Classification of Lung Diseases Using an Attention-Based Modified DenseNet Model

Abstract

Lung diseases represent a significant global health threat, impacting both well-being and mortality rates. Diagnostic procedures such as Computed Tomography (CT) scans and X-ray imaging play a pivotal role in identifying these conditions. X-rays, due to their easy accessibility and affordability, serve as a convenient and cost-effective option for diagnosing lung diseases. Our proposed method utilized the Contrast-Limited Adaptive Histogram Equalization (CLAHE) enhancement technique on X-ray images to highlight the key feature maps related to lung diseases using DenseNet201. We have augmented the existing Densenet201 model with a hybrid pooling and channel attention mechanism. The experimental results demonstrate the superiority of our model over well-known pre-trained models, such as VGG16, VGG19, InceptionV3, Xception, ResNet50, ResNet152, ResNet50V2, ResNet152V2, MobileNetV2, DenseNet121, DenseNet169, and DenseNet201. Our model achieves impressive accuracy, precision, recall, and F1-scores of 95.34%, 97%, 96%, and 96%, respectively. We also provide visual insights into our model’s decision-making process using Gradient-weighted Class Activation Mapping (Grad-CAM) to identify normal, pneumothorax, and atelectasis cases. The experimental results of our model in terms of heatmap may help radiologists improve their diagnostic abilities and labelling processes.

Introduction

Lung illness is a major global health concern, generating significant morbidity and mortality around the world. According to the World Health Organization (WHO), lung disorders such as chronic obstructive pulmonary disease (COPD), asthma, lung cancer, and pneumonia account for one out of every six fatalities worldwide, making them one of the main causes of death [1]. Because of the increasing prevalence of lung-related disorders, particularly in developing nations, proper diagnosis of such diseases is critical. Air pollution, smoking, occupational risks, and genetic factors are the principal causes of lung disease [2, 3]. These variables can cause lung tissue inflammation, damage, and scarring, leading to the development of various lung illnesses.

Two major lung diseases are (i) Atelectasis and (ii) Pneumothorax. Atelectasis is a condition caused by factors such as mucus obstructions, tumors, or pressure from outside the lung, which results in a partial or full collapse of lung tissues [4]. Pneumothorax, on the other hand, is generally caused by trauma or injury to the lungs or chest walls, resulting in lung collapse and breathing difficulties due to the presence of air or gas in the pleural cavity [5]. Accurate and timely diagnosis of these conditions from lung images is indispensable for the proper treatment.

Furthermore, early and accurate diagnosis of lung illnesses such as atelectasis might lower the healthcare burden and associated expenses dramatically. The key measures of lung illness detection are advanced screening technologies and diagnostic procedures such as Computed Tomography (CT) scans, X-rays, spirometry, and blood testing [6]. However, radiologists encounter considerable challenges when it comes to accurately analyzing patient health data. Manual identification of lung sickness requires time and is subject to radiologists’ expertise, resulting in delays in diagnosis and treatment [7–9].

X-ray imaging is the most widely used diagnostic technique among all available medical imaging methods. However, raw X-ray images often present several challenges in accurately diagnosing lung diseases. These images can suffer from issues such as low contrast, noise, artefacts, and variations in lighting. These imperfections can obscure critical details and make it challenging for radiologists and Artificial Intelligence (AI) based systems to provide precise diagnoses. To address these challenges, image enhancement techniques are an indispensable component in the medical imaging process. Image enhancement methods, such as contrast stretching, histogram equalization, and noise reduction, improve the quality and clarity of X-ray images. Enhancing the images allows subtle features and abnormalities to be more easily detectable.

In this study, we applied the contrast-limited histogram equalization (CLAHE) enhancement algorithm to improve the image pixel contrast. CLAHE and machine learning techniques were used by Islam and Nahiduzzaman [10] to detect COVID-19 from CT images to achieve an accuracy of 99.73%. A new contrast enhancement method based on brightness preserving dynamic fuzzy histogram equalization and singular value decomposition with discrete wavelet transform techniques was proposed by Mnassri et al. [11] to detect multiple sclerosis from MRI (Magnetic Resonance Imaging) images.

The transfer learning techniques using pre-trained models are playing a significant role in several medical diagnoses, including lung diseases. The models trained on vast datasets with various medical images have proven invaluable in automating the diagnosis process. By leveraging the deep learning capabilities of pre-trained models, computer-aided design systems can recognize complex patterns and abnormalities in X-ray images with high accuracy. This approach speeds up the diagnostic process and ensures a consistent level of precision, making it a vital tool in assisting radiologists and healthcare professionals. Studies have demonstrated the effectiveness of pre-trained deep learning models like ResNet, VGG16, VGG19, Xception, MobilenetV2, Inception, and DenseNet, and they are widely used for the detection and classification of lung diseases like pneumonia [12, 13], COVID-19 [14, 15], pneumothorax [16, 17]. Researchers have reported remarkable success rates in sensitivity and specificity, paving the way for integrating AI-driven solutions into clinical practice. The symbiotic relationship between image enhancement techniques and pre-trained models can revolutionize the field of diagnostic radiology, offering faster and more accurate diagnoses of lung diseases and improving patient health and early recovery [18].

However, very few works have been reported for the identification of lung collapse diseases such as atelectasis and pneumothorax. Motivated by the applications of pre-trained models in medical settings and the need for accurate lung disease diagnosis, the current article aimed to improve the classification of pneumothorax and atelectasis using chest X-ray images. Several benchmark pre-trained models were implemented to determine the best-performing model for improving the proposed model.

