Ethical approval

The dataset employed in this study comprises anonymized human CXR images that are publicly available on Kaggle (https://www.kaggle.com/datasets/fernando2rad/x-ray-lung-diseases-images-9-classes). As all personal identifiers were removed and the data were fully de-identified, the university research ethics committee waived the requirement for Institutional Review Board (IRB) approval. This study does not involve direct interaction with human participants or access to identifiable private information, thereby adhering to established ethical standards for human subjects research.

Dataset description

This study employs a widely used CXR lung disease dataset comprising 9 classes [23] and 6,743 X-ray images. The dataset includes Normal (1,340 images), Pneumonia (1,060 images), Higher Density (678 images), Lower Density (629 images), Obstructive Pulmonary Diseases (644 images), Degenerative Infectious Diseases (594 images), Encapsulated Lesions (658 images), Mediastinal Changes (596 images) and Chest Changes (544 images). This distribution is diverse, but the classes were not balanced. Therefore, we balanced the dataset, allowing the model to learn subtle variations across a broad spectrum of pulmonary conditions. The images were collected from a private imaging center (GRS Imagem) in Brazil. The Higher Density class includes radiological conditions such as pleural effusion, atelectatic consolidation, hydrothorax and empyema. The Lower Density class corresponds to conditions involving abnormal air accumulation, including pneumothorax, pneumomediastinum and pneumoperitoneum. Obstructive Pulmonary Diseases comprise emphysema, bronchopneumonia, bronchiectasis and pulmonary embolism. Degenerative Infectious Diseases include chronic progressive conditions such as tuberculosis, sarcoidosis, proteinosis and pulmonary fibrosis. Encapsulated Lesions cover well-defined localized abnormalities, including abscesses, nodules, cysts, tumor masses and metastases. Mediastinal Changes represent abnormalities such as pericarditis, arteriovenous malformations and lymph node enlargement. Chest Changes include structural abnormalities caused by conditions such as atelectasis, congenital malformations, agenesis and hypoplasia. It is important to note that the utilized dataset does not provide comprehensive patient-level information, including demographic and clinical details such as age, sex, ethnicity or medical background. In addition, the absence of patient identifiers restricts the ability to verify patient-wise independence during data partitioning. These limitations constrain the assessment of dataset diversity, representativeness and potential bias and should be considered when interpreting the generalizability of the proposed model. However, because the primary objective of this study is to evaluate the multiclass classification performance of the proposed model across diverse radiological categories, the dataset remains appropriate and technically suitable for benchmarking classification accuracy and comparative performance.

Data augmentation was applied to address class imbalance and improve model generalization while strictly preserving the diagnostic integrity of CXR images. All augmentation parameters were carefully selected to reflect realistic variations encountered during image acquisition and patient positioning, thereby ensuring medical plausibility. The augmentations included rotations (−5° to +5°) to account for minor patient misalignment during imaging, zoom-in scaling (1.05× to 1.15×) to simulate variations in source-to-detector distance and horizontal translations up to 5% of the image size to reflect slight shifts in patient centering. Horizontal flipping was selectively employed, as left-right anatomical symmetry in CXRs is commonly used in the literature and does not compromise diagnostic validity for the studied conditions. Vertical flipping was intentionally excluded because it would invert anatomical orientation and compromise diagnostic relevance. All transformations were performed using bilinear interpolation to maintain image quality and avoid artificial artifacts. These augmentation strategies enhance model robustness by enabling the network to learn stable and invariant features under minor variations in image orientation, patient positioning, scaling and centering that commonly occur during real-world CXR acquisition.

After augmentation, 1,000 images were randomly selected per class, ensuring that 50% (500 images) were original and the remaining 50% were augmented versions derived exclusively from the training pool. This resulted in a balanced dataset of 9,000 images (1,000 per class). For hold-out validation, 80% (7,200 images) were used for training, 10% (900 images) for validation, and 10% (900 images) for testing. We ensured that no test images or their augmented variants were present in the training or validation sets.

For five-fold cross-validation, the complete set of 9,000 images was stratified into five groups of 1,800 images each, preserving class distributions across folds, which is critical for reliable multiclass medical image evaluation. The hold-out dataset was utilized for ablation studies, hyperparameter optimization, class-wise performance analysis, computational complexity evaluation, comparison with recent lung image classification methods and interpretability analyses. In contrast, the cross-validation dataset was employed exclusively for 5-fold performance estimation and statistical significance testing. Sample images from the dataset are illustrated in Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Sample CXR image of nine classes used in this study.

https://doi.org/10.1371/journal.pone.0351762.g002

Reproducibility

The hold-out validation experiments were conducted using a fixed random seed (seed = 42) for Python, NumPy, and TensorFlow to ensure reproducibility across all models, including the proposed LXNet. The cross validation experiments were conducted using three random seeds (42, 123 and 999). Experiments were implemented in Python 3.10 using TensorFlow/Keras 2.x on a cloud-based environment (Kaggle) utilizing dual NVIDIA Tesla T4 GPUs.

