Next Article in Journal
Multi-Objective Routing and Categorization of Urban Network Segments for Cyclists
Previous Article in Journal
The Lightweight Fracture Segmentation Algorithm for Logging Images Based on Fully 3D Attention Mechanism and Deformable Convolution
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Applied Sciences
Volume 14
Issue 22
10.3390/app142210661
applsci-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

An Efficient One-Dimensional Texture Representation Approach for Lung Disease Diagnosis

by
Abrar Alabdulwahab
1,
Hyun-Cheol Park
2,
Heon Jeong
3,* and
Sang-Woong Lee
1,*
1
School of Computing, Gachon University, Seongnam-si 13120, Republic of Korea
2
Department of Computer Engineering, Korea National University of Transportation, Chungju-si 27469, Republic of Korea
3
Department of Fire Service Administration, Chodang University, Muan-gun 58530, Republic of Korea
*
Authors to whom correspondence should be addressed.
Appl. Sci. 2024, 14(22), 10661; https://doi.org/10.3390/app142210661
Submission received: 15 October 2024 / Revised: 14 November 2024 / Accepted: 15 November 2024 / Published: 18 November 2024
(This article belongs to the Section Computing and Artificial Intelligence)
Browse Figures
Versions Notes

Abstract

:
The remarkable increase in published medical imaging datasets for chest X-rays has significantly improved the performance of deep learning techniques to classify lung diseases efficiently. However, large datasets require special arrangements to make them suitable, accessible, and practically usable in remote clinics and emergency rooms. Additionally, it increases the computational time and image-processing complexity. This study investigates the efficiency of converting the 2D chest X-ray into one-dimensional texture representation data using descriptive statistics and local binary patterns, enabling the use of feed-forward neural networks to efficiently classify lung diseases within a short time and with cost effectiveness. This method bridges diagnostic gaps in healthcare services and improves patient outcomes in remote hospitals and emergency rooms. It also could reinforce the crucial role of technology in advancing healthcare. Utilizing the Guangzhou and PA datasets, our one-dimensional texture representation achieved 99% accuracy with a training time of 10.85 s and 0.19 s for testing. In the PA dataset, it achieved 96% accuracy with a training time of 38.14 s and a testing time of 0.17 s, outperforming EfficientNet, EfficientNet-V2-Small, and MobileNet-V3-Small. Therefore, this study suggests that the dimensional texture representation is fast and effective for lung disease classification.

1. Introduction

Viral and bacterial lung diseases are among the most common and dangerous diseases that have affected many people around the world, caused daily death, and spread quickly [1,2,3]. Viral and bacterial lung diseases have a very high mortality rate. The World Health Organization projected that 65 million people have chronic obstructive pulmonary disease (COPD) that resulted in 3 million deaths [4]. Around 808,694 children under the age of five died of pneumonia in 2017 [5,6], and 10 million people had tuberculosis, with 1.4 million deaths [7].
Thus, diagnosing viral and bacterial lung diseases precisely and early to provide the patient with the appropriate treatment is essential. Chest X-rays and computed tomography images are used to diagnose and detect lung diseases. They are considered one of the most efficient diagnostic tools, with over two million procedures performed yearly [8]. However, the diagnostic results are dependent on clinicians’ interpretations and experiences. Chest X-ray interpretations of radiologists’ disagreement are substantial and reported as high as 56% [9]. The lower resolution of chest X-ray images and the limited experience radiologists have while examining chest X-ray images can cause challenging and diagnostic errors. Moreover, viral and bacterial lung diseases are misdiagnosed because their symptoms and signs overlap, causing severe and critical medical complications that require intensive medical care. Thus, chest X-ray images have been merged with deep learning artificial intelligence techniques in an attempt to minimize human errors and facilitate prompt and more accurate detection and diagnosis of different lung diseases [10,11,12,13,14,15,16,17,18]. These artificial intelligence methods, such as K-means, the support vector machine, convolutional neural networks, recurrent neural networks, and generative adversarial networks, are reported to be effective in X-ray image analysis. They are increasingly used to detect lung abnormalities [19,20,21].
Deep learning (DL) algorithms have been trained on several widely used chest X-ray image datasets to help clinicians make accurate diagnostic decisions. The Japanese Society of Radio Logical Technology, in cooperation with the Japanese Radiological Society Chest X-ray database (JSRT), is a digital image database containing 154 subjects with lung problems and 93 subjects without lung problems [22,23]. Montgomery County, MC, USA, has a chest X-ray database for tuberculosis, with 80 healthy subjects and 58 tuberculosis patients [24]. The Guangzhou Women and Children’s Medical Center dataset [25] consists of 5863 JPEG-formatted chest X-ray images of children. It has three primary folders: train, test, and validation, each containing subfolders for the normal and pneumonia categories [26]. The COVID-19, pneumonia, and normal chest X-ray posteroanterior (PA) dataset [27] has 4575 samples and 1525 images of children and adults diagnosed with COVID-19, pneumonia, or a normal chest [27]. Due to the complex construction of DL algorithms, the training process on these datasets takes a long time.
The training time for deep convolutional neural networks on the Montgomery County (MC) chest X-ray database to detect tuberculosis ranges from several hours to a few days [28,29]. Using a single high-end computational power, the multitask learning model that involves segmentation and classification on the Montgomery County dataset takes approximately 24–48 h. The U-Net for segmentation and DenseNet for classification can extend this time further based on parameter tuning and reiterative enhancements [29,30]. Moreover, the training time for the U-Net architecture on the JSRT database for lung segmentation ranges from several hours to a few days. Extensive computational resources are required for preprocessing and careful model tuning of 247 high-resolution chest X-ray images to achieve optimal results [30].
Training the ResNet-34 model on the Guangzhou Women and Children’s Medical Center dataset for around 100 epochs on a high-end graphics processing unit can take approximately 10 to 20 h [31]. On the other hand, training GoogLeNet, ResNet-18, and DenseNet 121 models collectively on the COVID-19, pneumonia, and normal chest X-ray PA dataset [27] with resizing images to smaller dimensions took several hours to a few days [32]. Another study emphasized that transfer learning employing VGG-19 and ResNet-50 models demonstrated high accuracy in detecting pneumonia. However, training times on the COVID-19, pneumonia, and normal chest X-ray PA datasets [27] extend to a day or more in a powerful Graphics processing unit [33]. Such computational resources and advanced technology are unavailable in remote hospitals and underdeveloped countries. Thus, simplifying this technology and making it available will promote human well-being.
The model training time can vary significantly based on the complexity of the model, the computational resources available, the model architecture, the dataset size, the number of epochs, batch size, and the computational power. The deep-learning algorithms consist of multiple layers that integrate alternating convolutional and pooling processing layers for feature augmentation, enhancement of the image resolution, and extraction, followed by one or more fully connected layers for category classification [34,35,36]. These deep learning algorithms showed promising results in dealing with 2D medical images, spectral images, and video frames [34,35,36,37,38]. However, they require high computational complexity levels, where it takes a long time to analyze 2D medical images and an extensive training dataset for training the classifier to minimize overfitting problems. Based on the available literature, no study has converted a 2D chest X-ray into 1D data to train deep-learning algorithms to detect lung abnormalities but has not detected a cardiomegaly-level (enlarged heart). Previously, 2D chest X-rays were transformed to 1D to detect a cardiomegaly-level (enlarged heart) rapidly. It showed high accuracy, reaching 98.00% higher than the traditional 2D approach. However, the image transformation method used all the image pixels instead of utilizing a statistical representation to reduce image dimensions and carry meaningful information. Using statistical methods for feature extraction provides a more straightforward approach with more insights into the image. Moreover, it used pooling for downsampling to simplify data by taking maximum values, which may result in a loss of fine details [39]. Therefore, simplifying the complexity of image processing and classification is essential. Figure 1 illustrates our proposed method for achieving this by transforming 2D images from the Guangzhou Women and Children’s Medical Center dataset [25] and the COVID-19, pneumonia, and a normal chest X-ray PA dataset [27] into one-dimensional texture representation (1DTR) data. Then, the 1D data will train a deep learning classifier to classify lung diseases. Our contribution involves converting the 2D chest X-ray images into 1DTR, proposing an efficient, quick, and cost-effective method to classify lung diseases. Instead of extracting the features using moving windows or pooling, the texture representations consider a vertical extraction and histogram for dimensional reduction and distinctive data representation. This method could be suitable for remote clinics and emergency rooms for fast diagnosis in low-resource environments.