The major contributions of the current article are:

• Modification in the DenseNet201 model to improve identification of atelectasis and pneumothorax using chest X-ray images only.

• A channel attention block is proposed in between each dense block except the last block to better capture the salient features of the images for disease classification.

• A hybrid pooling layer is added to the proposed model to improve its learning capabilities.

• Furthermore, the performance of five established pre-trained models and their respective variants, VGG [19], Xception [20], MobileNet [21], ResNet [22], and DenseNet [23] are evaluated to classify lung collapse diseases, namely atelectasis, pneumothorax, and healthy lung conditions using X-ray images.

This paper is divided into seven sections. "Related Work" discusses the works related to our proposed approach. "Proposed Approach" section illustrates the proposed approach; "Experiment" section describes the experimental details; and "Results" and "Discussion" sections contain the results and discussion. "Conclusion and Future Work" section concludes the paper by summarizing the current research’s findings and seamlessly explores future avenues for improving the proposed method.

Related Work

Deep neural networks have gained significant prominence in medical imaging and are widely used in computer-aided design (CAD) systems. These systems leverage deep learning algorithms to enhance various aspects like image classification [14, 24, 25], object detection [26], image segmentation [24, 27, 28], and image restoration [29, 30], to name a few.

Chan et al. [31] introduced an approach to detect pneumothorax by combining local binary pattern (LBP) and support vector machine (SVM) techniques. They validated the efficacy of their proposed method using a private dataset. They employed gamma correction to enhance the original images’ mean intensity and five-fold cross-validation to prevent overfitting. Additionally, the authors performed lung region segmentation by eliminating background and noise, utilizing the Sobel edge detection technique to identify rib boundaries. The method achieved an accuracy of 82.2% through five-fold cross-validation. Filice et al. [17] addressed the issue of limited and time-consuming manual annotations for pneumothorax in the National Institutes of Health (NIH) Chest X-ray dataset [32]. The authors proposed a crowdsourcing method that employs machine learning annotations to accelerate and improve the annotation process. They leverage the power of existing machine learning models to generate initial annotations, which are then refined through a crowd-based iterative process. The authors showed that the utilization of the Machine Learning Annotation (MLA) technique expedites the annotation process with relatively high sensitivity at the expense of specificity. Wang et al. [9] proposed an attention-based variant of DenseNet to automatically learn and highlight informative features of pneumonia from X-ray images. The authors used a channel attention block named Squeeze and Excitation (SE) in the DenseNet model to highlight pneumonia information and achieved a classification accuracy of 92.8%, which is 2.4% higher than the existing DenseNet model. Xception model, along with a $5\times 5$ kernel Gaussian filter, was used by Guler and Polat [18] to classify pneumonia from chest X-ray images to achieve an accuracy of 95.73%. The authors compare the performance of different deep transfer learning methods for pneumonia detection from 5856 chest X-ray images.

CheXNeXt was designed by Rajpurkar et al. [33] to classify 14 lung diseases from chest X-ray images. They used the DenseNet121 model with a weighted binary cross-entropy loss function, where the per-class weights were determined according to the frequency of each class within the training set. Souid et al. [25] used the MobileNet V2 model along with global average pooling and a dense layer of 512 neurons to classify 14 lung diseases. The model achieved an accuracy of 90.11% and an average Area Under the Curve (AUC) of 0.801. Chutia et al. [34] conducted a study where they applied two image enhancement algorithms, namely Discrete Wavelet Transform (DWT) and Contrast-Limited Adaptive Histogram Equalization, to enhance lung X-ray images. During the pre-processing phase, the Conv2D layers are utilized to extract the main features of the improved images. These retrieved features are then input into a DenseNet201 model, which enables efficient and reliable classification.

Tian et al. [35] proposed a deep learning-based approach that combines transfer learning and multi-instance learning techniques for improved pneumothorax classification. The proposed approach used two public datasets to learn local features, followed by global features and achieved an accuracy of 94.4%. Albahli et al. [36] presented an approach that utilizes three deep learning algorithms: ResNet152V2, DenseNet121, and InceptionResNetV2. The proposed model is trained and tested on a dataset consisting of chest X-ray images to classify 14 different lung diseases. The study focuses on developing an effective classification system using these deep learning models for accurate disease diagnosis. Rajasenbagam et al. [13] classified pneumonia and healthy X-ray images using a Generative Adversarial Network (GAN)-based algorithm combined with a VGG19 network. The proposed model is trained on 15,900 X-ray images of the chest and obtained 99.34% accuracy with test images.

The study in [37] focused on enhancing the pre-trained models for the multi-label classification of X-ray images. Instead of using global average pooling, the authors employed adaptive dropout. Since X-ray images can exhibit multiple diseases, the classification problem becomes a multi-label challenge. The authors tackled it by dividing it into multiple binary problems, leading to improved outcomes. Bharti et al. [38] proposed a hybrid model named VDSNet using VGG, Spatial Transformer Network (STN), and data augmentation with a standard CNN. A tabular representation of prominent works done in the area of lung disease identification, highlighting the models, datasets, their performance, and limitations, is presented in Table 1. Aswiga and Shanthi [39] proposed a two-level framework that employs multilevel and feature extraction transfer learning for classifying breast tomosynthesis. Using this feature extraction and transfer learning approach, they achieved an AUC of 0.89.

Table 1.