Pre-processing

The preprocessing pipeline was designed to standardize and enhance the quality of input CXR images before model training. All images were first resized to a fixed dimension to ensure consistent input resolution, followed by contrast enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE) to emphasize relevant local structures while minimizing noise amplification and artifact generation. Finally, pixel intensities were normalized to a standard scale to stabilize optimization and improve training efficiency. Together, these preprocessing steps reduce variability caused by differences in image acquisition, illumination and contrast, thereby improving the robustness and consistency of feature extraction across heterogeneous CXR images.

Resizing

The CNN models require fixed-size inputs. Since the original CXR images varied in resolution, all images were resized to 224 × 224 pixels to match the input size expected by the pretrained networks. Resizing was performed using bilinear interpolation, where the intensity of each new pixel I(x, y) is computed as a weighted average of the four nearest pixels in the original image, as expressed in Eq. 1:

(1)

where are the interpolation weights determined by the relative distances between the new pixel and its neighboring pixels. This approach preserves image smoothness and minimizes distortion.

Contrast Enhancement using CLAHE

CLAHE enhances local contrast by dividing an image into small contextual regions (tiles) and independently applying histogram equalization to each. The clip limit parameter constrains the amplification of noise, while the tile grid size determines the number of regions processed. In this study, CLAHE was applied to improve the visibility of subtle features in CXRs.

Unlike conventional CLAHE implementations that use fixed parameter settings, our method employs a grid search optimization to systematically explore combinations of clip limit and tile grid size. For each image, the configuration yielding the highest local contrast enhancement was automatically selected, ensuring improved feature visibility without introducing excessive noise. The CLAHE transformation is mathematically expressed in Eq. 2:

(2)

where L denotes the number of gray levels, and CDF(r) represents the clipped and redistributed cumulative distribution function.

In Fig 3, shows the impact of CLAHE on the CXR images. The grid search-based CLAHE produced a noticeable improvement in local contrast, particularly in regions exhibiting subtle diagnostic details.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Adaptive CLAHE based enhancement of CXR images.

https://doi.org/10.1371/journal.pone.0351762.g003

Normalization

CNNs typically perform better when input intensities are scaled to a consistent range. The original pixel values were normalized to the interval [0, 1] by dividing each pixel value p by 255, shown in Eq. 3:

(3)

This normalization ensures numerical stability during training and accelerates model convergence.

Architecture of the Proposed LXNet

Fig 4 illustrates the overall LXNet architecture and its major components: (1) preprocessing and data ingestion, (2) the LXNet feature-extraction backbone and (3) the classification head. In the left-most block, the raw CXR image () is fed into an adaptive preprocessing stage composed of resizing, CLAHE-based contrast enhancement (with parameters selected via grid search per image) and intensity normalization. Then, the preprocessed images are fed into the LXNet backbone, which begins with a 7 × 7 stem convolution (32 filters) to capture low-level edges and textures. This is followed by three hierarchical convolutional blocks:

• Conv Block 1: Two 3 × 3 convolution layers (48 filters), each followed by Batch Normalization and Swish activation, then Max Pooling and SpatialDropout2D to reduce co-adaptation.
• Conv Block 2: Two 3 × 3 convolution layers (72 filters), with the same BN + Swish + Max Pool + Spatial Dropout pattern to capture mid-level shapes and contours.
• Conv Block 3: Two 3 × 3 convolution layers (128 filters) that prioritize high-level semantic features (no pooling or dropout here to retain spatial detail needed to discriminate subtle pathologies).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. LXNet pipeline and lightweight architecture for nine‑class chest X‑ray classification.

https://doi.org/10.1371/journal.pone.0351762.g004

After these blocks, a Global Average Pooling (GAP) layer compresses feature maps into a compact 128-dimensional vector, followed by a Dropout layer (rate = 0.3) and a final Dense Softmax classifier that outputs probabilities across nine disease classes. Fig 4 also highlights the small model footprint (∼0.35M trainable parameters) and the evaluation branch, where LXNet is benchmarked against DenseNet201, ResNet50V2 and InceptionV3 using accuracy, precision, recall, F1-score, confusion matrix, training time, 5-fold cross-validation and statistical significance testing. Finally, the XAI analysis (Grad-CAM, Score-CAM, LIME) produces heatmaps and superpixel-based visual explanations for qualitative analysis and interpretation.

Stem convolution layer

The feature extraction process begins with a 7 × 7 convolutional stem layer with 32 filters, responsible for capturing fundamental edge- and texture-level features as expressed in Eq. 4:

(4)

where , and denotes the Swish activation function, defined in Eq. 5:

(5)

Here, f(x) represents the Swish activation, x is the input value, and is the sigmoid function. The Swish activation provides a smooth, non-monotonic transformation that enhances gradient flow and improves feature representation during training.