2. Related Works

Bacterial and viral pulmonary infections, particularly pneumonia and COVID-19, have affected many individuals, especially in environments with low hygienic living conditions, overcrowding, and an inadequate medical infrastructure. Pneumonia and COVID-19 have a very high mortality rate, with around 65 million people having COVID-19 leading to 3 million deaths; 808,694 children died of pneumonia in 2017, and 10 million people had tuberculosis, with 1.4 million deaths [4,5,6,7]. Thus, early diagnosis of this infectious disease is crucial to providing proper treatment, increasing survival rates, and reducing the spread of infection. Chest X-ray imaging is one of the most frequently used methods for diagnosing these pulmonary infectious diseases. It is considered one of the most efficient diagnostic tools, with over two million procedures performed yearly [8]. However, reading chest X-ray images is challenging and subject to many variabilities, including radiologists’ interpretations and experiences and overlap of pulmonary symptoms and signs, causing severe and critical medical complications that require intensive medical care. Chest X-ray interpretations of radiologists’ disagreement are substantial and reported as high as 56% [9]. Thus, deep learning techniques have been used to interpret, read, and analyze chest X-ray images to minimize human errors and facilitate prompt and more accurate detection and diagnosis of different lung diseases [10,11,12,13,14,15,16,17,18]. Artificial intelligence methods are reported to be effective in 2D chest X-ray image analysis. They are increasingly used to detect lung abnormalities [19,20,21]. The large, publicly available 2D chest X-ray image datasets have encouraged scientists to advance the performances of deep learning algorithms to detect and classify various pulmonaria diseases automatically.

2.1. Lung Abnormalities Detection Methods

Dey et al. (2021) proposed a pneumonia detection method that customized VGG19 architecture computing the handmade features from the 2D chest X-ray, achieving a 97.94% classification accuracy [40]. Sahlol et al. (2020) used artificial Ecosystem-Based Optimization of the deep neural network features for detecting tuberculosis in 2D chest Radiographs in a Shenzhen dataset, achieving 90.2% classification accuracy [41]. Souid et al. (2021) classified and predicted pulmonary pathologies in chest X-rays using a modified model, MobileNet-V2, trained on the NIH Chest-Xray-14 database. They achieved above 90% classification accuracy and proposed a model that can be trained and modified equipment with low computing power that is implemented into smaller IoT devices. They suggested that deep neural network training in medicine is a practical and feasible option as more and more public databases are accessible [42].
Moreover, Alabdulwahab and Lee, 2023 [43], proposed the Swin Transformer and the Residual Neural Network’s performance to classify pneumonia, COVID-19, and healthy 2D chest X-ray images using the Guangzhou Women and Children’s Medical Center dataset [25], and C-19, pneumonia and normal chest X-ray PA datasets [27]. They concluded that the Swin Transformer outperforms the Residual Neural Network to assist radiologists in diagnosing pneumonia. Unnisa, 2023 [44], proposed a modified CNN model with an architecture consisting of 20 convolutional layers, 20 batch normalization layers, and 20 ReLU layers to improve the accuracy of detecting pneumonia. They run this model on the 2D chest X-ray images from the National Institutes of Health and Kaggle datasets [25,45]. This architecture detected pneumonia with an accuracy rate of 95.4% from the chest X-ray images. Furthermore, Hasan et al., 2023 [46], ran an ensemble of three CNN models, VGG16, MobileNet-V2, and DenseNet169, on the chest X-ray scan dataset from the Guangzhou Women and Children’s Medical Center dataset [25]. They reported successfully identifying pneumonia and normal cases with recall rates of 93% and 89%, a maximum accuracy of 92%, and precision rates of 93% and 89%.

2.2. Chest X-Ray Datasets

Many other studies used different DL techniques trained on various chest X-ray image datasets and showed different levels of detection accuracy [31,47,48,49,50,51,52,53]. Some of these datasets have many images with high-image pixels for many patients. The ChestX-ray8 dataset has 108,948 images with 1024 × 1024 pixels for 30,805 patients. The ChestX-ray14 dataset has 112,120 images with 1024 × 1024 pixels for 32,717 patients. The CheXpert dataset has 224,316 images for 65,240 patients. The MIMIC CXR dataset has 473,057 images with 2544 × 3056 pixels for 63,478 patients. The Indiana dataset has 7470 images with 512 × 512 pixels for 3996 patients. The Padchest dataset has 160,868 images for 67,000 patients. The PLCO dataset has 185,241 images for 56,071 patients. These large 2D chest X-ray datasets required intense supervision to make them suitable for multiple-layer DL techniques. They required annotated bounding boxes, adjusted image boundaries, a corrected batch size to minimize overfitting problems, and high computational power to feature augmentation, enhancement of image resolution, extraction, and one or more fully connected layers for the category classification of 2D images. These requirements are not available in many remote or underdeveloped countries’ hospitals. Therefore, to overcome these challenges, an attempt was made to convert a 2D chest X-ray into 1D data to train deep learning algorithms to detect a cardiomegaly-level (enlarged heart) rapidly [39]; it showed high accuracy, reaching 98.00% higher than the traditional 2D chest X-ray approach. The results of this study are promising and may be used for other chest problems. However, the image transformation method used all image pixels instead of utilizing a statistical representation to reduce the image dimensions, carry meaningful information, and provide a simpler approach with more insights into the image. Moreover, it used pooling for downsampling to simplify data by taking maximum values, which may result in the loss of fine details. Therefore, simplifying the complexity of image processing and classification is crucial. This study aims to simplify the image processing and classification complexity by converting 2D images in the Guangzhou Women and Children’s Medical Center dataset [25] and the COVID-19, pneumonia, and normal chest X-ray PA dataset [27] into 1DTR data. Then, the 1D data will train a deep learning classifier to classify pneumonia and COVID-19.

3. Methods

This section introduces our new method for diagnosing lung diseases quickly using 1D data. We used the Guangzhou Women and Children’s Medical Center dataset by Sherekar [25] and the COVID-19, pneumonia, and normal chest X-ray PA dataset created by Asraf [27] to convert the 2D data into 1D data and train the 1DTR model for lung disease classification.

