Multi-scale pyramid spatial atrous channel attention enhanced ResNet50V2 for explainable pulmonary fibrosis diagnosis

1 Introduction

Lung diseases are a major global health concern, with pulmonary fibrosis classified under lung tissue diseases. In 2020, over 3.5 million deaths were attributed to lung disorders like COVID-19, with 80% occurring in low- and middle-income countries (Shahbazi et al., 2025). Pulmonary fibrosis affects over 250,000 Americans, with 40,000 deaths and 50,000 new cases annually (Podolanczu et al., 2023). Idiopathic Pulmonary Fibrosis (IPF) is the most common and fatal form, with a median survival of 3–5 years. Current research includes trials of CC-90001, a JNK inhibitor, to slow disease progression (Popmihajlov et al., 2022). Diagnosis remains difficult due to symptom overlap, requiring high-resolution CT or biopsy.

The goal of the computer science field of Artificial Intelligence (AI) is to give computers like human intelligence so that it can motivate, gain knowledge, and overcome the difficulties (Kaul et al., 2020). Artificial intelligence algorithms, especially Machine Learning (ML) and Deep Learning (DL) (Mikołajczyk and Grochowski, 2018) have completely changed thoracic surgery techniques. These ML and DL methods simplify feature extraction (Hammad et al., 2024), selection, reduction, classification (Mahamud et al., 2024), and data preparation it enhancing the precision and effectiveness of medical image analysis. As a result, lung specialists now have much greater confidence in diagnosing fibrosis. DL techniques are widely used to detect pulmonary fibrosis, especially those that involve finetuning (Tajbakhsh et al., 2016) and transfer learning (Ahuja et al., 2021). Transfer learning is a deep learning technique where models pre-trained on large datasets are fine-tuned for specific tasks. CNN architectures, often trained on diverse labeled image datasets, are adapted to smaller medical image datasets to leverage their feature extraction capabilities. This approach reduces overfitting and speeds up training, making it practical even on standard laptops or personal computers. Numerous CNN models that have already been trained have been developed, including Visual Geometry Group (VGG) (Simonyan and Zisserman, 2014), Mobile Network (MobileNet) (Sandler et al., 2018), Residual Network (ResNet) (He et al., 2016), Dense Network (DenseNet) (Huang et al., 2017), and others.

Pulmonary fibrosis is a chronic progressive lung disease which is estimated to affect millions of people in the world and is characterized by inflammation and scarring of lung tissue with symptoms such as fatigue, coughing, and breathlessness. The delayed diagnosis is usually caused by the use of invasive or low sensitivity imaging modes (Vick et al., 2023). There has been no known cure and the present treatment is aimed at alleviating symptoms. The issue of early and correct diagnosis is one of the primary concerns, and this fact brings attention to the necessity of non-invasive, specific, and valid diagnostic method.

This study used eight pre-trained deep CNN models, VGG 19 (Khan et al., 2021), MobileNetV2 (Souid et al., 2021), ResNet-50 (Zakaria et al., 2021), ResNet 50V2, ResNet-152V2, DenseNet-169, DenseNet-121, and DenseNet-201 (Tekerek and Al-Rawe, 2023), to extract and identify pulmonary fibrosis features from CT images sourced from Kaggle. The models were changed structurally to fit binary classification (normal vs. pulmonary fibrosis), and a sigmoid activation function was used. The learning rate, layers, loss and epochs were hyperparameters that were optimized. ResNet50V2 was the most successful in terms of accuracy and loss and thus, it is beneficial in the rapid and accurate detection of pulmonary fibrosis. The objective of the study will be to enhance diagnostic accuracy and reliability particularly in persons who are aged 50 70, since they are the most affected by the disease.

ResNet50V2 is widely applied in disease classification tasks. It achieved 97% accuracy in kidney illness recognition (Gill et al., 2023), 79.63% accuracy in breast cancer detection using BreakHis dataset with data augmentation (Kaushik and Choudhary, 2024), and 92.93% accuracy in pulmonary disorder classification along with lightweight CNNs like MobileNetV2, EfficientNet-B0, and NASNetMobile (Khan et al., 2025). Even though ResNet50V2 is successful, it experiences the issue of false positives. This was overcome in pulmonary fibrosis detection by architectural improvements, namely, introducing batch normalization and dropout in block 5 and the significance of disease-specific model customization to achieve high-quality clinical results.

However, beyond false positive predictions, ResNet50V2 has inherent weaknesses in detecting pulmonary fibrosis-specific patterns in CT images. Pulmonary fibrosis is marked by slight, diffuse, and random texture changes, which demand multi-scale and feature representation. Conventional ResNet structures are mostly based on fixed receptive fields and hierarchical feature extraction, which might not satisfy fine-grained fibrosis textures or long-range contextual attributes. Also, the absence of explicit spatial attention mechanisms restricts the model to give attention to disease-relevant lung areas and inhibit irrelevant structures like ribs and background noise. Such limitations encourage the adoption of more sophisticated attention and multi-scale feature improvement modules to the proposed framework.

Batch Normalization (BN) stabilizes CNN activations by normalizing layer outputs, addressing gradient vanishing and poor weight initialization, and speeding up convergence. Although internal covariance bias caused by continuous updates of weights may exist, BN is used as preprocessing and optimization in CNNs (Li et al., 2018). Dropout is used in ResNet50V2 to boost generalization and reduce overfitting by randomly deactivating neurons during training. It is strategically placed after key convolutional and fully connected layers to regulate activation variance and enhance classification reliability (Achille and Soatto, 2018). The aim of this work is to enhance the accuracy and dependability of pulmonary fibrosis diagnostic techniques. The proposed pyramidal spatial atrous channel attention model is used for feature extraction to enable binary classification.

The Contribution of the research work is given below.

• ⁃ A novel pyramid spatial atrous channel attention (PSACA) architecture is proposed, where Spatial Attention is applied first, followed by ASPP and then Channel Attention for multi-scale fibrosis feature enhancement.
• ⁃ Integrate a lightweight ASPP module into CBAM to enhance the contextual learning of subtle PF patterns.
• ⁃ The PSACA is integrating the proposed block into Stage 3 and Stage 4 of ResNet50V2, strengthening both mid-level and high-level representations.
• ⁃ The PSACA serves as the principal feature extractor, enhancing edge definition, texture differentiation, and local contrast in CT images of pulmonary fibrosis. ResNet50V2 functions as the final classifier, efficiently utilizing the improved and attention-optimized features to achieve precise binary classification.
• ⁃ Explainable AI (XAI) using Grad-CAM++ is incorporated, presenting visual interpretation maps that emphasize clinically significant fibrosis regions, hence enhancing model transparency.