Convolutional Block 1 (48 filters)

The first block refines low-level feature maps through two consecutive 3 × 3 convolutions with Batch Normalization and Swish activations as expressed in Eq. 6:

(6)

where X⁽ⁱ⁾ is the output feature map of the i-th convolutional layer in the block, W⁽ⁱ⁾ and b⁽ⁱ⁾ are the convolution kernel and bias, is the input to the i-th convolution, * denotes the convolution operation, represents batch normalization, and is the Swish activation function defined in Eq. 5.

A max pooling operation reduces spatial dimensions while preserving dominant activations as expressed in Eq. 7:

(7)

where is the pooled output value at spatial position (p, q), represents the input feature map before pooling, and denotes the set of spatial indices (e.g., a 2 × 2 window) associated with each pooled output location.

Subsequently, SpatialDropout2D is applied to deactivate entire feature maps, mitigating co-adaptation as expressed in Eq. 8:

(8)

where is the resulting feature map after applying spatial dropout, denotes the pooled feature map, ⊙ indicates element-wise multiplication, and Z is a binary mask generated independently for each channel by a Bernoulli distribution with keep probability 0.7.

Convolutional Block 2 (72 filters)

The second block expands feature diversity and captures mid-level anatomical structures. Its formulation mirrors Block 1, given in Eq. 9:

(9)

where all symbols are as defined in Eq. 6, with an increased number of filters (72) to learn higher-level features in the second convolutional block.

Convolutional Block 3 (128 filters)

The final convolutional block is designed for high-level semantic abstraction, enabling the network to capture disease-specific radiographic signatures (e.g., opacities, lesions, density variations), as given in Eq. 10:

(10)

where all symbols are as defined in Eq. 6, with an increased number of filters (128). This block does not include pooling or dropout to preserve fine spatial details in the feature maps.

Global Average Pooling (GAP)

To reduce dimensionality while retaining class-discriminative features, a Global Average Pooling operation is performed as expressed in Eq. 11:

(11)

where denotes the global average pooled feature for channel k, X₄(i, j, k) is the value at spatial location (i, j) in channel k of the final feature map, H and W are the height and width dimensions, and g is the resulting feature vector of dimension 128.

Dropout Layer (rate = 0.3)

Dropout is applied to the pooled vector to further enhance generalization and reduce overfitting, expressed in Eq. 12:

(12)

where represents the feature vector after applying dropout, g is the globally pooled feature vector from Equation (9), ⊙ indicates element-wise multiplication, and Z is a binary mask vector with each element sampled from a Bernoulli distribution with probability 0.7.

Softmax Classifier

Finally, a fully connected dense layer with softmax activation produces probabilistic outputs for the nine disease categories, expressed in Eq. 13:

(13)

where denotes the predicted probability for class c, and are the weights and bias corresponding to class c in the final fully connected (dense) layer, is the feature vector after dropout, and the denominator sums the exponentiated scores over all nine classes to ensure the outputs form a valid probability distribution.

Baseline architectures for comparison

To evaluate LXNet fairly, we benchmarked it against three widely used CNN architectures. ResNet50V2 [33] represents modern residual networks that are commonly applied and easily fine-tuned for medical imaging tasks. ResNet50V2 is the updated version of the ResNet network family that performs better than ResNet50 and ResNet101 on the ImageNet dataset. DenseNet201 [21] serves as a high-capacity model that aggressively reuses features, often achieving strong accuracy at the cost of increased computational demand. InceptionV3 [34] captures multi-scale patterns efficiently through parallel convolutions, offering a balance between performance and computational efficiency. These baselines collectively span residual, densely connected and multi-scale design paradigms, enabling a comprehensive assessment of LXNet in terms of accuracy, generalization and efficiency under identical training conditions.

ResNet50V2

ResNet50V2 is a 50-layer deep CNN that leverages residual learning to enable the training of very deep architectures without suffering from vanishing gradients. The key component is the residual block, which introduces a shortcut (identity) connection allowing the input of a block to bypass one or more convolutional layers and be directly added to the block’s output:

(14)

where is the input, represents the residual mapping (convolution, batch normalization, and activation), are the weights, and is the output. ResNet50V2 differs from the original ResNet by using pre-activation, applying batch normalization and ReLU before convolution, which improves gradient flow and training stability. The network concludes with global average pooling and a softmax classifier, providing strong accuracy while maintaining efficient training.

DenseNet201

DenseNet201 is a 201-layer densely connected CNN. Unlike traditional CNNs, each layer receives inputs from all preceding layers, promoting feature reuse and mitigating the vanishing gradient problem [35]. Formally, the input to the l^th layer is:

(15)

where denotes the concatenation of feature maps from all previous layers, and is a composite function of batch normalization, ReLU, and convolution. This dense connectivity reduces the number of parameters, improves information flow, and enables accurate and efficient learning for complex image classification tasks.

InceptionV3

InceptionV3 is designed to efficiently capture spatial features at multiple scales using parallel convolutions of different kernel sizes within the same module. The core Inception module applies 1 × 1, 3 × 3, and 5 × 5 convolutions in parallel, then concatenates their outputs:

(16)

where y is the output of the module, , , and denote the outputs of the respective convolution branches, and concatenates them along the channel dimension. This multi-scale feature extraction enables the network to learn rich representations while keeping computational costs manageable.