3.1. Statistical Feature Extraction

In our approach, Figure 2 shows the 1DTR method that can diagnose lung diseases using 1D data for a short time processing. Thus, to achieve the intended goal of the study, we first collected a widely used dataset, the Guangzhou Women and Children’s Medical Center dataset [25], and the COVID-19, pneumonia, and normal chest X-ray PA dataset [27]. These datasets were then converted to a grayscale image to simplify the image data, reducing the computational complexity and focusing on essential features for analysis. Once the datasets are simplified, we extract statistical features. Instead of using window-sized downsampling to obtain statistical features, our method extracts the statistical features vertically, reducing the image dimensions into a 1DTR. This vertical extraction provides several benefits, such as computational efficiency, maintaining key features of the disease’s white areas, and getting a more comprehensive representation of the classes.
This step is essential as the mean gives information about the average intensity. It can normalize the image intensity, as chest X-rays can vary in contrast and brightness due to the equipment and the acquisition environment. Moreover, different tissue densities, such as bones, can affect pixel intensities. They typically appear brighter as they absorb more X-rays, unlike the lungs, which appear darker with a lower intensity as they are filled with air. The mean provides a consistent average value and distribution for the pixel, ensuring that relatively high pixel variables do not shadow the crucial features.
Choosing standard deviation came from the fact that it adds information about intensity distribution and variability. In pneumonia, the affected lung regions usually have a more diverse patch appearance with varying intensity, unlike healthy lungs with homogeneous intensity patterns. Accordingly, the higher standard deviation indicates a more significant variation, which can be interpreted as pneumonia. The mean and standard deviation of the pixel intensities can provide meaningful texture descriptions, such as homogeneity, contrast, and smoothness. These can capture the small patterns and changes between viral pneumonia, bacterial pneumonia, and normal lungs and help quantify the changes among them, which is essential for efficient training of the DL model.

3.2. Local Binary Pattern Features

The local binary pattern (LBP) feature extraction method was used to understand the characteristics of different classes better. The LBP initially used a 3 × 3 window size to downsample the image with a stride number of 1, where it looks at the texture of the grayscale images by comparing each pixel with its neighboring pixels. It calculates the LBP using eight points in a circular pattern with a radius of one pixel around each center. The comparison is performed by checking whether the neighboring pixels are greater or smaller than the central pixel. If the pixel has a higher value than the central pixel, it is assigned as 1; if not, it is assigned as 0. Subsequently, these binary values for each pixel are arranged into a sequence to form a binary number. These values are then used to make a histogram with 19 bins representing different uniform LBP patterns. This histogram provides a statistical summary of the X-ray texture patterns across the image. The decision to use a histogram came from the idea that it allows for the compact and comprehensive representation of the texture distribution in a 1D form, reducing the data size in the storage and the training and testing time. Additionally, to achieve high classification performance, the LBP histogram provides a density estimate to help capture the relative frequency of several texture patterns of the lung with and without a disease, making it constant to scale and lung condition changes. Instead of directly using downsampled images, this approach facilitates the extraction of one-dimensional features essential for high classification performance. Figure 3 is a normalized histogram representing LBP values’ frequency distribution. The values were normalized to focus on the texture distribution instead of the raw pixel count and to compare the textures across different image sizes.

3.3. Interpolation

The next step is to resize them to ensure consistency across all samples, as the original images may vary in size. This process is crucial because DL models require inputs of uniform dimensions for effective processing and learning. The interpolation technique was applied to achieve the right image size, allowing us to adjust the features to a unified size without losing essential attributes. Interpolation selects values from the original data at regular intervals and estimates intermediate values to fit the new given target size. This way ensures that the features are resized to the intended length and preserves the overall structure and distribution of the data. This method was selected over the padding because some images have large numbers of pixels, resulting in large rows of arrays, which increase the dataset size. Moreover, cropping the large images to fit the small ones can result in losing essential parts of the image that might contain the disease attributes. In this process, linear interpolation was used, where points are linearly spaced within a provided range, and the values between known data points are estimated. The interpolation function maps the original feature values to a new, uniformly spaced grid corresponding to the target size. Once we extract the features, they are added to the same CSV file. which combines the mean, standard deviation, and LBP, with a size of 219. By doing so, we create a resized version of the feature set that preserves the essential characteristics of the original data.

3.4. Feedforward Neural Network

The generated 1D dataset was first preprocessed using a standard scaler and then passed to a feed-forward neural network (FFNN). This network was selected over the recurrent neural networks because it efficiently handles the nonsequential dataset. Since our feature characteristics are one-dimensional and do not require deep computations, we created a relatively low number of layers. We maintain an efficient performance and low training and testing times; the FFNN model is configured by four layers with 128, 64, 32, and 16 units, respectively. We applied various learning rates during the training process, such as 0.1 and 0.05, along with different epoch values. However, we found that using a learning rate of 0.01 with low training epochs yielded the best performance. Initially, higher dropout rates caused underfitting in our small network. Therefore, the model utilized a dropout rate of 0.01 between the layers, along with 0.01 regularization, to prevent overfitting. For the output layer, we used SoftMax to measure the probability of the classes. For a fair comparison, we adapted the EffientNet-B0 model developed by Tan and Le (2019) [54] in our study to be trained and tested on the 2D dataset (Images). We modified the classifier for this task to have the same layers as our 1DTR model. Therefore, the sequence of dense layers and ReLU activations replaced the original fully connected layer. As with the 1DTR FFNN architecture, the input feature size is 128, 64, 32, and 16 before being classified into three categories. The training process was conducted using the Cross-Entropy Loss function and an Adam optimizer.

4. Datasets

This method used two types of datasets to utilize the performance of the proposed framework. The first dataset is the Guangzhou Women and Children’s Medical Center dataset [25], which consists of chest X-ray images of children formatted in JPEG with 5863 images. Each of the primary folders, train, test, and validation, has two subfolders: normal and pneumonia categories; the pneumonia images are divided into two labels, bacteria and virus. This dataset has already been split into 90% for training and 10% for validation. Figure 4 shows the samples of each class, normal bacteria and virus, respectively. On the other hand, the COVID-19, pneumonia, and normal chest X-ray PA dataset [27] contains three classes: covid, pneumonia, and normal. Figure 5 illustrates the class of the PA dataset, normal, covid, and pneumonia, respectively. This dataset [27] contains 4575 samples of children and adults, and it was split into 80:10:10 for training, testing, and validation, respectively.
In addition, in PA, pneumonia and COVID-19 could be differentiated radiologically. Pneumonia looks like a localized, condensed consolidation opacity in one or more lung lobes. COVID-19, in its early stages, demonstrates bilateral, peripheral ground-glass cloudiness distribution without lobar dense consolidation. At late stages, the ground glass cloudiness progresses into consolidations with an “irrational paving” shape of interlobular septal thickening. Moreover, pleural effusions are likely present in pneumonia but uncommon in COVID-19 [55]. In the Guangzhou dataset, bacterial and viral pneumonia exhibits distinct radiological characteristics. Radiographically, bacterial pneumonia is typically a focal lobar consolidation seen in one lobe. In contrast, viral pneumonia generally presents with a more diffuse “interstitial” pattern across both lungs [25].
Briefly, in both datasets, chest X-rays are taken with ordinary posteroanterior (PA) and anteroposterior (AP) projections for respiratory assessment to provide clear insights into lung health and the distribution and severity of respiratory involvement. Thus, these datasets mostly include X-ray images for screening and diagnostic purposes in clinics [25,26,27]. The X-ray techniques used in this dataset are digital radiography to have a higher resolution and more manageable dataset for computational analysis; standardized radiation settings tailored for patients to reduce the sensitivity to X-rays; and image preprocessing approaches to enhance the clarity of the images for DL applications and testing algorithms for automated respiratory disease classification. Thus, our proposed method will classify the label among three classes per dataset.