The subsequent sections of this research work are organized to present a deep learning model for detecting lung disease from CT images as follows: the Background work for this study is discussed in Section 2 and the proposed model is discussed in Section 3. Section 4 explains the experimental setup details are mentioned. The performance evaluation metrics are represented in Section 5. The simulation results analysis and discussions are presented in Section 6. A comparison analysis of the proposed model with the existing models are discussed in Section 7. Finally, this paper is concluded with a conclusion and reference.

2 Related work

This section explores research on methods of lung disease classification methods, highlighting various perspectives and approaches.

Detecting pulmonary diseases using simple and basic image processing methods can be time-consuming. However, with the significant advancements in GPUs and Convolutional Neural Networks (CNNs), the efficiency of lung disease diagnosis and decision support has greatly improved. The are different studies propose multiple deep learning techniques for the detection of pulmonary fibrosis and other different lung diseases.

Al-Sheikh et al. (2023) presented deep-learning models for image classification of lung disease detection. This classification was carried out using a modified CNN model, along with fine-tuned versions of the VGG16Net and AlexNet models and these models achieved an improved accuracy of 98.80% for CT scan images (Abdulahi et al., 2024). introduced the pulmoNet deep learning model which is designed to improve the diagnosis of lung disorders. pulmoNet was trained for large datasets of images from publicly available with enhanced performance and established robust capabilities of binary class and multi-class classification tasks (Mahapackialakshmi and Jaffino, 2024). described a CNN-driven SIFT (Scale-Invariant Feature Transform) method for identifying and categorizing the pneumonia through chest X-ray imagery. The findings highlight U-Net’s superior performance among CNN architectures, demonstrating less bias and enhanced generalization. Both U-Net and SegNet excel at leveraging limited yet intricate image datasets to achieve remarkable performance in pneumonia classification.

Bharati et al. (2020) described and emphasized the importance of identifying accurately and differentiate between the chronic obstructive pulmonary disease and lung cancer effectively to preserve the human life. The proposed hybrid RNN-based deep learning model gives better performance when the processed chest x-ray images are used especially when it comes to identify minute alterations in the respiratory system which includes the growth of tiny lobes or thicker capillaries (Patil et al., 2024). proposed hybrid approach that combines VGG19 and Capsule networks to improve the accuracy and dependability of lung cancer diagnosis from CT scans. Here, capsule Networks are used to enhance accuracy in lung cancer detection by effectively handling variations in image orientation, preserving spatial hierarchies, reducing false positives, and improving feature representation (Karla and Yalavarthi, 2024). described the recurrent neural networks to increase the classification accuracy of lung cancer and chronic obstructive disease by enhancing feature extraction. A hybrid system employing ResNet18 effectively classifies the X-ray pictures into normal, Chronic obstructive pulmonary disease, and lung cancer with increased accuracy (Gadikota et al., 2024). Examined the creation of a Computer-Aided Diagnosis (CAD) system utilizing convolutional neural networks for the early detection of lung cancer from CT scans. This system surpassed several state-of-the-art approaches in performance. Leveraging transfer learning, the CAD system enhances the CNN model trained on a substantial dataset of CT images to ensure precise and reliable cancer detection.

Nasra (2024) discussed the model of resnet50 as a feature extractor pre-trained model on ImageNet which is used to identify lung cancer and model performance which is greatly accurate by differentiating the malignant of lung tissues (Kanavati et al., 2021). developed a deep learning model for categorizing unspecified lung carcinoma in biopsy images to improve diagnosis and therapy. This is consistent with the growing trend of accurately detect the lung cancer with AI. For the analysis medical imaging, deep learning has promise for accurate and quick cancer diagnosis (Arun Prakash et al., 2023). described the real-time diagnostic capabilities demonstrated by the Stacked Deep Neural network model. The diagnosis of pediatric pneumonia is difficult because of chest X-rays have low radiation levels. A proposed computer-aided model that uses the deep learning-based characteristics for the improved image quality which achieves 98.62% accuracy (Yang et al., 2021). presented a deep learning model for diagnosing COVID-19 using CT and X-ray images. The VGG16 and ResNet models achieved high accuracy, with the classification results demonstrating that the enhanced CNN models consistently maintain an accuracy rate of 98%.

Sangeetha et al. (2024) proposed Multimodal Fusion Deep Neural Network (MFDN) combines the clinical data, genetics, and medical imaging to improve the accuracy of lung cancer diagnosis. With an accuracy of 92.5%, MFDN surpasses more established models like CNN, DNN, and ResNet. Deep CNN-based techniques for COVID-19 detection from chest CT scans are evaluated by Mishra et al. (2020). Performs several Deep CNN models that are combined in a decision fusion method which achieves an accuracy of above 86%. The results demonstrate the possibility of Deep CNNs for quick COVID-19 screening and epidemic containment (Vats et al., 2022). proposed a DenseNet system to use X-ray images and neural networks, such as CNNs to classify accurately. Using pre-trained models and a customized DenseNet architecture, the work outperformed of DenseNet model with an accuracy of 92.6% over 20 epochs with enhanced performance for early pneumonia detection.

Amyar et al. (2020) Presented an automatic classification tool for COVID-19 pneumonia screening using a CT scan. A suggested multitask deep learning model enhances the classification performance by leveraging shared feature representation. This model achieves the effective and correct identification COVID-19 instances by obtaining an area under the ROC curve which achieves 97% (Wang et al., 2017). introduced a Multi-scale Rotation-Invariant Convolutional Neural Network (MRCNN) which achieves an accuracy of 0.90. That same year (Kim et al., 2018), introduced a CNN architecture comprising four convolutional layers and two fully connected layers, achieving an improved performance with an accuracy of 0.9512.

Van der Velden et al (2022) examined fundamental XAI techniques such as saliency maps, Grad-CAM, and SHAP in the context of deep learning for medical image analysis, highlighting the necessity of validation for clinical trust. Building on this, van der Velden et al. (saliency overview) provided practitioner-oriented recommendations on saliency-based XAI methodologies to enhance interpretability in radiology processes (Barnes et al., (2023)). characterized machine learning as the “new frontier” in interstitial lung disease imaging, emphasizing AI’s capability for CT-based fibrosis categorization but underscoring transparency deficiencies that hinder its implementation (Colombi et al., (2025)). examined computer-aided techniques for ILD evaluation, illustrating textural biomarkers for diagnosis and prognosis while highlighting the limits of black-box models. The quantitative imaging review team elucidated AI-driven CT analysis for ILD patterns, underscoring XAI’s function in correlating predictions with pathophysiologic lung areas (Kaul et al., 2020).

More recent research has examined the use of state-of-the-art deep learning models that implement attention and feature optimization to analyze medical images (Xiao et al., 2025). have applied a modified version of the U-Net architecture that incorporated residual connections and dual attention mechanism in order to improve segmentation. Besides that (Fan et al., 2025), introduced a step gravitational search algorithm based on joint feature selection and parameter optimization, which can prove the significance of effective feature learning in deep models. The current developments underscore the increased focus on the multi-scale feature extraction, attention, and optimization, which is consistent with the design of the proposed PSACA-based framework.