Analysis of existing methods

Reference	Objective	Dataset	Method	Acc	Limitations
Wang et al. [9]	Classification	Chest X-ray2017	Attention-Based DenseNet	92.8%	Limited comparison against the state-of-the-art pneumonia classification models.
Rajasenbagam et al. [13]	Classification	NIH	VGG19 + DCGAN	99.34%	Only the comparison against the transfer learning technique is provided.
Filice et al. [17]	Segmentation	NIH Dataset	Mask R-CNN	78%	A proper workflow of machine learning annotation is not provided.
Souid et al. [25]	Classification	NIH Dataset	MobileNet V2	90.11%	Accuracy and sensitivity is less.
Chan et al. [31]	Segmentation	Private Dataset	LBP + SVM	82.2	A comprehensive comparison with existing medical image segmentation and feature analysis methods is not included.
Wang et al. [32]	Classification	NIH Dataset	ResNet50	80.9%	Sensitivity and specificity is less.
Guler and Polat [18]	Classification	Labeled Optical Coherence Tomography and Chest X-Ray Images for Classification	Xception	95.73%	The study doesn’t explore the customization of neural network architectures or the incorporation of domain-specific knowledge.
Tian et al. [35]	Classification	NIH and SAHZUSM	ResNet	94.4%	No comparison against the state-of-the-art methods.
Albahli et al. [36]	Classification	NIH	ResNet152 V2, DenseNet121, and InceptionResNetV2	41.02%	Model accuracy is very less.
Zhang et al. [37]	Montgomery County, Shenzhen, and NIH	modified ResNet		87.7%	Accuracy is not suitable for real-time implementation.
Bharti et al. [38]	Classification	NIH	VDSNet	73%	Accuracy is not suitable for real-time implementation.

The authors in [40] devised a framework that utilizes a deep learning algorithm to accurately identify lung cancer by classifying nodules as either benign or malignant. The integration of the Adaptive Tree Seed Optimization (ATSO) technique is used to alter the Convolution Neural Network (CNN) and create an Improved convolution neural network (ICNN) for classification purposes. The methodology demonstrated a peak accuracy of 96.5% and outperformed existing state-of-the-art technologies in terms of performance. Yuan et al. [41] introduced a multi-modal fusion approach using a multi-branch classification network to identify and categorize lung nodules. The data utilized consisted of (i) radiological data pertaining to pulmonary nodules, (ii) feature maps obtained from various depths of the network, each comprising distinct receptive field information, and (iii) multi-modal features. The experimental findings demonstrate that this model is capable of accurately categorizing benign and malignant pulmonary nodules for clinical diagnosis, achieving the highest levels of accuracy (94.89%), sensitivity (94.91%), and F1-score (94.65%), while maintaining the lowest false positive rate (5.55%).

Various studies have utilized machine learning and deep learning techniques to accurately classify and diagnose lung diseases based on chest X-ray images. However, the existing literature does not explicitly focus on classifying diseases that cause a collapsed lung. This highlights a potential research gap and suggests an opportunity for further investigation in this area.

Proposed Approach

The proposed approach consists of (i) a pre-processing phase where the CLAHE image enhancement algorithm is used to improve the visibility of image details and (ii) transfer learning with modified DenseNet to extract features and classify. The core of our approach lies in the modified DenseNet201, where we have introduced several enhancements to the existing models to address the challenges of lung disease classification. We have replaced the second layer of the original DenseNet model (a max pooling layer) with a hybrid pooling layer. This pooling approach empowers our model to extract more detailed information from the initial convolution layer, thereby improving feature extraction and representation. Furthermore, we have integrated a channel attention mechanism into each dense block of DenseNet201, except for the final dense block. This attention mechanism acts as a spotlight, selectively amplifying the information associated with disease areas while preserving the nuances related to healthy classes. This selective attention is crucial for the model’s ability to make accurate classifications. The architecture of our proposed model is visually represented in Fig. 5. In the following sections ("Effect of Image Enhancement", "Hybrid Pooling", "Channel Attention Blocks" and "Modified DenseNet-201"), we discussed the working of each component in detail.

Fig. 5.

Proposed architecture

Effect of Image Enhancement

In our study, we utilized the CLAHE technique to improve the visual quality and noise suppression of X-ray images by constraining the contrast enhancement applied to each pixel [42]. Specifically, we adopted a tile size of $8\times 8$ and set the clip limit to 3.0. Figure 1 visually represents (a) the original image, (b) the CLAHE-enhanced image, and (c) the delta image (illustrating the differences between the original and CLAHE-enhanced images). The delta image shown in Fig. 1(c) illustrates pixel-wise variations resulting from the CLAHE process, highlighting regions where the two images exhibit differences. The “delta image” serves as a visual map, helping us to understand the impact of CLAHE image enhancement.

Fig. 1.

Image enhancement and difference analysis: a shows the original image, b displays the CLAHE image, and c presents the delta image, illustrating differences between the original and CLAHE-enhanced images

Hybrid Pooling

In our proposed model, the first max pooling layer was replaced with a hybrid pooling layer, enhancing the feature extraction process for the initial dense block. The remaining pooling layers in the model remained unchanged. During the hybrid pooling process, both average pooling and max pooling operations were applied separately to the feature map, and the results of these operations were then added together. This hybrid approach combined the benefits of both pooling methods, capturing different aspects of the features in the image and enhancing the overall representation. The combined output was subjected to batch normalization. Batch normalization normalized the activation within a layer, ensuring stable and consistent training by reducing internal covariance shifts [43, 44].

Channel Attention Blocks

The channel attention module in DenseNet was used to capture and emphasize informative features within different channels of a convolutional neural network. Its purpose was to enhance the representation capabilities of the network by focusing on important channel-wise information while suppressing less relevant features [45, 46].

