles from which histogram equalization is applied. This technique can aid in enhancing local contrast and highlighting details in previously overshadowed areas of an image [28]. CLAHE is a base-variant of AHE that tries to prevent sounds from enhancing contrast [29].

3.3.1 Algorithm: AHE

Step 1: Divide the image into tiles or regions.

Step 2: Let the image be divided into m×n tiles.

Step 3: For each tile T_i,j, calculate the histogram $H_{i, j}(k)$ for intensity levels k.

Step 4: Calculate the Cumulative Distribution Function (CDF) for each tile.

$C D F_{i, l}(k)=\sum_{i=0}^k H_{i, j}(l)$          (8)

Step 5: Normalize the CDF.

$C D F_{i, j}^{\prime}(k)=\frac{C D F_{i, l}(k)-C D F_{i, l}(o)}{C D F_{i, l}(255)-C D F_{i, l}(o)} \times 255$          (9)

Step 6: Map the original pixel intensities in each tile to the new values using the normalized CDF.

$V_{i, j}^{\prime}(a, b)=C D F_{i, j}^{\prime}\left(V_{i, j}(a, b)\right)$          (10)

where, $V_{i, j}(a, b)$ is the original intensity and $I_{i, j}^{\prime}(a, b)$ is the new intensity.

Interpolate the pixel values at the boundaries of the tiles to ensure smooth transitions.

3.4 Segmentation with deep learning

The act of dividing an image into two or more sections is primarily emphasized by the variety of definitions [30]. The objective of segmentation is to give each pixel in an image a name so that those pixels that have the same label have similar properties shown in Figure 3.

Step 1: Choose a deep learning model architecture for segmentation.

Step 2: Train the model on labelled segmentation datasets.

Step 3: Predict segmentation masks for a fresh collection of lung images using the learnt model.

3.png

Figure 3. ResVGG-19 CNN block

3.4.1 Algorithm: Deep learning-based segmentation

Step 1: Region Proposal Network (RPN)

RPN may be a key component in present-day object location systems, such as Quicker R-CNN, which produces candidate object districts. These proposed calculations are at that point utilized by the discovery arrangement to classify objects and refine their bounding boxes [31].

Anchors:

Fixed-size reference boxes are placed uniformly across the lung image.

Anchors={(x,y,w,h)} for each position on the feature map.

Objectless Score and Bounding Box Prediction.

$P_{ {object }}=\sigma\left({Conv}\left(f_{ {map }}\right)\right)$          (11)

$t_x, t_y, t_w, t_h={Conv}\left(f_{ {map }}\right)$          (12)

Step 2: Feature Extraction based on Region of Interest (RoI)

RoI pooling is an essential factor required in object detection networks, including Faster R-CNN, which allows the selection of a fixed-size feature map from a variable-sized region of interest [32]. It is useful in this research to extract features properly, thereby having the capacity to take objects of varying sizes and aspect ratios.

$F_{{RoI }}= ROI\  Pooling\ \left(f_{{map }}\right., RoIs)$         (13)

Step 3: SoftMax Classification

Softmax classification is a method widely employed in machine learning and used mainly in neural networks to solve multi-class classification problems. The SoftMax function transform the raw class scores, or logits, through converting the vectors into probabilities that always sum to 1. This proposed network to predict the probability of each class and select the most likely class as the output [33].

Compute the raw class scores (logits) for each class.

Convert the logits into probabilities using the SoftMax formula.

$P_{{class }}={Softmax}\left(W_{c l s} \cdot f_{R o I}+b_{c l s}\right)$          (14)

Step 4: Bounding Box Regression

The essential objective of bounding box regression is to alter the facilities of the starting bounding box recommendations to adjust with the genuine boundaries of the objects within the lung image. This alteration is learned amid the preparation based on the contrasts (offsets).