From the literature review, classification, feature extraction, and disease detection techniques were analyzed, with many studies reporting enhanced performance metrics such as accuracy. These improvements highlight the advantages of the models used in the reviewed papers. However, some studies also reported a higher false positive rate. This metric was identified as a challenge in this work, and it has focused on reducing the false positive rate while simultaneously improving overall performance.

3 Materials and methods

This section explains the process of classifying CT Scan images into normal and pulmonary fibrosis. As illustrated in Figure 1a. Shows the general block diagram. This method refines eight models such as VGG 19, MobileNetV2, ResNet-50, ResNet-50V2, ResNet-152V2, DenseNet-169, DenseNet-121, and DenseNet-201 on CT scan image series from a CT scan pulmonary fibrosis image detection dataset for both feature extraction and the detection purposes using transfer learning with pre-trained ResNet and their modifications. These models were chosen over other state-of-the-art pre-trained DCNN architectures because of their minimal failure rates and computational efficiency. During inference, the models demonstrated exceptional performance, achieving top-tier accuracy on ImageNet. A crucial aspect of deep learning algorithms is transfer learning. By leveraging transfer learning and fine tuning techniques, various hyperparameters were utilized to enhance training and optimization processes. In order to improve the accuracy and decrease overall loss and Adam optimizer is essential for modifying the biases and learning rates of neural networks. A loss function evaluates an algorithm’s ability to fit the obtainable data in overall deep learning. The loss function uses an optimization function to help it gradually learn to minimize the incorrect detection. To rectify this specific problem, the binary cross-entropy loss function and Adam optimizer are used to identify the exact classification of normal and pulmonary fibrosis. Figure 1b. shows the workflow of the proposed model. Here, the input images contain various image sizes like 224x224, 256x256. hence resize process is necessary to standardize the size Subsequently, the dataset was divided into training, validation, and testing subsets. The CNN model was utilized to classify lung diseases, with the best model selected based on its performance evaluation, focusing on enhanced accuracy and a low-loss function.

3.1 Baseline model

A conventional Neural Network (NN)with 50 layers deep, ResNet50V2 stands for Residual Networks of 50 version 2 and is a key concept for many computer vision tasks. The original ResNet50 network is the source of ResNet50V2.

Researchers were able to successfully train very Deep Neural Networks (DNNs) with over 150 layers from the ResNet design. Training data with DNN architecture was challenging before ResNet of dying gradients. A deep residual learning approach was used to solve this issue. In order to create deep ResNet and connections with shortcuts that carry out for identifying the mappings were developed. These quick identity mappings had the advantage of not adding any additional parameters to the model and limiting the amount of time needed for computation. Figure 2a shows the modification of ResNetV2. Here, modifications are introduced in the form of an additional layer and a CBAM block to enhance the performance of the ResNet model. The CBAM was added between batch normalization and the activation function. The main goal of ResNet50V2 is to convey the output of addition function between the identity mapping and the residual function. A 224 × 224-pixel image is required as input for the ResNet50V2 network. Figure 2b illustrates the detailed architecture of ResNet50V2.

The original ResNet50V2 model is composed of stacked residual blocks, grouped hierarchically into stages to extract features. The proposed model also modified the architecture to incorporate an attention-enhanced residual block, as shown in Figure 2a. A CBAM-based attention module is specifically integrated during the transition between Conv4_block6 and Conv5_block1, representing the boundary between Stage 3 and Stage 4. This change retains the original residual learning framework and improves the recalibration of features using spatial and channel attention. CBAM module has standard configurations with mean and max pooling as channel attention and a convoluted layer with a 7x7 kernel size as spatial attention.

3.2 PSACA with ResNet50V2

This work introduced a new block inside the pretrained ResNet50V2 model to enhance the performance in detecting pulmonary fibrosis. The extra added block is called the attention improved residual block, i.e., Pyramid Spatial Atrous Channel Attention. The model consists of three primary components: Backbone, Neck, and Head. The Backbone employs ResNet50V2, comprising four consecutive convolutional stages for hierarchical feature extraction. At the output of designated stages (Stage 3 and Stage 4), a PSACA module restructured as Spatial → ASPP → Channel is utilized to augment fibrosis-specific characteristics via spatial attention and multi-scale context modelling employing Atrous Spatial Pyramid Pooling (ASPP). The Neck consolidates these enhanced feature maps using a multi-tiered fusion architecture. The Head does binary classification (Normal vs. Pulmonary Fibrosis), utilizing Grad-CAM to produce explainable images that emphasize clinically significant fibrotic areas.

PSACA module is strategically implemented at Stage 3 and Stage 4 of the ResNet50V2 architecture to improve the stage of feature representation with a higher level of semantic. The lower levels (Stage 1 and Stage 2) are the capture levels of low-level features like edges and simple textures that gives less informative when it comes to detect the pulmonary fibrosis. Conversely, Stage 3 learns mid-level structural features such as localized fibrosis texture, whereas Stage 4 learns high-level semantic and contextual features of abnormalities present in the lungs. When PSACA is used at these steps, it is possible to effectively refine the multi-scale features and enhance the capability of the model to reflect the local and global features of fibrosis without introducing the redundant computational complexity.

The initial CBAM design makes use of the average and Max Pooling along the channel to extract the feature to determine whether the CT images are normal or abnormal. The model is capable of dividing the input data correctly depending on the resulting feature maps. The given solution changes the pooling functionality to enhance the classification. The input images are frequently subject to processing to generate feature maps such that a range of low and high levels of activation occur. The adjusted high-M pooling selectively identifies the peak activations across channels and calculates their average. This high-M average pooling allows the model to differentiate between normal and abnormal cases by estimating the accumulated results against a specified threshold. Thus, the proposed model achieves enhanced classification precision by focusing on the most informative regions of the feature map.

The CBAM attention module is modified in the proposed new deep learning model, which is based on the ResNet50V2, to apply pyramid spatial atrous channel attention, resulting in a spatial-first attention strategy. By better focusing on lung regions that are crucial to the disease in CT images, this attention-driven framework enhances the ability to distinguish between normal and abnormal cases of pulmonary fibrosis.