Model training and optimal hyperparameters selection

LXNet was trained entirely from scratch using random weight initialization (Glorot Uniform), enabling the model to learn task-specific representations directly from the CXR images. In contrast, all baseline models (DenseNet201, ResNet50V2, and InceptionNet V3) were initialized with ImageNet pre-trained weights and fine-tuned using a transfer learning strategy, in which the convolutional backbones were retained and the original classification heads were replaced with task-specific heads corresponding to the target classes. To ensure a fair and consistent comparison, the baseline models were optimized within the same hyperparameter search space as LXNet, including optimizer type, learning rate, batch size, activation function, dropout rate and classification head configuration.

Table 1 summarizes the optimization spaces explored to identify the best-performing configurations for LXNet and the baseline models. For LXNet, systematic ablation and hyperparameter optimization experiments were conducted on the hold-out dataset, in which individual architectural components and training parameters, such as classifier head design, optimizer selection, learning rate, activation function, batch size and number of epochs, were independently varied while keeping other factors fixed. This ablation-driven optimization process enabled the identification of the most effective architectural and training configurations, ensuring that the final LXNet model achieved strong generalization performance while remaining computationally efficient. Subsequently, the best-performing configurations of all models were compared and the corresponding quantitative results are reported in the Results section.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Optimization and Hyperparameter Search Space for LXNet and Baseline Models.

https://doi.org/10.1371/journal.pone.0351762.t001

5-fold cross-validation

To ensure robust and unbiased performance evaluation, a 5-fold cross-validation strategy was employed using the cross-validation dataset, in which the images were partitioned into five mutually exclusive folds. Each fold was used once as a validation set, while the remaining four folds served as the training set. After training, performance metrics, including accuracy and macro F1-score, were computed for each fold and recorded. The fold-wise results were then aggregated and analyzed to compare the generalization performance of different models statistically.

The choice of a 5-fold configuration represents a practical balance between computational efficiency and statistical reliability. Using fewer folds (e.g., 3) could lead to higher variance in the results, while using more folds (e.g., 10) would significantly increase computational cost without a proportional gain in performance stability. The 5-fold approach thus provides a widely accepted trade-off suitable for deep learning tasks involving large image datasets. A stratified cross-validation strategy was employed to maintain proportional representation of all nine classes within each fold. This is particularly important for medical image datasets, which often exhibit varying class frequencies. Stratification ensured that the model was exposed to a representative sample of each class during both training and validation, preventing bias toward more prevalent disease categories and improving the reliability of performance estimates.

Ablation study and hyperparameter optimization results

A series of ablation experiments was performed to evaluate the effect of key hyperparameters on LXNet’s performance. Each parameter was varied independently, while the others were kept at their baseline settings. Model performance was assessed using validation and test accuracy and the summarized outcomes are presented in Tables 2–7.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. The impact of different classification head architectures on LXNet accuracy.

https://doi.org/10.1371/journal.pone.0351762.t002

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. The effect of optimizers on accuracy for training the LXNet.

https://doi.org/10.1371/journal.pone.0351762.t003

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. The effect of learning rates on accuracy.

https://doi.org/10.1371/journal.pone.0351762.t004

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Impact of the activation function on accuracy.

https://doi.org/10.1371/journal.pone.0351762.t005

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. The influence of batch size on accuracy.

https://doi.org/10.1371/journal.pone.0351762.t006

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Impact of training epochs on convergence and accuracy.

https://doi.org/10.1371/journal.pone.0351762.t007

Classifier head

Different classifier head configurations were examined to identify the most effective strategy for aggregating spatial features before the final dense layer. As shown in Table 2, Flatten and DualPooling resulted in lower accuracy, while GlobalMaxPooling2D provided modest improvement. The highest performance was achieved with GlobalAveragePooling2D, achieving test accuracy of 0.9944 and validation accuracy of 0.9922. This confirms that global average pooling retains the most discriminative spatial information and is optimal for LXNet’s classifier head.

Optimizer

The optimizer significantly influences convergence speed and training stability. LXNet was evaluated using SGD, RMSprop, Adamax, Nadam and Adam. As summarized in Table 3, traditional SGD underperformed (test accuracy < 0.90), while adaptive optimizers achieved superior results. Both Nadam and Adam achieved the highest test accuracy (0.9944), with Nadam selected for its smoother convergence and more stable validation performance.

Learning rate

The learning rate determines the balance between training stability and convergence efficiency. As shown in Table 4, a small rate (0.0001) led to underfitting, while larger rates () improved initial learning but caused instability. A learning rate of 0.0003 provided the best trade-off between convergence speed and final accuracy and was thus adopted as the optimal setting.

Activation function

Activation functions play a vital role in learning complex, non-linear patterns. Table 5 compares the performance of ReLU, LeakyReLU, ELU, and Swish. The results indicate that Swish consistently outperformed the others, achieving the highest test (0.9878) and validation (0.9900) accuracies. The smooth and non-monotonic nature of Swish enables better feature representation, making it the preferred activation for LXNet.