In the initial step of channel attention, we have applied Global Average Pooling (GAP) to the feature maps obtained from the dense block. The GAP layer reduces the spatial dimensions and aggregates information across the entire feature map, a step known as the “squeeze” operation. In the second step, the pooled features are passed through fully connected layers. The first fully connected layer acts as a bottleneck, reducing the dimensionality of the input data to extract the most relevant features. The size of the first fully connected layer was equal to the number of channels in the GAP, divided by 16. By decreasing the size of the first fully connected layer, the computational complexity and parameter space of the network are reduced. The size of the second fully connected layer equals the number of channels in the GAP. A “sigmoid” activation function is used to obtain attention weights. These weights represent the importance of each channel, reflecting their relevance to the task. This step is also referred to as the “excitation” operation. The second fully connected layer is the “attention” layer, as we multiplied the output feature map obtained from the dense blocks by $(F{C}_2)$ to get channel-refined features. Figure 2 demonstrates the architecture of the channel attention module. As shown, $F_x$ was the input feature map received from the dense blocks, where H and W represented the height and width, and C was the number of channels. In the next step, global average pooling was applied to $F_x$, resulting in $(1\times 1\times C)$, where $(1\times 1)$ was the dimension of each channel (C). $(F{C}_1)$ was the first fully connected layer with neurons set to 16, 32, and 112 for the first, second, and third attention blocks, respectively. Similarly, $(F{C}_2)$ was the second fully connected layer, with neurons set to 256, 512, and 1792 for the first, second, and third attention blocks, respectively.

Fig. 2.

Channel attention module

Modified DenseNet-201

The DenseNet [23] architecture consists of an initial $7\times 7$ convolution layer followed by a $3\times 3$ max pooling layer with a stride equal to 2, four dense blocks, and three transition layers. Each dense block consists of a series of $1\times 1$ and $3\times 3$ convolutional layers with the same output size, followed by a concatenation operation that combines the outputs of all previous layers. The transition layers are used to reduce the spatial dimensions of the feature maps between dense blocks. Each transition layer performs $1\times 1$ convolution and $2\times 2$ average pooling. This helps to control the network’s growth and improves model efficiency. The classification layer consists of a softmax activation function at the network’s end. Figure 3 demonstrates the overall architecture of DenseNet, unlike traditional CNN models, where N connections for N-layers, in DenseNet, there are $N(N+1)/2$ connections in an N-layer network, which ultimately helps in feature reuse and information propagation. Figure 4 demonstrates the number of connections in N-layers.

Fig. 3.

DenseNet architecture

Fig. 4.

Number of connections in a dense block

The proposed model takes an input image of size (224, 224, 3) and applies a $7\times 7$ convolution $(conv\_1)$ with a stride of 2. Next, max and average pooling operations are executed on the resulting feature map, each utilizing a $3\times 3$ pool size and a stride of 2. This yields feature maps from max and average pooling, each of size (56, 56, 64). These two feature maps are combined through element-wise addition using the “add” layer. The resulting feature map maintains its original dimensions, staying at (56, 56, 64), and proceeds to undergo batch normalization. The batch normalization step involves adjusting the feature map by subtracting the batch mean and dividing it by the batch standard deviation. This process is pivotal in preserving consistent scale and feature distribution across the network, enhancing learning. The first dense block $(D_1)$ consists of 6 repetitions batch normalization layer, ReLU activation, a $1\times 1$ convolution, another batch normalization layer, ReLU activation, and a $3\times 3$ convolution. The output feature map after $(D_1)$ has a shape of (56, 56, 256). The channel attention mechanism is applied after $(D_1)$, which includes a global average pooling layer, followed by two fully connected layers with “sigmoid” activation. The output of the channel attention block is obtained by multiplying the attention weights with the feature map, resulting in an output shape of (56, 56, 256). Afterwards, the transition layer $(T_1)$ reduces the number of filters by half through a $1\times 1$ convolutional layer. The resulting output feature map has a shape of (56, 56, 128). The transition layer further downsamples the feature map by applying average pooling with a pool size of $2\times 2$ and a stride of 2. This reduces the spatial dimensions to half and produces an output feature map with a shape of (28, 28, 128).

Another dense block $(D_2)$ with 12 repetitions is applied to the feature map obtained from the $(T_1)$. The output shape after the $(D_2)$ is (28, 28, 512). The channel attention mechanism is applied again to the output feature map of $(D_2)$, resulting in an output shape of (28, 28, 512). The second transition layer $(T_2)$ reduces the number of filters to 128 and downsamples the feature map to (14, 14) . So, the output shape after $(T_2)$ is (14, 14, 128). The third dense block ($D_3$) consists of 48 repetitions and produces an output of shape (14, 14, 1792) followed by the third attention block, which produces an output of shape (14, 14, 1792). The output of the third transition layer $(T_3)$ is (7, 7, 896). The dense block 4 ($D_4$) consists of 48 repetitions and produces output shape of (7, 7, 1920). The global average pooling layer is applied to the output produced by ($D_4$), which reduces the spatial dimensions of the feature map to $1\times 1$, resulting in an output shape of (1, 1, 1920). The output layer is a fully connected classification layer with “softmax” activation that produces the final classification probabilities. The output shape is (3), corresponding to the atelectasis, pneumothorax, and healthy images. The summary of the image sizes produced at various stages of the proposed model is presented in Table 2.

Table 2.

Details about the image sizes at various stages of the proposed model

Layers	Output Size	Details
Input Layer	$224\times 224\times 3$
Convolution Layer	$112\times 112\times 64$	$7\times 7$ conv, stride =2
Max Pooling	$56\times 56\times 64$	$3\times 3$ Max pool, stride =2
Average Pooling	$56\times 56\times 64$	$3\times 3$ Avg pool, stride =2
Hybrid Pooling	$56\times 56\times 64$	[ Both max & avg pooling]
Dense Block 1	$56\times 56\times 256$	[$1\times 1$ conv, $3\times 3$ conv] $\times \mathbf{6}$
Attention Block 1	$56\times 56\times 256$	GAP, Dense layers
Transition Layer 1	$56\times 56\times 128$	$1\times 1$ conv
	$28\times 28\times 128$	$2\times 2$ Avg pool, stride =2
Dense Block 2	$28\times 28\times 512$	[$1\times 1$ conv, $3\times 3$ conv] $\times \mathbf{12}$
Attention Block 2	$28\times 28\times 512$	GAP, Dense layers
Transition Layer 2	$28\times 28\times 256$	$1\times 1$ conv
	$14\times 14\times 256$	$2\times 2$ Avg pool, stride =2
Dense Block 3	$14\times 14\times 1792$	[$1\times 1$ conv, $3\times 3$ conv] $\times \mathbf{48}$
Attention Block 3	$14\times 14\times 1792$	GAP, Dense layers
Transition Layer 3	$14\times 14\times 896$	$1\times 1$ conv
	$7\times 7\times 896$	$2\times 2$ Avg pool, stride =2
Dense Block 4	$7\times 7\times 1920$	[$1\times 1$ conv, $3\times 3$ conv] $\times \mathbf{32}$
Global Average pooling	$1\times 1\times 1920$	$7\times 7$ global average pooling

Experiment

This section describes the dataset, implementation details, and evaluation metrics used in the proposed approach to classify pneumothorax, atelectasis, and healthy X-ray images.

Dataset Description

The NIH chest X-ray dataset is a large publicly available dataset of chest X-ray images [32]. The dataset consists of 112,120 X-ray images, out of which 9409 images were considered for analysis as shown in bold in Table 3. It includes 4215 atelectasis images, 3000 healthy images, and 2194 pneumothorax images.

Table 3.

Dataset details

Class	No. of images in each class
Atelectasis	4215
Pneumothorax	3000
Healthy	2194
Total	9409

The total dataset is split into 80% training and 20% testing ratios. Later, during the training phase, 20% of the data is considered for validation from the training set. Table 3 shows the count of images used in the evaluation process. Figure 6 shows sample images for each class, where “a” represents healthy images, “b” represents pneumothorax images, and “c” represents atelectasis images.

Fig. 6.

Sample X-rays used to train and test the proposed model for detecting lung diseases: a Healthy, b Pneumothorax, and c Atelectasis

Experimental Setup

All the experiments are performed on Python 3.7.12 using Keras, Numpy, Pandas and Tensorflow 2.0 libraries. The models trained using categorical cross-entropy as a loss function, an Adam optimizer with a learning rate of 0.0001 for 100 epochs; each epoch contains 187 steps, and each step processes 16 images. The hyperparameters used in the proposed model are provided in Table 4.

Table 4.

The hyperparameters used in the proposed approach

Hyperparameters	Values
Optimizer	Adam
Loss function	Categorical cross-entropy
Learning rate	0.0001
Epoch	100
Batch size	32
Step per epoch	187

Evaluation Metric

In this study, lung X-ray images are classified into three classes, and the performance of the proposed model is assessed using accuracy, precision, recall, F1-score, and AUC values.

The aforementioned matrices are calculated using True Positive $(T_{\mathit{pos}})$, False Positive $(F_{\mathit{pos}})$, True Negative $(T_{\mathit{neg}})$, and False Negative $(F_{\mathit{neg}})$ instances. $(T_{\mathit{pos}})$ refers to the number of correctly predicted positive instances, $(F_{\mathit{pos}})$ refers to the number of negative instances predicted as positive, $(T_{\mathit{neg}})$ refers to the number of negative instances predicted as negative, and $(F_{\mathit{neg}})$ refers to the number of positive instances predicted as negative. The following equations (1 to 5) denote the precision, recall, F1-score, and AUC.

$$\begin{array}{c}\hfill Accuracy=\frac{{\text{T}}_{\mathit{pos}}+T_{\mathit{neg}}}{T_{\mathit{pos}}+T_{\mathit{neg}}+F_{\mathit{pos}}+F_{\mathit{neg}}}\end{array}$$ 1

$$\begin{array}{c}\hfill Precision=\frac{{\text{T}}_{\mathit{pos}}}{T_{\mathit{pos}}+F_{\mathit{pos}}}\end{array}$$ 2

$$\begin{array}{c}\hfill Recall=\frac{{\text{T}}_{\mathit{pos}}}{T_{\mathit{pos}}+F_{\mathit{neg}}}\end{array}$$ 3

$$\begin{array}{c}\hfill F1-score=2\times \frac{Precision\times Recall}{Precision+Recall}\end{array}$$ 4

$$\begin{array}{c}\hfill AUC=\frac{1}{2}\left(\frac{T_{\mathit{pos}}}{\mathit{TotalPositive}}+\frac{T_{\mathit{neg}}}{\mathit{TotalNegative}}\right)\end{array}$$ 5

Results

Several experiments were conducted to assess the performance of the proposed model for disease classification in lung X-ray images. The results obtained are discussed in this section.

Classification Results of the Proposed Model

To better understand the model behavior, we created two scenarios: (i) a binary classification between healthy vs infected with any lung disease, and (ii) a Multi-class classification was performed for atelectasis, pneumothorax, and healthy.

Healthy vs Infected Diseases

For binary classification, three experiments were conducted. Table 5 shows the performance of each classification: pneumothorax vs healthy, atelectasis vs healthy and infected vs healthy. Our proposed model is quite accurate while classifying pneumothorax vs. healthy with an average precision of 98%, an average recall of 95%, an average F1-score of 96%, an average AUC value of 0.97, and an accuracy of 98.25%, as shown in bold in Table 5, compared to atelectasis vs. healthy having an accuracy of 95.21% only and infected vs. healthy having an accuracy of 93.21 only.

Table 5.

Performance of binary classifications performed on the proposed model

Classification	Pathology	Accuracy	Precision	Recall	F1-Score	AUC
	Pnemothourax		99%	96%	97%	0.98
Pnemothourax vs Healthy	Healthy	98.25%	96%	94%	95%	0.95
	Average		98%	95%	96%	0.97
	Atelectasis		96%	97%	96%	0.97
Atelectasis vs Healthy	Healthy	95.21%	94%	89%	91%	0.97
	Average		95%	93%	94%	0.97
	Infected		98%	96%	97%	0.94
Infected vs Healthy	Healthy	93.21%	85%	87%	86%	0.94
	Average		92%	92%	92%	0.94

Classification of Atelectasis, Pnemouthorax, and Healthy Classes

In this subsection, the result of multi-class classification is discussed. Table 6 depicts the performance of our proposed model on multi-class classification. Among all the classes, our model classified pneumothorax more accurately than other classes, with precision, recall, and an F1-Score of 98%, 98% and 97%. Atelectasis is identified with second best performance measure, having precision, recall, and an F1-Score of 98%, 96% and 95%, and the healthy class is identified as having the least performance with precision, recall, and an F1-Score of 95%, 93%, and 96%.

Table 6.

Performance of multi-class classifications performed on the proposed model

Pathology	Accuracy	Precision	Recall	F1-Score	AUC
Atelectasis		98% ($\pm 1.39$)	96% ($\pm 2.12$)	95% ($\pm 2.13$)	0.98 ($\pm 1.11$)
Pneumothorax	95.34% ($\pm 3.28$)	98% ($\pm 0.48$)	98% ($\pm 0.40$)	97% ($\pm 0.28$)	0.99 ($\pm 0.12$)
Healthy		95% ($\pm 3.64$)	93% ($\pm 2.8$)	96% ($\pm 3.6$)	0.99 ($\pm 0.52$)
Average		97% ($\pm .11$)	96% ($\pm 1.29$)	96% ($\pm 0.56$)	0.99 ($\pm 0.33$)

The proposed model shows a better accuracy of 98.25% on the binary classification of pneumothorax vs. healthy. However, the accuracy declines by 2.91% on multi-class classification. AUC (Area Under the Curve) is also used to evaluate the performance of the proposed model. Figures 7 and 8 illustrate the AUC and confusion matrix of the proposed model, respectively. The AUC of pneumothorax and healthy is 99%, and for atelectasis, the AUC is 98%. The values (\pm x) in Table 6 show the variations of the results across various runs. 

Fig. 7.

ROC curves for the proposed model over 100 epochs

Fig. 8.

Confusion Matrix of the proposed model

Comparison with Other Pre-trained Models

Twelve existing pre-trained models: VGG16, VGG19, InceptionV3, Xception, ResNet50, ResNet152, ResNet50V2, ResNet152V2, MobileNetV2, DenseNet121, DenseNet169, and DenseNet201, are trained with the current dataset and validated. The performance of each benchmark model is depicted in Table 7. The addition of channel attention (CA) blocks to the DenseNet201 model increases the precision by 6%, as shown in the second last row of Table 7. The use of CLAHE to improve the contrast of the images boosts the precision by 2%, as shown in the bold in the last row of Table 7.

Table 7.

The classification performance of the proposed model with CLAHE enhanced images compared to other CNN models

CNN Models	Precision	Recall	F1-score	Trainable Parameters
VGG16	40%	53%	48%	14,716,227
VGG19	42%	53%	47%	20,025,923
InceptionV3	76%	75%	75%	21,808,931
Xception	79%	63%	70%	20,867,627
ResNet50	69%	68%	68%	23,593,859
ResNet152	55%	52%	53%	58,377,091
ResNet50V2	64%	65%	64%	23,570,947
ResNet152V2	58%	55%	56%	58,337,795
MobileNetV2	61%	62%	61%	2,261,827
DenseNet121	76%	73%	74%	7,040,579
DenseNet169	79%	69%	74%	12,647,875
DenseNet201	89%	87%	88%	18,327,747
Proposed Model with raw images	95%	96%	95%	18,561,541
Proposed Model with CLAHE enhanced images	97%	96%	96%	18,561,541

Table 8 briefly compares the performance of the state-of-the-art and our proposed approach based on the AUC. All the existing works are multi-class classifications of 8 to 14 lung diseases, including atelectasis and pneumothorax and excluding the healthy class. The AUC values obtained by our model are shown in bold in the last column of Table 8.

Table 8.

Comparisons of AUC values of existing works and proposed approach

Pathology	Wang et al. [32]	Yao et al. [47]	Rajpurkar et al. [48]	Souid et al. [25]	Albahli et al. [49]	Guan et al. [50]	Proposed Approach
Atelectasis	0.71	0.77	0.86	0.80	0.74	0.85	0.98
Pneumothorax	0.80	0.84	0.94	0.87	0.89	0.92	0.99
Healthy	—	—	—	—	—	—	0.99
Average	0.76	0.81	0.90	0.84	0.82	0.89	0.99

Compression with Different Sampling Methods

We comprehensively evaluated our model using various sampling techniques, including stratified, random, and holdout sampling. The goal was to assess the model’s performance under different conditions. We also experimented with various batch sizes and learning rates to understand how they impact the model’s training and results. This extensive evaluation allowed us to gain valuable insights into the model’s robustness and adaptability across diverse data subsets and learning rate settings, providing a well-rounded perspective on its capabilities and limitations. Table 9 presents a thorough analysis of various sampling techniques in combination with different batch sizes (16, 32, 64) and learning rates (0.001, 0.0001, and 0.0001). As illustrated in Table 9, various sampling methods exhibit distinct performance characteristics. Among them, holdout sampling stands out as the top performer, achieving remarkable results with a batch size of 16 and a learning rate of 0.0001, boasting an accuracy of 96%, precision of 98%, recall of 96%, and F1-score of 95%, as shown in bold in the twentieth row of Table 9. Stratified sampling also demonstrates strong performance, particularly in accuracy, when using a batch size of 16 and a learning rate of 0.001, reaching an accuracy of 97%, precision of 97%, recall of 95%, and an F1-score of 94%, as shown in bold in the first row of Table 9. On the other hand, random sampling excels in accuracy with a batch size of 64 and a learning rate of 0.001, achieving an accuracy of 90%, precision of 92%, recall of 80%, and an F1-score of 72%.

Table 9.

Comparisons of the proposed model at different batch sizes and learning rates

Sampling Technique	Batch Size	Learning Rate	Accuracy	Precision	Recall	F1-score
		0.001	97%	97%	95%	94%
	16	0.0001	95%	94%	97%	94%
		0.00001	95%	92%	95%	92%
		0.001	97%	93%	97%	96%
Stratified	32	0.0001	96%	88%	96%	90%
Sampling		0.00001	95%	92%	95%	94%
		0.001	97%	83%	97%	90%
	64	0.0001	97%	83%	97%	90%
		0.00001	94%	82%	94%	85%
		0.001	89%	64%	89%	75%
	16	0.0001	89%	64%	89%	75%
		0.00001	89%	64%	89%	75%
		0.001	88%	74%	88%	76%
Random	32	0.0001	89%	64%	89%	75%
Sampling		0.00001	89%	64%	89%	75%
		0.001	90%	92%	80%	72%
	64	0.0001	85%	84%	85%	77%
		0.00001	89%	63%	87%	70%
		0.001	94%	95%	94%	93%
	16	0.0001	96%	98%	96%	95%
		0.00001	88%	68%	88%	83%
		0.001	89%	88%	89%	87%
Holdout	32	0.0001	89%	88%	86%	87%
Sampling		0.00001	89%	88%	89%	87%
		0.001	95%	84%	95%	80%
	64	0.0001	89%	83%	84%	85%
		0.00001	88%	81%	88%	85%

Model Interpretation and Statistical Significance Test

To interpret the predictions of the model, we generate heatmaps utilizing class activation mappings (CAMs), which visually depict the most significant regions within the image affected by the diseases [51]. To generate the CAMs, we input an image into the fully trained network and extract the feature maps that are produced by the last convolutional layer. Let the $C\_\mathit{conv}$ be the last convolutional layer, and $F_{map_K^{}}^{}$ the $k^{\mathit{th}}$ feature map and $W_{C\_\mathit{conv},k}$ be the weights of $C\_\mathit{conv}$ of $k^{\mathit{th}}$ feature map leading to pathology p.

Then CAM will generate $M_{C\_\mathit{conv}}$, which contains the most salient features used in classifying the image having pathology p. Equation 6 describes the working of CAMs:

$$\begin{array}{c}\hfill M_{C\_\mathit{conv}}=\sum _k{W}_{C\_\mathit{conv},k}F_{map_K^{}}^{}\end{array}$$ 6

Figure 9 shows the heat map generated by the proposed model in the classification task. As shown, the model identifies the most important features used for identifying pathology p. Along with the heat map generation, a probability prediction on random images is made to identify how accurately our proposed model predicts p.

Fig. 9.

Heat maps generated by the proposed model

We performed hypothesis testing and formulated the following hypotheses to ensure our results are statistically consistent across multiple runs. These hypotheses are pivotal in our statistical analysis to establish a meaningful conclusion from the results.

• Null hypothesis ($H_0$): The test results differ significantly among different runs.

• Alternative hypothesis ($H_1$): The test results do not differ significantly across different runs.

Our results are statistically significant; we performed a t-test at a 5% significance level (95% confidence level) and ran the experiments ten times to ensure robustness and eliminate potential training-related ambiguity. The calculated t-statistics for accuracy, precision, recall, and F1-score are as follows: 1.264, 0.2941, 1.263, and 0.902, respectively. All these values fall below the critical t-value of 1.833 at a significance level ($\alpha$) of 0.05. Consequently, we reject the null hypothesis. This statistical perspective indicates no significant differences in the results across the various experimental runs.

Discussion

One of the significant findings of the current research is a modified DenseNet model for the identification of diseases having collapsed lungs. The DenseNet201 model is modified by using hybrid pooling and incorporating attention blocks within each dense block except the last one. DenseNet201 is considered a base model after comparing its performance with other existing CNN models. The performance of each CNN model is discussed in Table 7. The addition of attention blocks within each dense block aimed to improve the model’s ability to focus on informative features and suppress irrelevant ones. Attention mechanisms have shown promising results in various computer vision tasks by allowing the model to selectively attend to important regions [9, 52]. By integrating attention blocks, the modified DenseNet201 exhibits enhanced discriminative power and better captures subtle patterns in the images. Figure 10 displays the visual analysis of the feature maps produced by the attention block. The other finding is the inclusion of combined features from both max pooling and average pooling on the second layer of the DenseNet model. This approach empowers the model to capture fine-grained details, leading to an increase in accuracy.

Fig. 10.

Heat maps generated by the proposed model at different channel attention blocks

The third finding is that the use of limited adaptive histogram equalization to enhance the input images allows the modified DenseNet201 to receive better visibility of image leading to improved performance. CLAHE improves precision by 2% and F1-score by 1%.

Comparing classification methods, our model performs better on binary class classification with a 98.25% accuracy for pneumothorax vs healthy, in contrast to the accuracy of 95.34% on multi-class classification. Choosing a proper batch size also impacts the efficiency of the model’s performance. After experimenting with various batch sizes, we observed that using a batch size of 16 yielded improved results for the proposed model.

In our proposed model, we also explored various activation functions, including Rectified Linear Unit (ReLU), Parametric Rectified Linear Unit (PReLU), Leaky ReLU, and Exponential Linear Unit (ELU), with their respective performances depicted in Fig. 11. As depicted, ReLU and PReLU exhibit nearly identical precision and F1-scores. However, PReLU outperforms all other functions in terms of accuracy, while ReLU has the highest recall among them. ELU and Leaky ReLU did not yield satisfactory results. Additionally, we investigated different optimizers, namely Adam, SGD, and RMSprop, while keeping the other hyperparameters constant, as depicted in Table 4. When compared to RMSprop and SGD, the result that Adam optimizer obtained is better, as shown in bold in Table 10. RMSprop achieved an accuracy rate of 94.20% and demonstrated a strong recall of 95%, showcasing its effectiveness in identifying relevant data. In contrast, SGD attained a lower accuracy rate of 85.01%. However, Adam outperformed all the optimizers with an impressive accuracy rate of 95.34%, with a precision of 97% and a recall of 96%. These results highlight Adam as the standout performer among the optimizers in terms of both accuracy and precision. Furthermore, the integration of attention blocks within different dense blocks is investigated and discovered that incorporating attention after the first, second, and third dense blocks results in improved performance. The performance of the model on various batch sizes and different attention mechanisms is presented in Table 11, where the best results are shown in bold.

Fig. 11.

The evaluation of various activation functions in our proposed model’s performance

Table 10.

Effect of different optimizer on the proposed model

Optimizer	Accuracy	Precision	Recall	F1-score
RMSprop	94.20%	88%	95%	96%
SGD	85.01%	82%	86%	85%
Adam	95.34%	97%	96%	96%

Table 11.

The performance of DenseNet201 with incorporated channel attention on different dense blocks for classification

CNN Models	Batch Size	Precision	Recall	F1-score
DenseNet201	16	89%	87%	88%
DenseNet201	32	87%	84%	85%
DenseNet201	64	82%	83%	82%
DenseNet201+ CA + SA	92	88%	90%	67%
DenseNet201+ CA on $4^{\mathit{th}}$ Block	16	91%	88%	89%
DenseNet201+ CA on $3^{\mathit{rd}}$ Block	16	91%	88%	89%
DenseNet201+ attention on $3^{\mathit{rd}}$ Block and $4^{\mathit{th}}$ block	16	94%	93%	93%
Proposed model without image enhancement	16	95%	96%	95%
Proposed model with CLAHE-enhanced image	16	97%	96%	96%

The modifications made to the DenseNet201 model by incorporating hybrid pooling, attention blocks within each dense block, and enhancing the images using CLAHE resulted in improved classification performance. The proposed model highlights the importance of attention mechanisms and pre-processing techniques in enhancing the discriminative power and feature representation of deep learning models. The choice of DenseNet201 as the base model is justified through a comprehensive comparison with other CNN architectures. The proposed work can contribute to the growing body of research in deep learning and provide valuable insights for improving the performance of CNN models in image classification tasks.

The proposed model is limited in terms of a large number of parameters. The proposed model has over 18 million parameters, making it a heavy-weight model. Another limitation is that the proposed model has to be evaluated on similar datasets. The proposed model has yet to be evaluated for other diseases, including pneumonia, mass, and nodules. Also, the model is not validated using other datasets.

Conclusion and Future Work

In this paper, we propose a modified DenseNet201 architecture that integrates attention mechanisms and hybrid pooling to promote classification performance. The input X-ray images are first pre-processed to remove noises and enhance the image visually using the CLAHE image enhancement algorithm. We have added a hybrid pooling layer by replacing the conventional max pooling in the DenseNet model. We have also generated a heatmap of the images to highlight the pixels helping in classifying an image into a specific class. We also did extensive experiments with different batch sizes, optimizers, and activation functions to fine-tune our model’s performance and ensure its optimal configuration for precise disease identification. We also used various sampling techniques to create our train and test samples and made various runs of the system to find that the system is yielding statistically similar performances across various runs.

The current research can be extended by using other explainability models such as saliency maps, feature visualization, and Grad-CAM. In future, the model can be used for other diseases with X-ray as an imaging technique. The same model may also be trained and tested with other imaging techniques, such as CT scans and MRIs; it would be interesting to see the performance of the proposed model in other domain datasets encompassing a variety of medical conditions and non-medical settings also.

Author Contributions

All the authors contributed equally.

Data Availability

The dataset for the current study is available at www.kaggle.com/nih-chest-xrays/data.

Declarations

Conflict of Interest

The authors declare no competing interests.