Dataset Visualization

For data feature visualization purposes, we reduced the dimensionality of the LBP features using t-Modified [56]. The t-SNE dimensionality reduction technique is highly efficient for visualizing high-dimensional data in a two-dimensional space. This technique is particularly effective in identifying similarities between data points, as it arranges similar points close together and dissimilar points far apart. Using Gaussian distribution in this process allows us to compute a conditional probability. In this context, the conditional probability reflects the likelihood of a data point being similar to another point, given its high-dimensional representation. This step could be represented as:
p j | i = exp ⁡ − x i   −   x j 2 2 σ i 2 ∑ k ≠ i exp ⁡ − x i   −   x k 2 2 σ i 2
where:
  • The two points x i and x j are two points in high-dimensional space.
  • The p j | i is the conditional probability.
  • The x i − x k 2   is the squared Euclidean distance between the two points.
  • The σ i is the variance of the Gaussian centered at x i   and x j , to the value of σ i   using the perplexity parameter that controls the effective number of neighbors.
Figure 6 shows that it contains each dataset’s classes, separating each class depending on its features. The graphs clearly show how this feature can capture a distinctive attribute of the classes, especially with the normal class in the Guangzhou Women and Children’s Medical Center dataset [25], since the features of the bacteria and virus of the first graph and pneumonia and COVID-19 in the second graph are similar to each other; this indicates the need for accurate classification of DL methods.

5. Results

Our experiment used the Guangzhou and PA datasets to evaluate the proposed approach using several performance metrics: precision, recall, F1 score, and accuracy. In this experiment, the system configuration included a CPU, Intel(R) Core(TM) i7-10700K, a GPU, NVIDIA GeForce RTX 3090, and a memory size of 64 GB. We compared our method with other convolutional neural network (CNN) models using the same datasets and number of layers. In our study, we chose EfficientNet-B0 [54], a type of CNN architecture designed to optimize accuracy and efficiency. It achieves high levels of accuracy while being computationally efficient, making it well suited for use in environments with limited computational resources. The B0 is the baseline model of the EfficientNet family and is the smallest among the larger models. We also evaluated EfficientNet-V2-Small [57], which marks a significant advancement in the design of convolutional neural networks. This model offers an optimal balance between performance and efficiency, achieving an impressive top-1 accuracy of 87.3% on the ImageNet dataset while maintaining a compact size of just 20.33 million parameters. We also evaluated the performance of MobileNet-V3-Small [58]. MobileNetV3-Small is an efficient convolutional neural network developed using a combination of neural architecture search (NAS) and manual optimizations. It reaches a top-1 accuracy of 67.5% on ImageNet, which is 6.6% more accurate than MobileNet-V2 while maintaining a similar level of latency. All the models classify the X-ray images directly, without employing our proposed feature extraction method. We aim to demonstrate that utilizing the proposed features with relatively small neural networks can achieve high accuracy and reduce both training and testing times, especially when compared to more complex models.
Table 1 evaluates the 1DTR and CNN models across two datasets. The 1DTR significantly outperformed the CNN model, achieving 96% and 99% accuracy. Although EfficientNet-B0, EfficientNet-V2-Small, and MobileNet-V3-Small performed well on the PA dataset, they could not achieve similarly high performance levels. Table 2 and Table 3 demonstrate the model’s classification performance for each class, showing its superior performance to other CNN models.
Figure 7 demonstrates the confusion matrix of each experiment for a clear summary of how the models performed. It shows the actual and predicted classification to help identify the model’s errors. Thus, the 1DTR (a and e) resulted in a few errors, indicating that the performance classification obtained the correct predictions. The other models produced more incorrect predictions, which may be attributed to overlapping features. However, when comparing the confusion matrixes of the CNN models (b), (c), (d), (f), (g), and (h) across different datasets, it is evident that its performance is significantly better on the Guangzhou dataset compared to the PA dataset. Furthermore, improving training and testing times can increase the efficiency of the model used for a disease-type diagnosis. Table 4 shows the training and testing times for different models using the PA and Guangzhou datasets. The 1DTR model is the fastest for both datasets, with training times of 38.14 s and 10.85 s on the PA and Guangzhou datasets, respectively. Its testing times are also very quick, at 0.17 and 0.19 s.
In comparison, models such as EfficientNet-B0, EfficientNet-V2, and MobileNet-V3 require significantly more time to train and test, particularly on the PA dataset. Among these, EfficientNet-B0 has the longest training time, taking 7776.78 s on the PA dataset. On the other hand, MobileNet-V3 is the quickest of the neural network models, with training times of 6425.22 s for the PA dataset and 4920.32 s for the Guangzhou dataset.
According to the results, the 1DTR significantly outperformed the neural network models in terms of both accuracy and speed of classification. Therefore, using 1DTR involves lower computational complexity and a reduced processing time.

6. Discussion

The results show that a 1D representation of the lung texture can provide enough meaningful features without needing a full image to train the FFNN DL for accurate and fast image classification with comparable results against CNN-based models. Moreover, it shows the importance of data manipulation for accurate performance.
Pneumonia and COVID-19 are communicable and contagious diseases that quickly spread among communities in crowded environments and are leading causes of morbidity and mortality globally. The World Health Organization estimated pneumonia incidence to be over 120 million early patients, with around 10 million severe cases. COVID-19 has affected billions of people since the pandemic started in late 2019, and as of mid-2023, there have been around 760 million confirmed cases, with over 6.9 million deaths worldwide [59]. Poor, inadequate, or delayed diagnosis contributes to their further spread and can lead to more severe outcomes, mismanagement of the infection, and increased transmission of the disease [59,60]. Additionally, as presented in the introduction and related studies, X-ray image datasets are large and require costly equipment and resources. Therefore, there are more likely more direct methods to handle 1D data as a rule-based model. For example, instead of employing DL, it is possible to utilize a simple algorithm to identify the pixels’ range. If the range falls between 0 and 127, it is considered normal; a range of 128 to 255 signifies pneumonia. However, this straightforward approach is not feasible for several reasons. As shown in Figure 7, the overlap between classes makes classifying based on pixel ranges challenging, and individual differences in chest bone thickness can affect pixel values. Thus, we are converting the X-ray vertically instead of pooling to 1D representation descriptive statistics, and the LBP can extract meaningful attributes about the image for the classifiers. Moreover, it significantly decreases the dataset size, which minimizes storage requirements and speeds up the training and testing processes. This reduction does not affect the DL model’s performance, facilitating quicker diagnoses and patient outcomes in resource-constrained hospitals and emergency rooms.
This model can enhance the vital role of technology in advancing healthcare, assisting clinicians globally in detecting respiratory diseases and preventing their spread. While the results of this study show a clear distinction between normal and infected lungs, further investigation is needed into the use of the DL model for more accurate classification. The confusion in classifying bacterial and viral pneumonia attributes indicates the need for ongoing research. This method could be more efficient for preliminary analysis or tasks with less critical detailed spatial relationships. However, this model may not be suitable for applications that require object segmentation and masking.