Batch size

Batch size influences both computational efficiency and model generalization. As shown in Table 6, smaller batches (16 and 32) yielded stable but slightly lower accuracy, while larger batches (64) reduced generalization. A batch size of 48 achieved the best performance, with a test accuracy of 0.9911 and strong validation stability, providing the most balanced configuration.

Number of epochs

The number of epochs controls the extent of model convergence. As reported in Table 7, test accuracy increased steadily with training duration from 0.88 at 10 epochs to nearly 0.99 at 30 epochs. Optimal convergence was observed at 40 epochs, achieving a test accuracy of 0.9922 and validation accuracy of 0.9967. Beyond 40 epochs, improvements were marginal, indicating adequate saturation. Overall, the ablation findings establish the following optimal configuration for LXNet: GlobalAveragePooling2D classifier head, Nadam optimizer, learning rate of 0.0003, Swish activation, batch size of 48, and 40 training epochs. This combination yielded superior accuracy, robust convergence and efficient computational performance.

Based on the ablation studies presented in Tables 2–7, the optimal hyperparameter configuration for LXNet was determined. Table 8 summarizes the optimal hyperparameters identified through the ablation studies. The final configuration for LXNet includes the GlobalAveragePooling2D classifier head (test accuracy = 0.9944, validation accuracy = 0.9922), Nadam optimizer (selected for its stable convergence and high accuracy), learning rate of 0.0003 (balancing performance and stability), Swish activation function (outperforming ReLU, ELU and LeakyReLU), batch size of 48 and 40 training epochs (ensuring convergence with negligible gains beyond this point). This combination achieved the most balanced trade-off between accuracy, training stability and computational efficiency and was adopted as the final configuration for all subsequent experiments and model comparisons.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Optimal hyperparameters for LXNet identified via ablation studies.

https://doi.org/10.1371/journal.pone.0351762.t008

Comparison with baseline models under hold-out evaluation

Tables 9–12 shows the comparative performance of LXNet against the baseline models, along with the class-wise performances. The micro, macro, and weighted average performances of the models were included in the tables. Micro, macro, and weighted-average accuracies differ in both their calculations and interpretations. Micro-average accuracy is calculated by aggregating correct predictions over all samples, so each instance contributes equally and majority classes dominate the result. Macro-average accuracy is computed by first calculating accuracy for each class separately and then taking their unweighted mean, giving equal importance to all classes. Weighted-average accuracy is obtained by averaging class-wise accuracies weighted by the number of samples in each class. For imbalanced medical datasets, macro-average accuracy is more reliable for fair class-wise evaluation, while weighted-average accuracy reflects overall practical performance. Since we used a balanced dataset for model evaluation, we mainly reported accuracy using the micro average in this paper.

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Class-wise performance of LXNet on the hold-out test set.

https://doi.org/10.1371/journal.pone.0351762.t009

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Class-wise performance of ResNet50V2 on the hold-out test set.

https://doi.org/10.1371/journal.pone.0351762.t010

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Class-wise performance of DenseNet201 on the hold-out test set.

https://doi.org/10.1371/journal.pone.0351762.t011

Download:

• PNG
  larger image
• TIFF
  original image

Table 12. Class-wise performance of InceptionNet V3 on the hold-out test set.

https://doi.org/10.1371/journal.pone.0351762.t012

Among the baselines, InceptionNetV3 demonstrated the lowest micro accuracy of 0.93, indicating limited generalization capability. DenseNet201 achieved moderately higher, more consistent performance (0.956 across all metrics) but required longer training due to its dense feature propagation. ResNet50V2 performed slightly better, reaching 0.974 with improved convergence stability. All the models are trained under identical settings to ensure fair comparison. LXNet achieved the highest micro accuracy of 0.982. This consistent improvement of 1–9% over the pretrained baselines demonstrates the effectiveness of LXNet’s lightweight design in capturing discriminative features while maintaining computational efficiency. The results confirm that LXNet delivers robust, efficient performance suitable for practical applications where accuracy and speed are both critical. Fig 5 shows the training and validation accuracy and loss curves for the LXNet.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Training and validation curves for LXNet under the final configuration.

https://doi.org/10.1371/journal.pone.0351762.g005

Class-wise performance of the models on the hold-out test

Table 9 shows the per-class accuracy, precision, recall and F1-score of LXNet across the nine classes. The model demonstrates outstanding performance, achieving perfect precision (1.00) in Higher Density, Degenerative Infection Diseases, Mediastinal Changes, and Chest Changes. Obstructive Pulmonary Diseases also reached an accuracy of 0.99, with a minor reduction in precision (0.98), indicating minimal misclassifications. The Normal class attained slightly lower values (accuracy = 0.98, precision = 0.94, recall = 0.95, F1 = 0.94), reflecting minor confusion with other classes that exhibit subtle radiographic patterns. Pneumonia had an accuracy of 0.98, a recall of 0.93, yielding an F1-score of 0.93, suggesting that a few positive cases were misclassified, likely due to overlapping radiographic features with other pulmonary conditions. Degenerative Infectious Diseases achieved an F1-score of 0.99, with perfect precision but slightly lower recall (0.99), indicating a small number of false-positive predictions. Overall, LXNet achieved macro and weighted-average accuracies of 0.99, highlighting consistent and balanced performance across all classes. These results confirm that LXNet can reliably discriminate subtle and challenging thoracic patterns, demonstrating robustness and minimal misclassification even in the most difficult categories.