$t_x, t_y, t_w, t_h=W_{c l s} . f_{R o I}+b_{c l s}$          (15)

Step 5: Mask Prediction 

Mask R-CNN is an expansion of the Quicker R-CNN protest location system that includes the capability to create pixel-level covers for each question recognized in a lung image. It coordinates the qualities of Speedier R-CNN for protest location with a semantic division assignment to create profoundly exact protest covers.

$mask\ Head\ M_{mask}=\sigma\left({Conv}\left(f_{{RoI }}\right)\right.$          (16)

$\begin{gathered}{ Binary\ Mask\ Prediction }\ M_{ {binary }}=  { throshold }\left(M_{ {mask }}\right)\end{gathered}$          (17)

3.5 Classification 

(1) VGG-19 classification

The input of the VGG19 model is a 224 × 224, 3-channel picture with its average RGB value eliminated. Its 19 weighted layers (Figure 4) are made up of 16, and the minimal slot size of 3×3. The architecture is exactly the same as in the previous array with five max-pooling layers, but this time it has a two-pixel stride and a 2×2 core size.

(2) Resnet50 classification

ResNet50 is a residual network model that has a maximum of 26 million parameters and 50 layers. In actuality, Microsoft unveiled the deep convolutional neural network model in 2015. Instead of learning features, we learn residuals in the residual network by subtracting taught features from the layer inputs.

4.png

Figure 4. VGG-19 architecture

3.5.1 Proposed hybrid ResVGG-19 classification

We began our trials by choosing the ResVGG-19 convolutional neural network, pre-trained, and actually enhanced by freezing some of the layers in order to avoid data overlearning, all too often the case given the size of the picture selection used by us. The number of convolutional layers used in ResVGG-19 model is 69. Concerning the network’s input picture, it has a resolution of 512×512×3. It also has 5 layers of Max pooling of size 2×2 on the whole network and 69 layers of convolutional to a fixed size filter of 3×3. There is no disconnection in the two levels, though, at the top, there is a SoftMax output layer. The ResVGG-19 model has more than 138 million parameters, which shows that the given network is a real behemoth. To construct the complex structural architecture, outward convolution layers are used to extend deeply the ability to learn emergent properties; it places multiple.

5.png

Figure 5. Flowchart for ResVGG-19 classification

A flow graph of the classification is shown in Figure 5. Vertical flip isn’t discussed because flipping the X-ray lung images would not compromise practical views as they are not vertically aligned. Following this, the machine learning classifier utilizes augmented images [34] where the test sample is only 20% of the entire dataset, while the remaining 80% is for the training function. Offering the classifier a diverse and properly preprocessed set of images, this study enhances the classifier’s learning and generalization capability.

Matric evaluation:

Convolution algorithms are often evaluated using a number of metrics to see how effectively the model is doing on classification tasks. It can be used to identify its advantages and disadvantages. The following are some important CNN classification performance measures.

(1) Accuracy

Accuracy $=\frac{\text { Amount of accurate forecaste }}{\text { Total number of forecaste }}$          (18)

(2) Precision

One way to measure positive figure exactness would be precision. The ratio of accurately predicted positive perceptions to all predicted positives is displayed.

Precision $=\frac{\text { True positives }}{\text { True positives }+ \text { False positives }}$          (19)

(3) Recall

Review is the ratio of correctly predicted positive views to all perceptions in the actual lesson, and it gauges the show's ability to identify positive moments.

Recall $=\frac{\text { True positives }}{\text { True positives }+ \text { False Negatives }}$          (20)

3.5.2 Confusion matrix

A lattice that tracks how many classifications the show made precisely and conversely is termed a disarray lattice. It enlightens on the kinds of mistakes that are being made. The symbols are used, including wrong positive (FP), untrue negative (FN), genuine positive (TP), and genuine negative (TN).

In cross-validation, the dataset is separated into various subsets, the demonstrate is prepared on a couple of these subsets, and it is at that point assessed on the remaining subsets. It helps in assessing the vigor and generalizability of the demonstration. By reducing the impacts of overfitting and information changeability, it offers a more precise appraisal of the model's execution.