7. Conclusions

Viral and bacterial lung diseases, such as pneumonia and COVID-19, significantly threaten global public health. These highly communicable and contagious diseases, particularly dangerous in crowded environments, spread rapidly and are significant contributors to illness and death worldwide. The proposed model has the potential to significantly impact global public health by facilitating the timely and accurate diagnosis of these diseases. To treat the patient appropriately, it is essential to promptly and accurately diagnose viral and bacterial lung diseases. Various artificial intelligence methods, such as K-means, support vector machines, convolutional neural networks, recurrent neural networks, and generative adversarial networks, have proven effective in analyzing X-ray images. These approaches are increasingly used to identify lung abnormalities, and they have been trained on widely used chest X-ray image databases to assist clinicians in making improved diagnostic decisions, where these deep learning algorithms showed promising results in dealing with 2D medical images. Training these methods on large X-ray image datasets requires significant time, costly equipment, and resources for fast disease diagnosis. Therefore, we have proposed a model that simplifies image processing and classification by converting 2D images into 1DTR using descriptive statistics. The LBP can extract meaningful image attributes for the classifier. Our method significantly reduces dataset size and required storage and expedites the training and testing processes without compromising the DL model’s performance. With this achievement, our method can facilitate speedy diagnosis in resource-constrained environments, aiding clinicians worldwide in detecting and preventing the spread of respiratory diseases. Our experiment results indicate that the 1DTR is fast and effective for lung disease classification. The 1DTR achieved 99% with a training time of 10.85 s and 0.19 s for testing, and it achieved 96% accuracy with a training time of 38.14 s and a testing time of 0.17 s on both datasets in Guangzhou, PA, respectively.

Author Contributions

Conceptualization, A.A.; Data curation, A.A.; Formal analysis, A.A.; Funding acquisition, H.J. and H.-C.P.; Investigation, S.-W.L.; Methodology, A.A.; Project administration, S.-W.L.; Resources, S.-W.L.; Software, A.A.; Supervision, S.-W.L.; Validation, A.A. and S.-W.L.; Visualization, A.A.; Writing—original draft, A.A.; Writing—review and editing, A.A. and S.-W.L. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the “Regional Innovation Strategy (RIS)” through the National Research Foundation of Korea (NRF) and funded by the Ministry of Education (MOE) (No. 2021RIS-002). It was supported by the Gachon University research fund of 2022. (GCU-202206100001).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The datasets presented in this study are available in Kaggle in Chest X-ray (COVID-19 & Pneumonia) https://www.kaggle.com/datasets/prashant268/chest-xray-covid19-pneumonia (accessed on 5 August 2024), reference number [25], and in Mendeley Data in COVID19, Pneumonia and Normal Chest X-ray PA Dataset at https://data.mendeley.com/datasets/jctsfj2sfn/1 (accessed on 5 August 2024), reference number [27].