The pyramid spatial atrous channel attention mechanism and the overall feature recalibration process: Spatial Attention First Equation 1 is given by

M S A F F = σ F k * k H i g h − M P o o l F ⊕ H i g h − M P o o l F (1)

Where, High-M Pool is used to perform average pooling along the channel dimension and it denotes the average of maximum activations across channels, designed to emphasize dominant feature responses, ⊕ is denotes the concatenation along the channel axis. F^k*k is represented convolution operation with a kernel size. σ is represented bounded activation function. The feature map f1 is a detailed function with below Equation 2,

F 1 = M S A F F ⊙ F (2)

⊙ represents the element-wise multiplication. Channel Attention, compute the channel descriptor A £ R c by average pooling. Where MSAF(F) denotes the spatial attention map of an input feature map F. In particular, spatial attention is calculated by performing high-M pooling operations along the channel dimension and then concatenating the results and convincing them with a k × k kernel. A sigmoid activation function is then applied to the resulting feature map in order to acquire the weights of the spatial attention. The input feature map F is element-wise multiplied with this attention map to create the refined feature map F¹, highlighting spatial regions of fibrosis-relevant features Equation 3.

A C = 1 H x W ∑ x = 1 H ∑ y = 1 w F 1 c , x , y (3)

The channel attention passes through a two-layer network with a reduction rating is Equation 4 w 2 £ R c * c r w 1 £ R c r * c are learned weights. δ , ∅ both are activation functions.

M c A = ∅ w 2 δ w 1 A (4)

The channel attention map M c £ R c is expanded back and applied to F¹ Equation 5

F o u t = M c ⊙ F A S P P (5)

Overall modified CBAM output is given in the below Equation 6,

F o u t = M c ⊙ F A S P P M S A F F ⊙ F (6)

This spatial First CBAM operation recalibrates feature maps by sequentially applying spatial attention and then channel attention. These equations summarize how the proposed model spatial-first CBAM recalibrates intermediate features to focus on pulmonary fibrosis regions, improving binary classification accuracy and interpretability.

The proposed work inserts one CBAM-enhanced residual block between Conv4_block6out and Conv5_block1out in ResNet50V2. This CBAM has 2 ways to learns the features, channel attention and spatial attention. Channel attention is used to learns which features channels are most important. Spatial attention is used to learns which spatial locations are most important. These both combine attention and give reduced distraction from irrelevant boundaries like ribs, shadow, and improve the features focused on fibrosis areas.

Even though PSACA is used at different stages, the design does not involve progressive over-filtering, but the hierarchies of semantically distinct features. Stage 3 represents the mid-level structural patterns like local fibrosis textures, whereas Stage 4 represents global contextual representations. ASPP incorporation also provides multi-scale diversity assuring that there is no redundancy and the refinement of features is complementary at different stages.

3.3 Proposed pyramid spatial atrous channel attention (PSACA) architecture

Figure 3 depicts the multi-scale feature responses of the ASPP module employing atrous rates of 1, 5, 10, and 15. Lower rates capture intricate structural details, including edges and vascular patterns, whereas higher rates encode more extensive contextual information related to patch-level and global lung architectures. The fused ASPP output integrates complementary information across all scales, providing a rich representation for the analysis of pulmonary fibrosis. This is the feature map used to easily identify the affected part of the lung for classification. The step-by-step process equations are provided in the description below.

The selection of dilation rates (1, 5, 10, and 15) in the ASPP module is intended to achieve a balance between the extraction of local and global features. The smaller dilation rates are more sensitive to fine-grained texture features like subtle fibrosis patterns whereas the bigger dilation rates are more sensitive to broader contextual information like global lung structure. The combination allows efficient multi-scale modeling features of pulmonary fibrosis. Even though dilation rate configurations can be different, the selected one offers a practical trade-off between feature diversity and computational efficiency.

The input feature map is derived from, spatial attention block, Equations 7, 8,

F 1 = M S A F F ⊙ F (7)

F A S P P = A S P P F 1 (8)

ASPP applies parallel convolutions with different dilation rates. For K*K atrous convolution with dilation d, the operation is Equation 9

Y d i , j = ∑ m = 1 k ∑ n = 1 k F i + d * m , j + d * n W m , n d (9)

Where, W^(d) is the convolution kernel for dilation rate d, d is dilation rate (1, 5, 10, 15), (i, j) is spatial position. 1x1 Convolution branch Equation 10 are given below,

F 1 x 1 = σ W 1 x 1 * F 1 + b 1 x 1 (10)

In this equation, dilation 1 is the default value. Then the Atrous Convolution Branches Equation 11 is written below, for dilation rates d = 1 , 5 , 10 , 15

F d = σ W 3 x 3 d * d F 1 + b d (11)

Where * d is dilated convolution. After that, Image-Level Pooling Branch is represented by the Equation 12 below

F p o o l = σ W p * U p s a m p l e G A P F 1 (12)

Image-level pooling generates a global context vector, which is then expanded and projected back to the original spatial resolution, restoring the spatial structure for subsequent processing.

Feature Concatenation, All the functions are concatenated channel-wise are denoted with the given below the Equation 13

F c o n c a t = C o n c a t F 1 x 1 , F 6 , F 12 , F 18 , F p o o l (13)

Final Fusion (1 × 1 Conv) Equation 14 is written below

F A S P P = σ W f * F c o n c a t + b f (14)

Where, W f is the fusion 1x1 convolution. After that, the Final Formulation of the proposed model is described with the given Equation 15

F A S P P = σ W f * C o n c a t F 1 x 1 , F 1 F 6 , F 12 , F 18 , F p o o l (15)

This final formulation is the complete function of the ASPP layer output.

In this work applies transfer learning to eight pre-trained models namely VGG 19, MobileNetV2, ResNet-50, ResNet-50V2, ResNet-152V2, DenseNet-169, DenseNet-121, and DenseNet-201. The ImageNet baseline dataset served as the original training set for these variation models. The previously mentioned models endure explicitly fine-tuned CT scan images of pulmonary fibrosis datasets of lung disease classification. The Customized ResNet50V2 model architecture is shown in Figure 4. The proposed model used initial convolution layers and 3x3 intermediate convolution layers in a bottleneck of the residual block. Fully connected layers are essential for classification and the last 20 layers are used to organize the fine-tune modification based on all models. Initializing the base model using pre-trained ImageNet weights serves as the foundation for fine-tuning the pre-trained ResNet as well as the remaining models. A Global Average Pooling (GAP) layer is added to the ReseNet50V2 with other models in order to minimize the dimensionality, during each block convolutional basis in all model architecture remains unchanged. Furthermore, the GAP layers reduced the feature map size before passing it to the dense layers while sustaining the highest accuracy. A flattening layer was then included. After the data has been flattened, it is sent to a dense layer with 512 neurons and the rectified linear unit (ReLU) activation function is used to learn complex, so the layers are activated. To prevent overfitting, a dropout layer with a 50% dropout rate is applied following the layer containing 512 neurons. Finally, the labels are shown by a dense layer with a single neuron with sigmoid activation to provide the detections and make it appropriate for binary classification tasks such as pulmonary fibrosis and normal. The pulmonary fibrosis dataset of lung disease detection is then used to retrain the entire architecture. The last 20 layers are modified to enhance the performance. It is mathematically represented by Equation 16,