The comparison of class-wise performances from Table 9, 10, 11 and 12 shows that LXNet consistently achieves the highest or near-highest scores across all classes. For the Normal class, LXNet attains an F1-score of 0.945, outperforming ResNet50V2 (0.906), DenseNet201 (0.900), and InceptionNet V3 (0.870). A similar trend is observed for Pneumonia, where LXNet achieves an F1-score of 0.939, compared to 0.907 (ResNet50V2), 0.863 (DenseNet201), and 0.762 (InceptionNet V3), indicating a clear margin of improvement. For structurally distinct classes such as Higher density, Mediastinal changes, and Chest changes, both LXNet and ResNet50V2 achieve perfect F1-scores of 1.000, while DenseNet201 and InceptionNetV3 show slightly lower or comparable results. In intermediate-complexity classes, including Lower density and Obstructive pulmonary diseases, LXNet maintains F1-scores above 0.98, marginally exceeding ResNet50V2 and clearly outperforming DenseNet201 and InceptionNetV3. Overall, LXNet demonstrates the most balanced and robust class-wise performance, particularly for difficult and relevant categories, as reflected in its highest micro accuracy (0.982) among all evaluated models.

Fig 6 and 7 show the Receiver operating characteristic (ROC) curves and confusion matrices of the models for the test set, providing a detailed view of class-wise prediction performance. LXNet demonstrated the most balanced performance across all nine classes. For the Normal class, LXNet correctly classified 95 samples, with only minor confusion into Pneumonia (1) and Lower Density (2), whereas ResNet50V2, DenseNet201, and InceptionNet-V3 misclassified 4, 10 and 13 cases, respectively, indicating progressively weaker discrimination of healthy tissue. A similar pattern is observed for Pneumonia, where LXNet achieved 93 correct predictions with limited confusion, compared to higher misclassification in ResNet50V2 (16 cases) and substantially greater dispersion in DenseNet201 and InceptionNet V3, the latter misclassifying 31 samples across multiple categories. For Obstructive Pulmonary Diseases and Degenerative Infectious Diseases, LXNet maintained strong robustness with 98 and 99 correct predictions, respectively. At the same time, DenseNet201 and InceptionNet V3 showed higher cross-class confusion, reflecting reduced stability in differentiating overlapping pathological patterns.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. ROC curve of the Models.

https://doi.org/10.1371/journal.pone.0351762.g006

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Confusion matrix of the Models for test data(N = Normal, P = Pneumonia, HiD = Higher Density, LoD = Lower Density, Ob = Obstructive Pulmonary, Deg = Degenerative Infectious Diseases, En = Encapsulated Lesions, Med = Mediastinal Changes, Ch = Chest Changes).

https://doi.org/10.1371/journal.pone.0351762.g007

While ResNet50V2 achieved comparable performance to LXNet, particularly for structurally distinct classes, it displayed higher confusion in Normal–Pneumonia and Pneumonia–Degenerative pairs compared to LXNet. DenseNet201 and InceptionNet V3 showed the highest inter-class confusion, especially among Normal, Pneumonia, Obstructive and Degenerative classes, indicating reduced robustness in ambiguous cases. Consistent with the confusion matrix analysis, LXNet achieved superior class-wise consistency without sacrificing performance across any category, as further supported by its uniformly high AUC values across all classes (over 0.99). Collectively, these results demonstrate that LXNet not only matches or exceeds baseline models in overall accuracy but also provides more reliable and interpretable class-level predictions, particularly in challenging scenarios.

5-Fold cross-validation analysis

We performed 5-fold cross-validation experiments three times independently, using different randomly generated seeds (42, 123 and 999). This provides a more rigorous assessment of robustness and reproducibility. As shown in Table 13, LXNet achieved consistently strong performance across all folds, with fold-wise accuracies ranging from 0.952 to 0.973. The overall average accuracy was 0.961 ± 0.007, demonstrating both high predictive performance and low variability across folds and seeds. The relatively small standard deviations across folds further indicate that LXNet is stable with respect to random initialization and data shuffling, confirming its reliable generalization without signs of overfitting. In comparison, the baseline pretrained models exhibited either slightly lower mean performance or greater instability. ResNet50V2 achieved a comparable average accuracy of 0.960 ± 0.007; however, its fold-wise fluctuations suggest marginally less consistent behavior across different splits and seeds. DenseNet201 obtained a lower average accuracy of 0.903 ± 0.004, while InceptionNetV3 achieved 0.889 ± 0.004, both indicating substantially weaker performance. Moreover, DenseNet201 showed relatively larger per-fold standard deviations (up to 0.034), highlighting increased sensitivity to data partitioning and random initialization. Overall, the repeated cross-validation protocol (multiple folds combined with multiple seeds) strengthens the reliability of the findings. The results confirm that LXNet maintains competitive accuracy with improved stability across diverse training conditions, supporting its robustness and suitability for practical deployment.