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Ahmad, W.S.H.M.W.; Zaki, W.M.D.W.; Fauzi, M.F.A.; Tan, W.H. Classification of infection and fluid regions in chest X-ray images. In Proceedings of the 2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA), Gold Coast, Australia, 30 November–2 December 2016; IEEE: New York, NY, USA, 2016; pp. 1–5. [Google Scholar] [CrossRef]
  2. Brown, S.-A.W.; Padilla, M.; Mhango, G.; Powell, C.; Salvatore, M.; Henschke, C.; Yankelevitz, D.; Sigel, K.; De-Torres, J.P.; Wisnivesky, J. Interstitial lung abnormalities and lung cancer risk in the national lung screening trial. Chest 2019, 156, 1195–1203. [Google Scholar] [CrossRef] [PubMed]
  3. Cha, M.J.; Chung, M.J.; Lee, J.H.; Lee, K.S. Performance of deep learning model in detecting operable lung cancer with chest radiographs. J. Thorac. Imaging 2019, 34, 86–91. [Google Scholar] [CrossRef]
  4. Marciniuk, D.; Schraufnagel, D.; Society, E.R. The Global Impact of Respiratory Disease; European Respiratory Society: Lausanne, Switzerland, 2017; Available online: https://static.physoc.org/app/uploads/2019/04/22192917/The_Global_Impact_of_Respiratory_Disease.pdf (accessed on 1 July 2024).
  5. World Health Organization. Pneumonia in Children. 2022. Available online: https://www.who.int/news-room/fact-sheets/detail/pneumonia (accessed on 1 July 2024).
  6. Khoiriyah, S.A.; Basofi, A.; Fariza, A. Convolutional neural network for automatic pneumonia detection in chest radiography. In Proceedings of the 2020 International Electronics Symposium (IES), Surabaya, Indonesia, 29–30 September 2020; IEEE: New York, NY, USA, 2020; pp. 476–480. [Google Scholar] [CrossRef]
  7. World Health Organization. Tuberculosis. Available online: https://www.who.int/news-room/fact-sheets/detail/tuberculosis (accessed on 1 July 2024).
  8. Raoof, S.; Feigin, D.; Sung, A.; Raoof, S.; Irugulpati, L.; Rosenow, E.C. Interpretation of plain chest roentgenogram. Chest 2012, 141, 545–558. [Google Scholar] [CrossRef] [PubMed]
  9. Herman, P.G.; Gerson, D.E.; Hessel, S.J.; Mayer, B.S.; Watnick, M.; Blesser, B.; Ozonoff, D. Disagreements in chest roentgen interpretation. Chest 1975, 68, 278–282. [Google Scholar] [CrossRef] [PubMed]
  10. Chandrasekar, K.S. Exploring the deep-learning techniques in detecting the presence of coronavirus in the chest X-ray images: A comprehensive review. Arch. Comput. Methods Eng. 2022, 29, 5381–5395. [Google Scholar] [CrossRef] [PubMed]
  11. Kieu, S.T.H.; Bade, A.; Hijazi, M.H.A.; Kolivand, H. A survey of deep learning for lung disease detection on medical images: State-of-the-art, taxonomy, issues and future directions. J. Imaging 2020, 6, 131. [Google Scholar] [CrossRef]
  12. Priyadarsini, M.J.P.; Kotecha, K.; Rajini, G.K.; Hariharan, K.; Raj, K.U.; Ram, K.B.; Indragandhi, V.; Subramaniyaswamy, V.; Pandya, S. Lung diseases detection using various deep learning algorithms. J. Health Eng. 2023, 2023, 3563696. [Google Scholar] [CrossRef]
  13. Ahmed, S.T.; Kadhem, S.M. Using machine learning via deep learning algorithms to diagnose the lung disease based on chest imaging: A survey. Int. J. Interact. Mob. Technol. (iJIM) 2021, 15, 95–112. [Google Scholar] [CrossRef]
  14. Sun, J.; Liao, X.; Yan, Y.; Zhang, X.; Sun, J.; Tan, W.; Liu, B.; Wu, J.; Guo, Q.; Gao, S.; et al. Correction to: Detection and staging of chronic obstructive pulmonary disease using a computed tomography–based weakly supervised deep learning approach. Eur. Radiol. 2022, 32, 5785. [Google Scholar] [CrossRef]
  15. Bharati, S.; Podder, P.; Mondal, M.R.H. Hybrid deep learning for detecting lung diseases from X-ray images. Inform. Med. Unlocked 2020, 20, 100391. [Google Scholar] [CrossRef]
  16. Feng, Y.; Wang, Y.; Zeng, C.; Mao, H. Artificial intelligence and machine learning in chronic airway diseases: Focus on asthma and chronic obstructive pulmonary disease. Int. J. Med Sci. 2021, 18, 2871–2889. [Google Scholar] [CrossRef] [PubMed]
  17. Kumar, S.; Bhagat, V.; Sahu, P.; Chaube, M.K.; Behera, A.K.; Guizani, M.; Gravina, R.; Di Dio, M.; Fortino, G.; Curry, E.; et al. A novel multimodal framework for early diagnosis and classification of COPD based on CT scan images and multivariate pulmonary respiratory diseases. Comput. Methods Programs Biomed. 2023, 243, 107911. [Google Scholar] [CrossRef]
  18. Saygılı, A. Analysis and segmentation of X-ray images of covid-19 patients using the k-means algorithm. Veri Bilim. 2021, 4, 1–6. [Google Scholar]
  19. Park, B.; Park, H.; Lee, S.M.; Seo, J.B.; Kim, N. Lung segmentation on HRCT and volumetric CT for diffuse interstitial lung disease using deep convolutional neural networks. J. Digit. Imaging 2019, 32, 1019–1026. [Google Scholar] [CrossRef] [PubMed]
  20. Toğaçar, M.; Ergen, B.; Cömert, Z. Detection of lung cancer on chest CT images using minimum redundancy maximum relevance feature selection method with convolutional neural networks. Biocybern. Biomed. Eng. 2020, 40, 23–39. [Google Scholar] [CrossRef]
  21. Golan, R.; Jacob, C.; Denzinger, J. Lung nodule detection in ct images using deep convolutional neural networks. In Proceedings of the 2016 International Joint Conference on Neural Networks (IJCNN), Vancouver, BC, Canada, 24–29 July 2016; IEEE: New York, NY, USA, 2016; pp. 243–250. [Google Scholar] [CrossRef]
  22. Japanese Society of Radiological Technology. Digital Image Database-jsrt Database. 2004. Available online: http://db.jsrt.or.jp/eng.php (accessed on 1 July 2024).
  23. Shiraishi, J.; Katsuragawa, S.; Ikezoe, J.; Matsumoto, T.; Kobayashi, T.; Komatsu, K.-I.; Matsui, M.; Fujita, H.; Kodera, Y.; Doi, K. Development of a digital image database for chest radiographs with and without a lung nodule. Am. J. Roentgenol. 2000, 174, 71–74. [Google Scholar] [CrossRef]
  24. Jaeger, S.; Candemir, S.; Antani, S.; Wáng, Y.-X.J.; Lu, P.-X.; Thoma, G. Two public chest X-Ray datasets for computer-aided screening of pulmonary diseases. Quant. Imaging Med. Surg. 2014, 4, 475–477. [Google Scholar] [CrossRef]
  25. Sherekar, P. Chest X-Ray Images (pneumonia). Kaggle, v2. 2021. Available online: https://www.kaggle.com/datasets/prashant268/chest-xray-covid19-pneumonia (accessed on 5 August 2024).
  26. Kermany, D. Labeled Optical Coherence Tomography (oct) and Chest X-ray Images for Classification. Mendeley Data, v3. 2018. Available online: https://data.mendeley.com/datasets/rscbjbr9sj/2 (accessed on 5 August 2024).
  27. Asraf, A. COVID-19, Pneumonia and Normal Chest X-Ray pa Dataset. Mendeley Data, v1. 2021. Available online: https://data.mendeley.com/datasets/jctsfj2sfn/1 (accessed on 5 August 2024).
  28. Jangam, E.; Annavarapu, C.S.R.; Elloumi, M. Deep Learning for Lung Disease Detection from Chest X-Rays Images; Springer: Berlin/Heidelberg, Germany, 2021; pp. 239–254. [Google Scholar] [CrossRef]
  29. Imran, A.A.Z.; Terzopoulos, D. Semi-supervised multi-task learning with chest X-ray images. In Proceedings of the Machine Learning in Medical Imaging: 10th International Workshop, MLMI, Held in Conjunction with MICCAI 2019, Shenzhen, China, 13 October 2019; Springer: Berlin/Heidelberg, Germany, 2019; pp. 151–159. [Google Scholar]
  30. Nasser, A.A.; Akhloufi, M.A. A review of recent advances in deep learning models for chest disease detection using radiography. Diagnostics 2023, 13, 159. [Google Scholar] [CrossRef]