L = L 0 , L 1 , L 2 , ⋯ , L M − 1 (16)

This is the general form layer modification. L is the set of layers from the model and M is the total number of layers. This layer-modified equation can be Equation 17 based on the proposed model of unfreezing the layer,

∀ L i ∈ L M − 20 , L M − 19 , L M − 18 , ⋯ , L M − 1 (17)

The final 20 layers of ResNet50V2 usually include batch normalization and ReLU activation layers after the final convolutional layers from the last residual block. These are essential for enhancing and stabilizing the model’s functionality. Following these layers, the network has a global average pooling layer that preserves significant feature information while reducing the spatial dimensions. In order to build the final probability distribution across the target classes, the model ends with a sigmoid activation layer after a dense layer for classification. Applying the Global Average Pooling (GAP) operation channel-wise, each feature map’s average across all spatial locations (x, y) is calculated by Equation 18,

G A P C = 1 H x W ∑ x = 1 H ∑ y = 1 w P x , y , C (18)

GAP_c is represented by output for the cth channel. The height and width of the feature are denoted by H & W. P_x,y,c is the activation of the spatial position (x,y) of the channel c. Following GAP the final output form is (1, 1, C) which is flattened to (C). ReLU and sigmoid activation functions can be represented by the Equation 19 given below, ReLU Function

f M = max 0 , M (19)

Where M denotes the input of the activation function, max(0, M) represents the maximum value between 0 & M. If the input value is positive, it simply returns the value; if it is negative, it returns 0. Sigmoid Function Equation 20

σ S = 1 1 + e − s (20)

Where S is the input (logits) from the neural network, and outputs denote a probability value between 0 and 1. The output yields a probability score between 0 and 1, representing the probability that the input CT scan corresponds to a patient with Pulmonary Fibrosis. The ASPP CBAM-ResNet50V2 is the proposed model for the classification of pulmonary fibrosis from lung diseases. Classifier output is after passing through the modified ResNet50V2 backbone enhanced with CBAM. The final output logits for the binary classification Y £ R 2. The Equation 21 is given below, F l a s t is the final feature map layer from the backbone, w D and b D are the dense layer weight and bias. Y £ R 2 is the predicted distribution over the Normal versus Pulmonary Fibrosis classes.

Y = s i g m o i d w D . F l a t t e n G A P F l a s t + b D (21)

3.4 Algorithm for pulmonary fibrosis classification

Algorithm 1 Algorithm: Pulmonary Fibrosis Classification.

4 Experimental setup

This section begins with a detailed overview of the pulmonary fibrosis classification dataset utilized in this study. To enhance model performance and generalization, various image preprocessing techniques which include image normalization were applied and addressing the large volume of data effectively. Following this the software requirements for training and evaluating the model are discussed in detail. Lastly, the optimization of several hyperparameters is examined.

4.1 Dataset description

In this research work, the publicly available dataset was used. Pulmonary fibrosis Dataset gathered from Kaggle. The dataset distribution of chest CT images within each class across the training, validation, and testing data is shown in Table 1.

TABLE 1. Dataset distribution of CT images.

Bold values indicate the total counts.

There are 14,390 balanced samples of CT images in the collection, 7,195 of which are pulmonary fibrosis, and the remaining 7,195 are normal pulmonary CT images. The normal class images represent the without fibrosis and the abnormal class represents the images with pulmonary fibrosis these are the two separate classes that the distribution creates inside the dataset. The dataset was split into three sets, with the training sample consists of 8,634 images (4,317 pulmonary fibrosis images and 4,317 normal images), the validation sample consists of 2,878 images (1,439 pulmonary fibrosis images and 1,439 normal images), and the testing sample consists of 2,878 images (1,439 pulmonary fibrosis images and 1,439 normal images).

To visualize the dataset class distribution split, a pie chart can effectively represent the distribution across training, validation, and test sets, as shown in Figure 5.

The total balanced samples of the data are divided into 60% for training (8,634 samples), 20% for validation (2,878 samples), and 20% for testing (2,878 samples). This makes it easier to understand how the dataset is distributed for the creation and evaluation of models. The dataset is split into training, validation, and testing sets to serve distinct purposes: the training set is used to learn weights and features; the validation set is used to tune hyperparameters and monitor overfitting; and the testing set is used for final performance evaluation. The images were resized to have a pixel resolution of (224x224). Sample images are shown in Figures 6, 7.

4.2 System and software requirements

The model was established and trained using 64-bit version of Windows 11 with 6 GB of memory and powered by 13th Gen Intel core i7-13650HX processor operating at 2.60 GHz. To enhance the training efficiency an Nvidia GeForce RTX 3050 GPU was also used to speed up the process. TensorFlow was the main deep learning package used for model construction, training, and assessment while Python was used as the main development tool for the implementation. The system requirements tabulated in Table 2. The initial layers of the base model were frozen when the pre-trained ResNet model was imported from Keras. The entire model was then re-trained after the proposed layers were customized using chest CT images of the pulmonary fibrosis. Compare the model with various CNN models were carried out in order to validate this work experiment using the particular validation procedures described in Section 7.

TABLE 2. System and software requirements.

4.3 Hyper-parameter optimization and settings

This section discusses the hyper-parameter modifying process and the ultimate configurations used to optimize the deep learning model performance for binary classification. Several hyperparameters are empirically adjusted to achieve the best model training configuration and the anticipated outcomes which includes the batch size, selection of optimizer, learning rate, loss function, and epochs. Optimizing hyper parameter is an essential process to improve the performance of models which are used to classify lung diseases. Grid search (Shekar and Dagnew, 2019) is one of the most popular techniques to investigate a predetermined portion of the hyper-parameter space. However, the number of alternative configurations increases exponentially with the number of hyper parameters, making this approach computationally costly. Random search (Mantovani et al., 2015) is an alternative method that chooses the hyper-parameter combinations at random assessment. This approach gives a stochastic nature frequently results in huge variations, even though it can observe a wide range of values. The effectiveness of Grid Search and Random Search in optimizing models for lung disease categorization is limited subsequently neither use data from previously established hyper-parameter settings to inform subsequent searches.