Download:

• PNG
  larger image
• TIFF
  original image

Table 13. 5-Fold cross-validation accuracy of models over three random seeds (Avg = average; Std = standard deviation).

https://doi.org/10.1371/journal.pone.0351762.t013

Further, a paired t-test was conducted on the repeated 5-fold cross-validation results (5 folds × 3 seeds). The analysis showed that LXNet significantly outperformed DenseNet201 and InceptionNetV3 (p < 0.05), while no statistically significant difference was observed between LXNet and ResNet50V2 (p > 0.05), indicating comparable performance between these two models.

Robustness analysis

To evaluate model robustness under common image degradation conditions in CXRs, we performed a systematic analysis considering Gaussian blur, Gaussian noise, Gamma correction (non-linear transformation) and linear intensity adjustments. Gaussian blur images were generated with standard deviations from 0 to 2.0, Gaussian noise with standard deviations from 0 to 0.10, Gamma values ranged from 0.50 to 1.50 and linear intensity shifts were simulated by adding or subtracting fixed values to pixel intensities. The analysis revealed that LXNet maintained high performance under moderate Gaussian blur, while ResNet50V2, DenseNet201 and InceptionV3 showed sharper degradation. Under Gaussian noise, ResNet50V2 was most stable, whereas LXNet was more sensitive at higher noise levels. LXNet showed reduced performance at very low Gamma values but recovered near a Gamma of 1.0, while ResNet50V2 and DenseNet201 remained stable across Gamma variations. All models tolerated linear intensity shifts reasonably well, with LXNet and ResNet50V2 performing slightly better. Within moderate degradation ranges, LXNet’s performance remained stable, highlighted as the optimal performance zone in Fig 8.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Performance of the models for digitally simulated test sets.

https://doi.org/10.1371/journal.pone.0351762.g008

We further evaluated LXNet on an external dataset collected from multiple sources and laboratories to assess domain shift, as shown in Table 14. The model achieved a micro-averaged accuracy of 0.706, performing well on normal, pneumonia, high-density, mediastinal and chest classes, but showing reduced accuracy for less common or complex classes, such as obstructive pulmonary diseases and degenerative infectious diseases. These results indicate that while LXNet is robust under moderate degradation and generalizes well to common patterns, its performance decreases for rarer or more complex pathological variations.

Download:

• PNG
  larger image
• TIFF
  original image

Table 14. Class-wise performance of LXNet on external dataset.

https://doi.org/10.1371/journal.pone.0351762.t014

Statistical significance

To evaluate whether the observed performance differences between LXNet and the baseline models (ResNet50V2, DenseNet201 and InceptionNetV3) were statistically meaningful, a non-parametric analysis was conducted using the results of repeated 5-fold cross-validation across three random seeds (42, 123 and 999). Parametric tests such as the paired t-test assume independence and normality of samples; however, these assumptions are violated in k-fold cross-validation due to overlapping training sets and shared data distributions across folds. Therefore, we employed the Wilcoxon signed-rank test, which is suitable for paired, small-sample comparisons and does not rely on distributional assumptions. For each baseline model, fold-wise accuracies were paired with those of LXNet and the accuracy difference for each fold was computed as

(17)

where denotes the classification accuracy on the i-th fold. Zero differences were excluded and the remaining absolute differences were ranked in ascending order. Each rank was assigned the sign of the corresponding difference. The Wilcoxon signed-rank test statistic was then computed as:

(18)

and evaluated using a two-sided hypothesis test with the null hypothesis

(19)

The Wilcoxon signed-rank test revealed clear performance patterns between LXNet and the baseline models. Compared with ResNet50V2, LXNet achieved a negligible mean accuracy gain of +0.001; this difference was not statistically significant (p = 0.3125), indicating that LXNet and ResNet50V2 perform comparably. In contrast, LXNet significantly outperformed DenseNet201, with a mean accuracy gain of +0.058 (p = 0.03125) and InceptionNetV3, with a larger mean accuracy gain of +0.072 (p = 0.03125). These results confirm that LXNet consistently provides a meaningful advantage over DenseNet201 and InceptionNetV3, while achieving parity with ResNet50V2. To further strengthen robustness, performance was averaged across the three random seeds in addition to the 5 folds, yielding an overall mean ± standard deviation of 0.961 ± 0.007 for LXNet, compared to 0.960 ± 0.007 for ResNet50V2, 0.903 ± 0.004 for DenseNet201 and 0.889 ± 0.004 for InceptionV3. These results highlight LXNet’s low variability across both folds and seeds. Moreover, LXNet maintained consistently high class-wise accuracy across all nine classes, whereas baseline models showed greater inter-class confusion. Collectively, these findings demonstrate that LXNet provides more stable, reproducible and superior generalization performance under repeated cross-validation conditions.