  31. Guo, K.; Cheng, J.; Li, K.; Wang, L.; Lv, Y.; Cao, D. Diagnosis and detection of pneumonia using weak-label based on X-ray images: A multi-center study. BMC Med Imaging 2023, 23, 209. [Google Scholar] [CrossRef]
  32. Kundu, R.; Das, R.; Geem, Z.W.; Han, G.-T.; Sarkar, R. Pneumonia detection in chest X-ray images using an ensemble of deep learning models. PLoS ONE 2021, 16, e0256630. [Google Scholar] [CrossRef]
  33. Trivedi, A.; Robinson, C.; Blazes, M.; Ortiz, A.; Desbiens, J.; Gupta, S.; Dodhia, R.; Bhatraju, P.K.; Liles, W.C.; Kalpathy-Cramer, J.; et al. Deep learning models for COVID-19 chest X-ray classification: Preventing shortcut learning using feature disentanglement. PLoS ONE 2022, 17, e0274098. [Google Scholar] [CrossRef] [PubMed]
  34. Wang, X.; Peng, Y.; Lu, L.; Lu, Z.; Bagheri, M.; Summers, R.M. ChestX-Ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases; IEEE: New York, NY, USA, 2017; pp. 2097–2106. [Google Scholar]
  35. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition; IEEE: New York, NY, USA, 2016; pp. 770–778. [Google Scholar]
  36. Drozdzal, M.; Vorontsov, E.; Chartrand, G.; Kadoury, S.; Pal, C. The Importance of Skip Connections in Biomedical Image Segmentation; Springer: Berlin/Heidelberg, Germany, 2016; pp. 179–187. [Google Scholar] [CrossRef]
  37. Ullah, A.; Anwar, S.M.; Bilal, M.; Mehmood, R.M. Classification of arrhythmia by using deep learning with 2-D ECG spectral image representation. Remote Sens. 2020, 12, 1685. [Google Scholar] [CrossRef]
  38. Long, J.; Shelhamer, E.; Darrell, T. Fully Convolutional Networks for Semantic Segmentation; IEEE: New York, NY, USA, 2015; pp. 3431–3440. [Google Scholar]
  39. Lin, C.-H.; Zhang, F.-Z.; Wu, J.-X.; Pai, N.-S.; Chen, P.-Y.; Pai, C.-C.; Kan, C.-D. Posteroanterior chest X-ray image classification with a multilayer 1D convolutional neural network-based classifier for cardiomegaly level screening. Electronics 2022, 11, 1364. [Google Scholar] [CrossRef]
  40. Dey, N.; Zhang, Y.-D.; Rajinikanth, V.; Pugalenthi, R.; Raja, N.S.M. Customized VGG19 Architecture for pneumonia detection in chest X-rays. Pattern Recognit. Lett. 2021, 143, 67–74. [Google Scholar] [CrossRef]
  41. Sahlol, A.T.; Elaziz, M.A.; Jamal, A.T.; Damaševičius, R.; Hassan, O.F. A novel method for detection of tuberculosis in chest radiographs using artificial ecosystem-based optimisation of deep neural network features. Symmetry 2020, 12, 1146. [Google Scholar] [CrossRef]
  42. Souid, A.; Sakli, N.; Sakli, H. Classification and predictions of lung diseases from chest X-rays using mobilenet V2. Appl. Sci. 2021, 11, 2751. [Google Scholar] [CrossRef]
  43. Alabdulwahab, A.; Lee, S.W. Comparative analysis of swin transformer and residual neural network for pneumonia classification. J. Korea Inst. Next Gener. Comput. 2023, 19, 7–17. [Google Scholar]
  44. Unnisa, S. Svnet for detecting pneumonia from chest X-ray images. Indian J. Nat. Sci. 2023, 13, 52978–52982. [Google Scholar]
  45. National Institutes of Health. Nih Clinical Center Provides One of the Largest Publicly Available Chest X-Ray Datasets to the Scientific Community. National Institutes of Health (NIH). 2021. Available online: https://www.nih.gov/news-events/news-releases/nih-clinical-center-provides-one-largest-publicly-available-chest-x-ray-datasets-scientific-community (accessed on 1 July 2024).
  46. Hasan, M.R.; Ullah, S.M.A.; Hasan, M. Deep learning in radiology: A transfer-learning based approach for the identification and classification of covid-19 and pneumonia in chest X-ray images. In Proceedings of the 2023 Fourth International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE), Bengaluru, India, 8–9 December 2023; IEEE: New York, NY, USA, 2023; pp. 1–6. [Google Scholar] [CrossRef]
  47. Varshni, D.; Thakral, K.; Agarwal, L.; Nijhawan, R.; Mittal, A. Pneumonia detection using cnn based feature extraction. In Proceedings of the 2019 IEEE International Conference on Electrical, Computer and Communication Technologies (ICECCT), Coimbatore, India, 20–22 February 2019; IEEE: New York, NY, USA, 2019; pp. 1–7. [Google Scholar] [CrossRef]
  48. Liang, G.; Zheng, L. A transfer learning method with deep residual network for pediatric pneumonia diagnosis. Comput. Methods Programs Biomed. 2020, 187, 104964. [Google Scholar] [CrossRef]
  49. Ahamed, K.U.; Islam, M.; Uddin, A.; Akhter, A.; Paul, B.K.; Abu Yousuf, M.; Uddin, S.; Quinn, J.M.; Moni, M.A. A deep learning approach using effective preprocessing techniques to detect COVID-19 from chest CT-scan and X-ray images. Comput. Biol. Med. 2021, 139, 105014. [Google Scholar] [CrossRef]
  50. Kaur, P.; Harnal, S.; Tiwari, R.; Alharithi, F.S.; Almulihi, A.H.; Noya, I.D.; Goyal, N. A hybrid convolutional neural network model for diagnosis of COVID-19 using chest X-ray images. Int. J. Environ. Res. Public Health 2021, 18, 12191. [Google Scholar] [CrossRef] [PubMed]
  51. Maharjan, J.; Calvert, J.; Pellegrini, E.; Green-Saxena, A.; Hoffman, J.; McCoy, A.; Mao, Q.; Das, R. Application of deep learning to identify COVID-19 infection in posteroanterior chest X-rays. Clin. Imaging 2021, 80, 268–273. [Google Scholar] [CrossRef] [PubMed]
  52. Ozturk, T.; Talo, M.; Yildirim, E.A.; Baloglu, U.B.; Yildirim, O.; Acharya, U.R. Automated detection of COVID-19 cases using deep neural networks with X-ray images. Comput. Biol. Med. 2020, 121, 103792. [Google Scholar] [CrossRef] [PubMed]
  53. Hussain, E.; Hasan, M.; Rahman, A.; Lee, I.; Tamanna, T.; Parvez, M.Z. CoroDet: A deep learning based classification for COVID-19 detection using chest X-ray images. Chaos Solitons Fractals 2021, 142, 110495. [Google Scholar] [CrossRef] [PubMed]
  54. Tan, M.; Tan, M. Efficientnet: Rethinking Model Scaling for Convolutional Neural Networks. In Proceedings of the 36th International Conference on Machine Learning, Long Beach, CA, USA, 9–15 June 2019; pp. 6105–6114. [Google Scholar]
  55. Duzgun, S.A.; Durhan, G.; Demirkazik, F.B.; Akpinar, M.G.; Ariyurek, O.M. COVID-19 pneumonia: The great radiological mimicker. Insights Imaging 2020, 11, 118. [Google Scholar] [CrossRef]
  56. der Maaten, L.V.; Hinton, G. Visualizing data using t-sne. J. Mach. Learn. Res. 2008, 9, 2579–2605. [Google Scholar]
  57. Tan, M.; Le, Q. EfficientNetV2: Smaller Models and Faster Training. In Proceedings of the International Conference on Machine Learning, Online, 18–24 July 2021; pp. 10096–10106. [Google Scholar]
  58. Howard, A.; Sandler, M.; Chu, G.; Chen, L.-C.; Chen, B.; Tan, M.; Wang, W.; Zhu, Y.; Pang, R.; Vasudevan, V.; et al. Searching for MobileNetV3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Republic of Korea, 27 October–2 November 2019; pp. 1314–1324. [Google Scholar]
  59. Gong, Z.; Song, T.; Hu, M.; Che, Q.; Guo, J.; Zhang, H.; Li, H.; Wang, Y.; Liu, B.; Shi, N. Natural and socio-environmental factors in the transmission of covid-19: A comprehensive analysis of epidemiology and mechanisms. BMC Public Health 2024, 24, 2196. [Google Scholar] [CrossRef]
  60. Jung, J.; Kim, J.Y.; Park, H.; Park, S.; Lim, J.S.; Lim, S.Y.; Bae, S.; Lim, Y.-J.; Kim, E.O.; Kim, J.; et al. Transmission and infectious sars-cov-2 shedding kinetics in vaccinated and unvaccinated individuals. JAMA Netw. Open 2022, 5, e2213606. [Google Scholar] [CrossRef]