The model which is used a more sophisticated method for identifying the ideal hyperparameters are discussed in Bayesian Optimization proposed by Nguyen (2019). Previous studies have shown that Bayesian Optimization quickly finds the best hyperparameters with less iterations than Grid Search. A surrogate model is used by Bayesian Optimization is to direct the search process based on observed results, in contrast to Grid Search which thoroughly evaluates every possible combination. The count of neurons in the dense layer, dropout rate, optimizer type, learning rate, and batch size were among all the hyperparameters that were optimized in the model investigation. The objective function for optimization was to minimize the negative evaluation accuracy. Table 4 gives the search space for the hyperparameters identified in this research work. After completing the 18th iteration of Bayesian Optimization, the optimal hyperparameters were identified. Since a model with lower computed loss value is more efficient, minimizing loss is the primary objective. Binary Cross-Entropy (BCE) is used to compute the predicted probability distribution that matches the actual class labels (0 or 1). 0 represents the Normal and 1 represents the disease. The low computation loss function of binary cross entropy is represented by Equation 22,

B C E = − 1 / N ∑ 1 N K log M + 1 − K log ⁡ 1 − M (22)

Where N denotes the quantity of data samples, K denotes the binary value (1 or 0) and M denotes the probability. The binary cross-entropy loss function is chosen because it interprets the output as probability value between 0 and 1. It is especially well-suited models that uses sigmoid activation function. Adam was chosen because of its simplicity with effective memory management and capacity and for faster learning as the optimizer for model training. Adaptive Moment Estimation (Adam) integrates the advantages of both RMSProp and momentum optimization techniques. It is ideal for intricate deep-learning models since it uses first-order and second-order moment estimates to modify the learning rate for every parameter separately. In comparison, the optimizers like SGD, Adam also works well across a variety of architectures which is resilient to noisy gradients and usually needs less modifying. Because of these features, it is a recommended effectively to optimize the deep learning models. All six models use Adam optimizer which has been initialized as 0.001 learning rate.

In order to facilitate effective information processing to preserve computational memory, a batch size of 32 was used. With the pre-trained ImageNet weights are retained with a dropout rate of 0.5 was used to improve the regularization during modification. To ensure steady optimization, images from the training set were processed with 32-image mini-batches during training. In order to attain optimal performance, the refined ResNet model was additionally trained for 30 epochs. Consistent hyper parameter settings and experimental setups all chosen various models such as VGG 19, MobileNetV2, ResNet-50, ResNet-50V2, ResNet-152V2, DenseNet-169, DenseNet 121, and DenseNet-201 are trained and evaluated. Tables 3, 4 present an overview of optimal values for each hyperparameter which used in this study. The model was trained with a batch size of 32, utilizing the Adam optimizer with an initial learning rate of 0.001, and binary cross-entropy was chosen as the loss function. Training was carried out over 30 epochs.

TABLE 3. Search space and best hyperparameters identified.

TABLE 4. Hyper-parameter values.

5 Performance evaluation metrics

5.1 Accuracy

In the lung disease classification accuracy defined as the proportion of correctly classified cases, distinguish between normal and pulmonary fibrosis lung conditions. It is calculated as the ratio of correctly identified normal and pulmonary fibrosis cases to the total number of cases. Mathematically, accuracy is expressed as Equation 23:

A c c = TP 11 + TN 10 TP 11 + TN 10 + FP 01 + FN 00 (23)

Where,

TP (True Positives): Abnormal cases are appropriately classified as pulmonary fibrosis.

TN (True Negatives): Normal cases are appropriately classified as normal.

FP(False Positives): Normal cases are inaccurately classified as pulmonary fibrosis.

FN (False Negatives): Abnormal cases are inaccurately classified as normal.

A higher accuracy indicates a more effective model to distinguish between normal and abnormal lung conditions.

5.2 Recall

Recall, also referred to as Sensitivity or True Positive Rate, evaluates the model’s capacity to accurately detect pulmonary fibrosis lung conditions. It represents the fraction of actual abnormal cases that are correctly identified as pulmonary fibrosis. Mathematically, it is defined as Equation 24:

R e c a l l = TP 11 TP 11 + FN 00 (24)

A higher recall indicates that the model effectively detects the abnormal lung conditions to reduce the risk of missing diseased cases.

5.3 Precision

Precision is also known as Positive Predictive Value (PPV), which measures the accuracy of the model to predict the pulmonary fibrosis lung conditions. It represents the proportion of appropriately identified pulmonary fibrosis cases out of all cases classified as pulmonary fibrosis. Mathematically, precision is defined as Equation 25:

P R E C = TP 11 TP 11 + FP 01 (25)

Greater precision signifies that the model commits fewer false positive errors, enhancing its reliability in accurately confirming abnormal lung conditions.

5.4 F1 score

The F1 Score represents the harmonic mean of precision and recall, offering a balanced evaluation of a model’s performance. Mathematically, the F1 Score is defined as Equation 25:

F 1 = 2 * PREC * RECALL PREC + RECALL (26)

A high F1 Score signifies that the model achieves a good balance between precision and recall, ensuring accurate detection of abnormal lung conditions while reducing false classifications.

5.5 Mean average precision (mAP)

The performance metric Mean Average Precision (mAP) is frequently employed in task involving the object detection and classification. It assesses how well the model can classify and rank predictions, which makes it helpful for classifying the lung dis eases while evaluating both normal and pulmonary fibrosis instances. mAP, or mean Average Precision, represents the average AP scores across all categories, including normal and pulmonary fibrosis. AP is determined by calculating the area under the Precision-Recall (PR) curve, which visualizes precision in relation to recall across varying classification thresholds. Mathematically, AP can be defined as Equation 27

A P = ∫ 0 1 P R d R (27)

Where, P(R) is the precision as a function of recall. The mAP is calculated as the mean of the Average Precision (AP) across all classes Equation 28.

m A P = 1 c ∑ c = 1 c A P c (28)

C = number of Classes, APc = Average Precision for class c, mAP measures both precision and recall together. Higher mAP means better object detection performance. A higher mAP indicates that the model is more effective distinguish between normal and abnormal lung conditions with better trade-off between precision and recall.

6 Results and discussion

This section gives the findings of the optimized ResNet Versions testing on the binary classification task and an ablation study to investigate how these elements, modifications, and hyperparameters of the models affect their performances. Tweaking such parameters as batch size, layer adjustments, optimizer options, learning rate and loss function a more robust model with increased classification accuracy may be obtained. A study on ablation was carried out to investigate these changes and the findings of these analyses are discussed in the subsequent subsections.

6.1 Quantitative performance

The proposed ResNet50V2 with Pyramid Spatial Atrous Channel Attention architecture was estimated on the Pulmonary fibrosis dataset using a 60/20/20 train-validation-test division. Table 5 encapsulates the comprehensive classification performance for Normal versus Pulmonary Fibrosis. The integration of multi-scale feature extraction and attention techniques, which improve fibrosis-relevant feature representation, is responsible for the outstanding performance. Nonetheless, this almost flawless performance also requires a close consideration of model generalization on different data sets.

TABLE 5. Overall classification performance (test set).

The classification of pulmonary fibrosis disease detection of customized models on unidentified test illustrations is summarized. It indicates accuracy sensitivity or recall, precision, specificity, F1-score, and Mean Average Precision (mAP). Table 6 presents a performance comparison of various deep learning models which includes VGG19, MobileNetV2, ResNet-50, ResNet-152V2, DenseNet-121, DenseNet 169, DenseNet-201, and the proposed model ResNet-50V2, on a designated task focused on classifying images. The evaluation metrics include accuracy, recall, precision, F1 score, and mean average precision (mAP), which provide insights into the models’ ability to correctly classify instances, detect true positives, avoid false positives, and perform well across multiple classes. The proposed model ResNet-50V2 achieves the highest accuracy of 0.9981, with competitive results across other metrics, including recall (0.9964), precision (0.9988), F1 score (0.9976), and mAP (0.9999), demonstrating its superior performance compared to other models. The model performance both before and after customizing is discovered through the accuracy and loss graphs and both the graphs are shown in Figures 8, 9. The accuracy graph demonstrated a gradual improvement in validation and training accuracy before modifying when just the fully connected layers were trained while keeping the RseNet50V2 back bone frozen. The loss graph shows that validation loss plateaued early, indicating little feature adaption whereas training loss decreased. The final 20 layers were unfrozen after fine-tuning the accuracy graph, which exhibited a sharper improvement, indicating that the model was effectively picking up task-specific features. The loss graph established a more stable decay in both validation and training loss indicating higher generalization and enhanced performance.

TABLE 6. Performance comparison of customized models.

Although it has a high classification performance, the dataset is based on one publicly available source that can be a limitation to generalization. In spite of the dataset being even and well divided, a real-life clinical data can be much more varied in terms of protocols of acquisition, type of scanners, and patient groups. In this way, the next step of work will be the validation of the proposed model on multi-institutional datasets to make it robust and clinically applicable.

The graph represents the training and validation accuracy before and after modifying over 30 epochs. Initially, before modifying, both training and validation accuracy (Depicted by the solid blue line and the dashed green line, respectively) shows the fluctuations, particularly in the early epochs. This indicates that the model is still adapting and identifying patterns within the data. However, as training progress, accuracy steadily increases and stabilizes close to 1.0, suggesting improved model performance. After modifying, the accuracy (shown by the orange solid line for training and the red dashed line for validation) rapidly converges to near-perfect values with both the training and validation accuracy maintaining consistency without any major fluctuations. The minimal difference between training and validation accuracy further confirms that the model has generalized well and reduced the risk of overfitting. This indicates that modifying has significantly enhanced the model ability to classify lung diseases such as pulmonary fibrosis with high precision and reliability. The graph (Figure 9) illustrates the training and validation loss trends before and after modifying over 30 epochs. Initially, during the pre-modifying phase, both the training and validation losses are relatively high with noticeable fluctuations which indicates that the model is still learning and adjusting its parameters. The validation loss, represented by the green dashed line shows some instability, which suggests that the model is not yet fully optimized. However, after modifying the losses in both training (orange solid line) and validation (red dashed line) decreased significantly and stabilized at very low values which indicates that the model has effectively learned and generalized well to the validation data. The absence of large difference between training and validation losses suggests that overfitting is minimal, and the model is now more robust for lung disease classification. This highlights the significant impact of the modification in enhancing model performance and achieving better optimization.

6.2 Confusion matrix

The proposed model performance was evaluated using a confusion matrix (CM) to detect the correctly classified and incorrectly classified image data from the test dataset. Figure 10 demonstrates that the model correctly recognized 1,438 pulmonary fibrosis images, missing one in the pulmonary fibrosis class (Pulmonary_Fibrosis). The model was able to correctly identify 1,435 normal CT images, with four being overlooked in the no fibrosis class (Normal).

6.3 Receiver operating characteristics (ROC) curve

The effectiveness of the pulmonary fibrosis model is illustrated by the ROC curve which is shown in Figure 11. An AUC value of 1.00 exhibits robust discriminatory ability, as the curve distinctly illustrates that the model attains an effective equilibrium between sensitivity and specificity.

6.4 Qualitative results

Figure 12 shows the classification results. After training, the test images are passed through the trained model. If an image is correctly identified as normal, it is classified as normal. Similarly, if an image is correctly identified as pulmonary fibrosis, it is classified accordingly. However, some misclassifications occurred, which are illustrated in the confusion matrix.

The CT images of Figures 13a–f show different CT scans of lungs affected by Pulmonary Fibrosis and their corresponding Grad-CAM++ visualizations, which highlight the regions most indicative of disease based on a proposed model. Figure 13a shows the original chest CT image used as input to the model. Figure 13b depicts the binary lung mask derived from lung segmentation, distinguishing the left and right lung sections from the background. A lung-only CT picture is produced using this mask in Figure 13c to guarantee that subsequent analysis is confined solely to the lung parenchyma. Figure 13d displays the Grad-CAM++ heatmap generated from the trained pulmonary fibrosis classification model, with regions of higher intensity indicating greater contributions to the fibrosis prediction. To improve clinical interpretability, the Grad-CAM++ map is superimposed on the lung-only CT image in Figure 13e, emphasizing fibrosis-related activation areas restricted to the lung borders. Figure 13a illustrates the lesion mask derived from thresholding the Grad-CAM++ response in the lung region, with green representing fibrosis-affected lung tissue and red signifying very confident lesion cores. This Grad-CAM++-based visualization provides a weakly supervised and interpretable localization of pulmonary fibrosis patterns, enabling effective model explainability and aiding radiologist validation.

6.5 Ablation study for proposed model

An ablation study was conducted to illustrate the efficacy of the proposed feature extraction block, and this is shown in Table 7. To continue the discussion of the role of individual components in relation to the PSACA module, the role of spatial attention, ASPP and channel attention are discussed with the help of an architectural design presented in Figure 4. Spatial attention concentrates on the emphasis of fibrosis-relevant regions, ASPP concentrates on multi-scale contextual information, and channel attention concentrates on discriminative feature channels. The initial results show that when these components are used separately, the false positive is higher and the localization is worse. Conversely, the joint PSACA module yields a much lower false-positive rate through the joint use of spatial, contextual, and channel-wise refinement of features. This proves the complementary character of the proposed components and validates the success of the combined design.

TABLE 7. Ablation results impact of PSACA (Stage 3 & 4).

The best-performing values are highlighted in bold.

The proposed model significantly improves the multi-scale fibrosis capture, while PSACA strengthens mid-level and high-level feature in Stages 3 and 4, resulting in a notable increase in sensitivity and facilitating more precise detection of subtle fibrotic changes. The combined improvement from Stage 3 and Stage 4 confirms that the feature refinement, which gives the complementary information rather than redundant information, at each stage contributes distinct contextual and structural information.

The ablations research was carried out to determine how various model elements and hyperparameters are affected the proposed model performance. In this work, different components of the suggested ResNet50V2 model were changed, and the ensuing variations in metrics for performance were observed.

6.6 Case study i: chosen of batch size

The evaluation metrics for each batch size are captured in order to examine the impact of various batch sizes (16, 32, and 64) on the model performance. Batch size of 32 provides high accuracy of 99.81% while batch size 16 gives in reduction with the accuracy of 98.23% and batch size 64 takes more computational complexity with overfitting. The modified Batch size is shown in Table 8.

TABLE 8. Batch size comparison.

The best-performing values are highlighted in bold.

6.7 Case study ii: selection of best optimizer

The model was trained using various optimizers such as RMSprop, SGD, and Adam. The evaluation metrics for each optimizer were noted to identify the best performance for classification function. Among these three optimizers, Adam outperforms other optimizers in terms of effectiveness. The comparison of different optimizers is summarized in Table 9.

TABLE 9. Comparison of optimizers.

The best-performing values are highlighted in bold.

6.8 Case study iii: chosen of learning rate

To find the best learning rate for model training and testing was performed by the 0.01, 0.001, and 0.0001 learning rates. 0.001 gives the better performance compared to other learning rates. The modified Learning rate is tabulated in Table 10.

TABLE 10. Learning rate comparison.

The best-performing values are highlighted in bold.

The learning rate plays a vital role in model convergence and performance. When the learning rate (0.01) is increased, the updates become unstable and the convergence is suboptimal which reduces the accuracy. On the other hand, a very small learning rate (0.0001) causes the training process to be slow and it might take additional epochs to actually converge. The learning rate of 0.001 was chosen to offer the balance between the convergence speed and the stability of the validation to support efficient optimization and to reach the optimal classification results.

6.9 Case study iv: chosen of loss function

The evaluation of different loss functions for pulmonary fibrosis classification revealed significant differences in performance metrics. Binary Cross-Entropy (BCE) loss achieved the best result compared to both Focal Loss and Hinge Loss because it shows poor convergence. These results indicate that BCE is the most effective loss function for the classification task which ensures both high accuracy and stable training. The modified loss function is summarized in Table 11.

TABLE 11. Loss function comparison.

The best-performing values are highlighted in bold.

6.10 Case study v: selected layer modification with customization

In this study, the performance metrics were recorded after freezing and unfreezing of different layers. Unfreezing the last 10 layers, the model performance was a slight improvement with limited feature extraction. But unfreezing the last 20 layers, the model performance improved with high accuracy and minimal over-fitting, so the last 20 layers were chosen to modify based on the proposed model. The modified layers are given in Table 12.

TABLE 12. Layer modification with customization.

The best-performing values are highlighted in bold.

According to the ablation study, every element and hyperparameter is essential to the proposed ResNet50V2 model functionality. It determines the best arrangement for precisely detecting pulmonary fibrosis from CT images by systematically analyzing various setups. This through examination guarantees the proposed model effectiveness in clinical applications by highlighting its resilience and adaptability.

Overall, the proposed model demonstrates strong performance; however, further validation on external datasets is necessary to confirm its robustness in real-world clinical environments.

7 Comparison of the proposed model with the existing model

The performance evaluation of the current study as well as the previous research initiatives that have used Machine Learning and Deep Learning techniques for the diagnosis of pulmonary fibrosis using CT images which is shown in Table 13. The performance of various deep learning models for lung disease classification using CT images has improved significantly over the years (Wang et al., 2017). introduced a Multiscale Rotation-invariant Convolutional Neural Network (MRCNN), which achieves an accuracy of 0.9000. In their study (Kim et al., 2018), used a CNN architecture (4 convolutional layers and 2 fully connected layers) for medical image classification and obtained a high accuracy of 0.9512, indicating the model’s potent ability to identify anomalies. More recently (Shekar and Dagnew, 2019), employed a VGG16 model with image enhancement filters, further boosting accuracy to 0.9700. In comparison, the proposed method for 2025 utilizes the Modified CBAM-ResNet50V2 architecture, demonstrating superior performance with an impressive accuracy of 0.9983. This significant improvement can be attributed to the advanced residual connections and optimized feature extraction capabilities of Modified CBAM-ResNet50V2, which allow more efficient learning from complex medical imaging data. The proposed model effectively captures intricate patterns in CT images of pulmonary fibrosis and leads to highly accurate classification. The comparison between Dou et al. (2025), Selman et al. (2026) illustrates that recent models with the use of radiomics, transformer-based models, and hybrid models are characterized by high performance. Nevertheless, the suggested PSACA-ResNet50V2 model is more successful in combining multi-scale feature extraction, attention, and explainability, which leads to a higher classification accuracy than these approaches.

TABLE 13. Comparison of the proposed method with existing approaches.

8 Conclusion

This paper introduces an automated fibrosis detection system of the lungs, using a ResNet50V2 architecture, but with certain modifications, where the primary goal was to improve the learning of intermediate features. The main innovation of the described method is in the fact that it involves the inclusion of a Pyramid Spatial Atrous Channel Attention (PSACA) module that integrates the concept of spatial Atrous Spatial Pyramid Pooling (ASPP) with the idea of channel attention and is inserted in Stage-2 and Stage-3 of the ResNet50V2 backbone strategically. Such multi-scale placement facilitates multi-scale contextual modelling and also allows the selective focus of areas of fibrosis-relevant features at key feature levels, which promotes superiority of discriminating between normal and fibrotic lung patterns.

Although Stage 3 and Stage 4 yield similar accuracy when used separately, their combined integration results in enhanced sensitivity and specificity, demonstrating complementary feature refinement instead of redundancy. The proposed design enhances the feature representation and strength with progressive enlarging of receptive fields and adaptively reweighting the informative feature channels, and overfitting is reduced by the methods of training. The proposed model achieved overall evaluation performance with a classification accuracy of 0.9983, recall of 0.9993, specificity of 0.9972, precision of 0.9972, F1-score of 0.9983, and mean average precision of 0.9999 for a deep learning model used in the automated detection of pulmonary fibrosis from CT images. The experimental outcomes illustrate that the proposed architecture is able to provide better classification results than traditional backbone architecture, thereby supporting the fact that it is effective and has clinical implications in the reliable use of computer-assisted pulmonary fibrosis assessment using the CT images. Moreover, Grad-CAM++ visualizations validate that the network consistently focuses on fibrotic areas, enhancing interpretability and clinical confidence. Future work can extend this framework to multi-class classification and integrate with real-time diagnostic systems for broader clinical deployment. However, the current study is limited by the use of a single-source dataset, and further validation on external clinical datasets is required to confirm the generalizability of the proposed model.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Sonia M. Rocha-Sanchez, Creighton University, United States

Reviewed by: Raimondo Fanale, University of the System of the Italian Chambers of Commerce, Italy

Chaodong Fan, Hainan University, China