Comparison with existing methods

Table 15 shows the comparison of LXNet with recent lung disease classification models. The method proposed by Aldamani et al. [31] used EfficientNetB2 and achieved 98.50% accuracy. The Versatile ML Architecture System of Rodriguez-Carpio [41] achieved 97.20% accuracy and Gutta’s RespiraAI [42] achieved 95.00% accuracy. LXNet outperformed these recently methods, reaching 99.62% (macro average accuracy in hold out test), while maintaining consistently high per-class F1-scores (0.97–1.00) and statistically significant improvements over baselines (p < 0.000001, large Cohen’s d). These results highlight the effectiveness of LXNet’s lightweight design, optimized hyperparameters and feature recalibration, demonstrating both high predictive accuracy and robust generalization suitable for practical use.

Download:

• PNG
  larger image
• TIFF
  original image

Table 15. Comparison of LXNet with recent CNN-based lung disease classification methods.

https://doi.org/10.1371/journal.pone.0351762.t015

Interpretability analysis

To enhance trust and support real-world adoption, we evaluated the interpretability of LXNet using three complementary XAI techniques: Grad-CAM [16], Score-CAM [43] and LIME [44] across all nine lung disease categories. Representative examples are shown in Fig 9.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Visual explanations of LXNet predictions using Grad-CAM, Score-CAM, and LIME.

https://doi.org/10.1371/journal.pone.0351762.g009

Grad-CAM highlights class-discriminative regions by weighting convolutional feature maps with the gradient of the predicted class score. In the heatmaps (second column of Fig 9, Grad-CAM consistently emphasized relevant thoracic regions, such as diffuse opacities in Pneumonia, localized Encapsulated Lesions and altered density patterns in Obstructive Pulmonary Disease. Score-CAM offers a gradient-free alternative that weights activation maps based on their contribution to the classification score. Compared to Grad-CAM, Score-CAM maps were smoother and less noisy, enabling clearer localization of affected regions and reducing sensitivity to gradient saturation, particularly useful for subtle radiographic abnormalities. LIME approximates local decision boundaries by perturbing superpixels. The third column in Fig 9 shows the super-pixels most influential to each prediction. While less spatially precise than CAM-based methods, LIME provided complementary insights by highlighting meaningful subregions, such as lobar clusters or mediastinal areas, that contributed to the model’s confidence.

To further illustrate disease-specific feature learning, we analyzed how LXNet’s attention maps align with established radiological patterns. For Pneumonia, Grad-CAM and Score-CAM highlighted patchy opacities, particularly in the lower lung zones, consistent with infectious infiltrates. In Encapsulated Lesions, activations were highly localized around well-defined nodular regions, indicating sensitivity to sharply bounded abnormalities. For Obstructive Pulmonary Disease, the model emphasized radiolucent regions and altered lung textures over broader areas, reflecting hyperinflation-related changes. In Normal CXRs, attention was diffusely distributed across clear lung fields, suggesting reliance on the absence of pathology. LIME-based superpixel analysis further supported these findings by identifying coherent anatomical regions, such as lobar and perihilar areas, contributing to predictions. Overall, these results indicate that LXNet captures disease-specific imaging features rather than spurious correlations.

Across all classes, LXNet consistently focused on meaningful pulmonary regions rather than spurious background artifacts, demonstrating reliable attention behavior. By integrating multiple XAI techniques, the analysis confirms that the model’s predictions are interpretable and grounded in relevant features. This not only validates the robustness of LXNet across diverse thoracic conditions but also supports its key contribution as an accurate and transparent AI tool for assisting radiologists in complex CXR interpretation. Future work may incorporate complementary explainability approaches, such as SHapley Additive exPlanations (SHAP) and Prediction Difference Analysis (PDA), to provide both global and local interpretability and further enhance reliability.

Time complexity analysis

Table 16 presents a detailed comparison of training times for LXNet and three baseline pretrained CNN models, including per-epoch and total training durations, along with their micro average test accuracy on the hold out set. LXNet exhibits remarkable computational efficiency, requiring only 32 seconds for the first epoch and an average of 7.7 seconds per subsequent epoch, resulting in a total training time of 308 seconds. Despite its lightweight design, LXNet achieved the highest test accuracy of 0.982 (micro average test accuracy). In comparison, ResNet50V2 and InceptionNet V3 required 721 and 668 seconds in total, achieving 0.974 and 0.930 micro average test accuracy, respectively, while DenseNet201 incurred the longest training time of 1320 seconds with a slightly lower accuracy of 0.956. These results demonstrate that LXNet delivers superior predictive performance while maintaining fast training, making it highly suitable for resource-constrained environments and rapid deployment.

Download:

• PNG
  larger image
• TIFF
  original image

Table 16. Comparative training duration and accuracy of LXNet versus baseline CNNs (Avg = average; Sec = seconds; Acc = accuracy).

https://doi.org/10.1371/journal.pone.0351762.t016