Applsci 14 10661 g001
Figure 1. The overall methodology for 1DTR uses statistical methods and LBP to extract the texture features of the input images.
Figure 1. The overall methodology for 1DTR uses statistical methods and LBP to extract the texture features of the input images.
Applsci 14 10661 g001
Applsci 14 10661 g002
Figure 2. The process involves extracting features from an input image by calculating its mean, standard deviation, and binary pattern. Interpolation is used to ensure all features are the same length. The 1D data are then preprocessed using a standard scaler, and the string labels are converted into integers using a label encoder. Finally, the preprocessed data are passed to an FFNN model.
Figure 2. The process involves extracting features from an input image by calculating its mean, standard deviation, and binary pattern. Interpolation is used to ensure all features are the same length. The 1D data are then preprocessed using a standard scaler, and the string labels are converted into integers using a label encoder. Finally, the preprocessed data are passed to an FFNN model.
Applsci 14 10661 g002
Applsci 14 10661 g003
Figure 3. We normalized the histogram of LBP values to use it as a 1D feature. The x-axis represents the number of bins and the y-axis represents the frequency.
Figure 3. We normalized the histogram of LBP values to use it as a 1D feature. The x-axis represents the number of bins and the y-axis represents the frequency.
Applsci 14 10661 g003
Applsci 14 10661 g004
Figure 4. Guangzhou dataset samples: Normal, Bacteria, and Virus. Where R is the right side of the body.
Figure 4. Guangzhou dataset samples: Normal, Bacteria, and Virus. Where R is the right side of the body.
Applsci 14 10661 g004
Applsci 14 10661 g005
Figure 5. PA dataset samples: Normal, COVID, and Pneumonia. Where R is the right side of the body and L is the left side of the body.
Figure 5. PA dataset samples: Normal, COVID, and Pneumonia. Where R is the right side of the body and L is the left side of the body.
Applsci 14 10661 g005
Applsci 14 10661 g006
Figure 6. Feature visualization of Guangzhou dataset and PA dataset.
Figure 6. Feature visualization of Guangzhou dataset and PA dataset.
Applsci 14 10661 g006
Applsci 14 10661 g007aApplsci 14 10661 g007b
Figure 7. Confusion matrix of each experiment, comparing the performance of 1DTR with the neural network models (EfficientNet, EfficientNet-V2-Small, and MobileNet-V3-Small). (a) 1DTR Guangzhou dataset confusion matrix. (b) EfficientNet Guangzhou dataset confusion matrix. (c) EfficientNet-V2-Small Guangzhou dataset confusion matrix. (d) MobileNet-V3-Small Guangzhou dataset confusion matrix. (e) 1DTR PA dataset confusion matrix. (f) EfficientNet PA dataset confusion matrix. (g) EfficientNet-V2-Small PA dataset confusion matrix. (h) MobileNet-V3-Small PA dataset confusion matrix.
Figure 7. Confusion matrix of each experiment, comparing the performance of 1DTR with the neural network models (EfficientNet, EfficientNet-V2-Small, and MobileNet-V3-Small). (a) 1DTR Guangzhou dataset confusion matrix. (b) EfficientNet Guangzhou dataset confusion matrix. (c) EfficientNet-V2-Small Guangzhou dataset confusion matrix. (d) MobileNet-V3-Small Guangzhou dataset confusion matrix. (e) 1DTR PA dataset confusion matrix. (f) EfficientNet PA dataset confusion matrix. (g) EfficientNet-V2-Small PA dataset confusion matrix. (h) MobileNet-V3-Small PA dataset confusion matrix.
Applsci 14 10661 g007aApplsci 14 10661 g007b
Table 1. Overall model performance of dataset with 1DTR and CNN models (Global Accuracy represents the overall accuracy across all classes for a given model and dataset).
Table 1. Overall model performance of dataset with 1DTR and CNN models (Global Accuracy represents the overall accuracy across all classes for a given model and dataset).
MethodGlobal AccuracyF1PrecisionRecall
1DTR (PA dataset)96%0.960.960.96
EfficientNet-B0 (PA dataset)91%0.910.920.91
EfficientNetV2 (PA dataset)93%0.930.930.93
MobileNetV3 (PA dataset)91%0.910.910.91
1DTR (Guangzhou dataset)99%0.990.990.99
EfficientNet-B0 (Guangzhou dataset)81%0.800.820.80
EfficientNetV2 (Guangzhou dataset)76%0.750.770.75
MobileNetV3 (Guangzhou dataset)77%0.760.790.76
Table 2. Class-wise performance metrics for the PA dataset (Class-wise performance represents the accuracy calculated separately for each class (Normal, COVID, and Pneumonia) in the PA dataset).
Table 2. Class-wise performance metrics for the PA dataset (Class-wise performance represents the accuracy calculated separately for each class (Normal, COVID, and Pneumonia) in the PA dataset).
NormalCOVIDPneumonia
Acc.F1Pre.Rec.Acc.F1Pre.Rec.Acc.F1Pre.Rec.
1DTR0.96%0.960.930.980.96%0.970.990.950.96%0.950.960.95
EfficientNet-B00.95%0.890.830.960.87%0.930.990.880.89%0.920.950.89
EfficientNet-V20.96%0.910.870.960.93%0.950.970.960.89%0.930.960.90
MobileNet-V30.86%0.900.870.930.93%0.950.960.940.93%0.890.850.94
Table 3. Class-wise performance metrics for the Guangzhou dataset (Class-wise performance represents the accuracy calculated separately for each class (Normal, Virus, and Bacteria) in the Guangzhou dataset).
Table 3. Class-wise performance metrics for the Guangzhou dataset (Class-wise performance represents the accuracy calculated separately for each class (Normal, Virus, and Bacteria) in the Guangzhou dataset).
NormalVirusBacteria
Acc.F1Pre.Rec.Acc.F1Pre.Rec.Acc.F1Pre.Rec.
1DTR0.99%0.990.991.000.99%0.970.990.950.99%0.990.981.00
EfficientNet-B00.68%0.810.980.690.76%0.720.680.760.96%0.870.800.96
EfficientNet-V20.63%0.750.930.630.69%0.680.660.700.92%0.820.730.93
MobileNet-V30.59%0.600.950.730.72%0.690.660.720.96%0.840.740.97
Table 4. Training and testing time of both models using PA and Guangzhou datasets.
Table 4. Training and testing time of both models using PA and Guangzhou datasets.
Training TimeTesting Time
1DTR (PA dataset)38.14 s0.17 s
EfficientNet-B0 (PA dataset)7776.78 s8.56 s
EfficientNet-V2 (PA dataset)7128.57 s7.55 s
MobileNet-V3 (PA dataset)6425.22 s7.24 s
1DTR (Guangzhou dataset)10.85 s0.19 s
EfficientNet-B0 (Guangzhou dataset)6251.16 s7.02 s
EfficientNet-V2 (Guangzhou dataset)5393.67 s6.20 s
MobileNet-V3 (Guangzhou dataset)4920.32 s5.79 s
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Alabdulwahab, A.; Park, H.-C.; Jeong, H.; Lee, S.-W. An Efficient One-Dimensional Texture Representation Approach for Lung Disease Diagnosis. Appl. Sci. 2024, 14, 10661. https://doi.org/10.3390/app142210661

AMA Style

Alabdulwahab A, Park H-C, Jeong H, Lee S-W. An Efficient One-Dimensional Texture Representation Approach for Lung Disease Diagnosis. Applied Sciences. 2024; 14(22):10661. https://doi.org/10.3390/app142210661

Chicago/Turabian Style

Alabdulwahab, Abrar, Hyun-Cheol Park, Heon Jeong, and Sang-Woong Lee. 2024. "An Efficient One-Dimensional Texture Representation Approach for Lung Disease Diagnosis" Applied Sciences 14, no. 22: 10661. https://doi.org/10.3390/app142210661

APA Style

Alabdulwahab, A., Park, H. -C., Jeong, H., & Lee, S. -W. (2024). An Efficient One-Dimensional Texture Representation Approach for Lung Disease Diagnosis. Applied Sciences, 14(22), 10661. https://doi.org/10.3390/app142210661

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop