Early Stage Detection of Pulmonary Fibrosis using Deep Learning

Early Stage Detection of Pulmonary Fibrosis using Deep Learning

Byrisetty Aasreeja Suryaprakash, Dr. J. Jayaprakash, Dr. P. Dhivya

Department of Computer Science and Engineering, Dr. M.G.R. Educational and Research Institute, Chennai 600095, India

Abstract – Pulmonary Fibrosis (PF) is a progressive interstitial lung disease characterized by irreversible scarring of lung tissue, leading to respiratory failure if not detected at an early stage. Accurate and timely diagnosis using High-Resolution Computed Tomography (HRCT) scans is essential but remains challenging due to visual similarity between normal and pathological patterns and high inter-observer variability among radiologists. This paper presents a deep learningbased framework for automated pulmonary brosis detection from chest CT images using transfer learning. A labeled CT image dataset is preprocessed and parti- tioned into training, validation, and testing sets to ensure reliable evaluation. Pre-trained convolutional neural network models, including EfcientNetB0, ResNet50, and DenseNet121, are ne- tuned for binary classication of Normal and Pulmonary Fibrosis cases. The models are evaluated using precision, recall, F1-score, and accuracy metrics to address clinical reliability requirements. Experimental results demonstrate that the DenseNet121 model achieves superior performance with an accuracy of 99.95% and an F1-score of 0.9995 on the test dataset. To enable real- world applicability, the proposed framework is deployed as an end-to-end system using a Flask-based API and a web-based user interface for real-time CT image prediction. The developed system provides a reliable, fast, and deployment-ready solution to support radiologists in early pulmonary brosis screening.

Index TermsPulmonary brosis detection, chest CT imaging, deep learning, transfer learning, convolutional neural networks, DenseNet121, medical image classication, Flask deployment, web-based diagnosis system.

Highlights

• A deep learningbased framework is proposed for early detection of pulmonary brosis using chest CT images.

• Transfer learning models including EfcientNetB0, ResNet50, and DenseNet121 are ne-tuned for medical image classication.

• DenseNet121 achieved superior performance with an accuracy of 99.95% and F1-score of 0.9995 on the test dataset.

• The proposed system is deployed using a Flask-based API and web interface for real-time pulmonary brosis prediction.

  1. Introduction

     Pulmonary Fibrosis (PF) is a chronic and progressive in- terstitial lung disease characterized by irreversible scarring of lung tissue, which leads to a gradual decline in respiratory function [1], [2]. The disease signicantly reduces oxygen exchange efciency and often results in respiratory failure if not diagnosed and managed at an early stage. High-Resolution

     Computed Tomography (HRCT) is currently the primary imag- ing modality used for PF diagnosis, as it enables visualization of characteristic patterns such as reticulation, honeycombing, and ground-glass opacities [3].

     Despite advances in medical imaging, PF diagnosis remains challenging due to the subtle visual differences between normal and brotic lung tissues and the subjective nature of radiological interpretation [4]. Manual assessment of CT images is time-consuming and highly dependent on radiologist expertise, leading to considerable inter-observer variability. In many healthcare settings, the shortage of experienced radiolo- gists further exacerbates the problem, resulting in delayed or inconsistent diagnoses. These limitations motivate the need for automated and reliable computer-aided diagnostic systems to assist clinicians in early-stage detection.

     Recent advancements in deep learning have demonstrated remarkable success in medical image analysis tasks, including disease classication and segmentation [5], [6]. Convolutional Neural Networks (CNNs) have proven particularly effective in extracting hierarchical features from medical images without the need for handcrafted descriptors [7]. Transfer learning, which leverages knowledge from models pre-trained on large- scale image datasets, enables efcient training even when labeled medical datasets are limited [8]. Models such as EfcientNet, ResNet, and DenseNet have shown promising results in chest imaging applications by capturing both low- level texture patterns and high-level structural features [9].

     However, many existing pulmonary brosis detection ap- proaches remain conned to experimental environments and lack practical deployment frameworks [10]. Several studies focus solely on model accuracy without addressing real- world usability, scalability, or clinical integration. Further- more, black-box behavior of deep learning models and inade- quate evaluation practices limit trust and adoption in clinical workows. There remains a need for an end-to-end framework that not only achieves high diagnostic accuracy but also en- sures reproducibility, interpretability through reliable metrics, and deployment readiness.

     To address these challenges, this paper proposes a deep learningbased pulmonary brosis detection framework using transfer learning models applied to chest CT images. The proposed system incorporates standardized data preprocessing, robust model training, and rigorous evaluation using clinically relevant metrics such as precision, recall, and F1-score. Multi-

     ple CNN architectures are investigated, and the most optimized model is selected based on performance on unseen test data. In addition, the trained model is deployed using a Flask-based API and a web-based user interface to enable real-time image- based diagnosis. The major contributions of this work are summarized as follows:

• Development of a transfer learningbased classication framework for automated pulmonary brosis detection from CT images.

• Comparative analysis of EfcientNetB0, ResNet50, and DenseNet121 models for performance optimization.

• Rigorous evaluation using clinically meaningful metrics to ensure diagnostic reliability.

• Deployment of an end-to-end prediction system with a web interface for practical usability in clinical environ- ments.

The proposed framework aims to reduce diagnostic work- load, improve early detection accuracy, and provide a fast and objective second opinion for radiologists in pulmonary brosis screening.

1. Literature Survey

   Pulmonary brosis detection has received signicant atten- tion in recent years due to its clinical importance and the increasing availability of medical imaging data [11], [12]. With the widespread use of High-Resolution Computed Tomog- raphy (HRCT), researchers have explored machine learning and deep learning techniques to automatically identify brotic lung patterns and reduce diagnostic dependency on manual interpretation. Advances in articial intelligence for medical imaging have enabled automated feature extraction and im- proved diagnostic consistency compared to traditional manual assessment.

   1. Traditional Image Processing and Machine Learning Ap- proaches

      Early studies primarily relied on handcrafted feature ex- traction techniques such as texture analysis, histogram-based descriptors, and edge detection methods [13]. Statistical tex- ture features derived from Gray Level Co-occurrenc Matrices (GLCM), Local Binary Patterns (LBP), and wavelet transforms were commonly used to characterize lung tissue patterns [14]. These features were subsequently classied using conventional machine learning models such as Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), and Random Forest classiers [15]. Although these approaches demonstrated mod- erate success, their performance heavily depended on feature engineering quality and lacked robustness against variations in imaging conditions and disease severity.

      Hybrid systems combining image segmentation and ma- chine learning were also proposed, where lung regions were rst extracted using thresholding or region-growing techniques before classication [16]. However, these methods suffered from limited generalization capability and were sensitive to noise, scanner variability, and anatomical differences among patients.

   2. Deep Learning-Based Approaches

      With the emergence of deep learning, Convolutional Neural Networks (CNNs) have become the dominant paradigm for medical image classication tasks [17], [18]. CNN-based models automatically learn hierarchical representations from raw images, eliminating the need for manual feature design. Several studies applied custom CNN architectures for pul- monary disease classication, demonstrating improved accu- racy compared to traditional methods.

      Transfer learning has further enhanced performance by utilizing pre-trained networks such as VGG16, ResNet, In- ception, and DenseNet [19], [20]. These models leverage knowledge learned from large-scale datasets such as ImageNet and adapt it to medical imaging tasks with limited labeled samples. Research has shown that DenseNet-based architec- tures are particularly effective in capturing ne-grained texture variations in lung tissue, making them suitable for pulmonary brosis detection.

      More recent works have explored advanced architectures such as attention-based CNNs, explainable deep learning models, and transformer-based networks [22]. Biomedical segmentation architectures such as U-Net and UNet++ have inuenced modern medical imaging pipelines [36], while train- ing stabilization techniques including Batch Normalization and Dropout improve model convergence and generalization [37], [38].

   3. 2D Slice-Based vs. 3D Volumetric Analysis

      Most existing pulmonary brosis detection systems operate on individual 2D CT slices due to lower computational cost and simpler data handling [39]. However, this approach may ignore inter-slice contextual information that is important for comprehensive disease assessment. To address this limitation, some studies have proposed 3D CNNs to process entire CT volumes [40]. These volumetric models capture spatial continuity across slices and improve detection reliability but signicantly increase training complexity and inference time.

   4. Model Evaluation and Deployment Challenges

      Although many studies report high classication accuracy, several methodological limitations remain. Unrealistic evalua- tion strategies, such as random data splitting without patient- wise separation, can lead to data leakage and overestimated performance [26]. Furthermore, many works emphasize accu- racy as the primary metric while neglecting clinically relevant measures such as precision, recall, and F1-score. In addition, several high-performing models remain conned to experi- mental setups and lack deployment-ready frameworks for real- time clinical use [27].

   5. Research Gap and Motivation

   Despite substantial progress, existing approaches often suf- fer from reliance on handcrafted features, excessive compu- tational complexity, lack of realistic evaluation protocols, or absence of end-to-end deployment frameworks. There remains

   a clear need for a balanced system that combines high diagnos- tic accuracy with practical usability and real-time prediction capability.

2. Proposed Methodology

   The proposed pulmonary brosis detection framework is designed as an end-to-end deep learning system that performs automated classication of chest CT images into Normal and Pulmonary Fibrosis categories. The overall methodology consists of four main stages: data preprocessing, model con- struction using transfer learning, model training and optimiza- tion, and deployment for real-time prediction. The system architecture is illustrated in Fig. 1.

3. Dataset Description and Comparison

   1. Proposed Dataset

      The dataset used in this study consists of High-Resolution Computed Tomography (HRCT) images collected for pul- monary brosis classication. The images include both normal and brotic lung patterns and were preprocessed through re- sizing, normalization, and augmentation techniques to improve model generalization. Data augmentation methods such as ro- tation, ipping, and scaling were applied to reduce overtting and enhance model robustness.

      Each CT image was resized to a xed resolution before being input into the deep learning models. The dataset was divided into training, validation, and testing subsets to ensure unbiased performance evaluation. Patient-wise splitting was applied to prevent data leakage between training and testing samples.

   2. Previously Used Datasets in Literature

      Several studies have utilized publicly available CT datasets and clinical imaging repositories for pulmonary disease clas- sication. Earlier works employed HRCT image collections and interstitial lung disease datasets to train convolutional neural networks for automated diagnosis [1], [3]. Transfer learningbased approaches also relied on benchmark medical imaging datasets to compensate for limited labeled samples [6], [11]. Some researchers utilized large-scale medical image repositories combined with weakly supervised learning strate- gies to improve classication accuracy [20], [25].

      Despite the availability of these datasets, challenges such as class imbalance, limited annotation quality, and variability in imaging protocols remain signicant limitations in previous studies.

      TABLE I

      COMPARISON BETWEEN PREVIOUS DATASETS AND PROPOSED DATASET

      Study Dataset Class Size Limitation

      [1] HRCT Multi Med Limited deploy
      [3] Lung CT Binary Small Class imbalance
      [6] Transfer Binary Med Limited metrics Proposed HRCT PF Binary Large Balanced eval

   3. Dataset and Preprocessing

      A labeled dataset of chest CT scan images is utilized for model development. The dataset contains two classes: Normal and Pulmonary Fibrosis. All images are resized to a xed spatial resolution of 150 × 150 pixels and converted to three-channel RGB format to ensure compatibility with pre- trained convolutional neural networks. Pixel intensity values are normalized to the range [0, 1] to improve numerical stability during training.

      The dataset is partitioned into training, validation, and testing subsets using a stratied splitting strategy to preserve class distribution across splits. Data augmentation techniques, including random rotation, horizontal ipping, width and height shifting, and zooming, are applied to the training set to increase data diversity and reduce overtting. The validation and test sets are processed using only rescaling to ensure unbiased evaluation.

      TABLE II Dataset Distribution

      Category	Number of Images
      Normal	1500
      Pulmonary Fibrosis	1500
      Total	3000

   4. Model Training and Optimization

   The models are trained using the binary cross-entropy loss function and the Adam optimizer with an initial learning rate of 103. To improve training efciency and revent overt-ting, Early Stopping is employed to terminate training when validation loss ceases to improve, and ReduceLROnPlateau is applied to dynamically adjust the learning rate. Model Checkpointing is used to save the best-performing model based on validation accuracy.

   TABLE III HYPERPARAMETER CONFIGURATION

   Parameter	Value
   Image Size	150 × 150
   Batch Size	32
   Optimizer	Adam
   	103
   Learning Rate
   Epochs	50
   Loss Function	Binary Cross-Entropy

4. System Architecture

   The overall system architecture of the proposed pulmonary brosis detection framework is illustrated in Fig. 1. The system follows a modular pipeline consisting of data acquisition, pre- processing, deep learning-based classication, and deployment for real-time prediction.

   Chest CT images are rst collected from the dataset and passed to the preprocessing module, where image resizing, normalization, and augmentation are performed to standardize the input and improve generalization. The preprocessed images

   are then forwarded to the deep learning model built using transfer learningbased convolutional neural networks.

   The classication module employs pre-trained architectures such as EfcientNetB0, ResNet50, and DenseNet121, which are ne-tuned to discriminate between Normal and Pulmonary Fibrosis cases. The trained model outputs class probabilities, which are post-processed to generate the nal prediction label.

   For real-world usage, the trained model is deployed using a Flask-based API. A web-based user interface allows users to upload CT images and receive prediction results along with condence scores. This end-to-end architecture ensures scalability, fast inference, and seamless integration into clinical or laboratory environments.

   Fig. 1. System architecture of the proposed pulmonary brosis detection framework.

5. Proposed Deep Learning Models

   1. EfcientNetB0 Architecture

      EfcientNetB0 is a convolutional neural network architec- ture that employs compound scaling to balance network depth, width, and resolution efciently. The model utilizes Mobile Inverted Bottleneck Convolution (MBConv) blocks combined with squeeze-and-excitation optimization to enhance feature extraction while maintaining computational efciency. Due to its lightweight structure and strong performance on medical imaging tasks, EfcientNetB0 is suitable for pulmonary bro- sis classication from CT images.

      In this work, CT slices are resized and passed through the pre-trained EfcientNetB0 backbone. The extracted deep features are followed by global average pooling and fully connected layers for binary classication.

   2. ResNet50 Architecture

      ResNet50 is a deep residual learning network that introduces skip connections to address the vanishing gradient problem in deep architectures. Residual blocks allow the network to learn identity mappings, improving convergence and enabling deeper feature representation. The model extracts hierarchical visual features from CT images, capturing both low-level textures and high-level structural patterns.

      In the proposed framework, a pre-trained ResNet50 back- bone is ne-tuned using transfer learning. The nal classica- tion layers are modied to adapt the network for pulmonary brosis detection.

   3. DenseNet121 Architecture

   DenseNet121 introduces dense connectivity between layers, where each layer receives feature maps from all preceding layers. This design encourages feature reuse, reduces param- eter redundancy, and improves gradient ow during training. DenseNet architectures are particularly effective in medical imaging tasks where subtle texture variations must be cap- tured.

   In this study, DenseNet121 is used as a transfer learning backbone. The dense blocks extract ne-grained lung tissue features from HRCT images, followed by pooling and dense layers for classication.

6. Proposed Algorithm

   The proposed pulmonary brosis detection algorithm is designed as a sequential pipeline that processes chest CT images and produces a binary classication output indicating Normal or Pulmonary Fibrosis. The algorithm integrates image preprocessing, transfer learningbased feature extraction, and probability-based classication.

   1. Algorithm Steps

      Input: Chest CT image I

      Output: Class label y {0, 1}, where 0 denotes Normal and 1 denotes Pulmonary Fibrosis

      1. Acquire chest CT image dataset containing Normal and Pulmonary Fibrosis classes.

      2. Resize each image to a xed resolution of 150 × 150

         pixels and convert to three-channel RGB format.

      3. Normalize pixel intensity values to the range [0, 1].

      4. Apply data augmentation techniques such as rotation, ipping, zooming, and shifting on the training set to improve generalization.

      5. Load pre-trained convolutional neural network models (EfcientNetB0, ResNet50, and DenseNet121).

      6. Replace the original classication layer of each network with a custom binary classication head.

      7. Freeze the convolutional base layers and train the newly added layers using the training dataset.

      8. Fine-tune selected higher layers of the network using a reduced learning rate to improve feature adaptation.

      9. Compute prediction probabilities using the sigmoid ac- tivation function.

      10. Assign the nal class label based on a decision threshold of 0.5.

      11. Evaluate the trained model using accuracy, precision, recall, F1-score, and ROC-AUC metrics.

      12. Deploy the optimized model using a Flask-based API for real-time image prediction.

   2. Algorithm Description

   The algorithm begins by standardizing input CT images through resizing and normalization to ensure uniform fea- ture representation. Transfer learning is employed to leverage knowledge from large-scale image datasets, enabling effec- tive training with limited medical data. The DenseNet121

   model is selected as the nal classier based on superior performance during evaluation. The trained model outputs probability scores that are mapped to diagnostic labels and delivered to users through a web-based interface.

7. Problem Formulation

   Pulmonary brosis detection from chest CT images can be formulated as a binary classication problem. Given a set of CT scan images {I1, I2,…, IN }, each image Ii is associated with a class label yi, where yi = 1 denotes the presence of pulmonary brosis and yi =0 represents a normal lung condition. The objective is to learn a mapping function f (Ii) yi that accurately predicts the class label for unseen CT images.

   Let xi RH×W ×C represent the input image after pre- processing, where H and W denote the image height and width, and C denotes the number of channels. Each input image is resized to a xed spatial resolution and normalized to ensure numerical stability during training. The classier aims to minimize the binary cross-entropy loss function:

   N

   1 X

   TABLE IV

   Model Performance Comparison

   Model	Accuracy	Precision	Recall	F1-score
   EfcintNetB0	98.60%	0.986	0.985	0.985
   ResNet50	99.51%	0.995	0.994	0.994
   DenseNet121	99.95%	0.9995	0.9996	0.9995

   DenseNet121 achieved the highest classication performance with an accuracy of 99.95%, outperforming EfcientNetB0 and ResNet50. The superior performance of DenseNet121 can be attributed to its dense connectivity mechanism, which enables efcient feature reuse and improved gradient ow during training.

   In addition to performance evaluation, an analysis of datasets used in pulmonary brosis detection studies is illus- trated in Fig. 2. Previous research primarily relied on HRCT datasets and general lung CT image collections. Transfer learning approaches also utilized benchmark medical imaging repositories to compensate for limited labeled samples. In contrast, the proposed work employs a balanced HRCT dataset containing equal representation of normal and pulmonary

   brosis cases, ensuring reliable model training and evaluation.

   L =

   N

   i=1

   [yi log(yi)+ (1 yi) log(1 yi)] (1)

   HRCT Dataset (Walsh et al.)

   where yi denotes the predicted probability of pulmonary

   brosis for image Ii.

   The dataset is imbalanced, with pulmonary brosis samples 30

   forming a minority class compared to normal cases. In such 25

   scenarios, accuracy alone is not a sufcient performance indicator, as a classier biased toward the majority class may

   Lung CT Dataset (Kim et al.) Transfer Learning Dataset (Goy Proposed Dataset

   still achieve high accuracy while failing to detect brotic cases. Therefore, detection-oriented performance metrics such as precision, recall, F1-score, and area under the receiver operating characteristic curve (ROC-AUC) are emphasized to ensure clinical relevance.

   The problem further involves identifying a suitable deep learning architecture that can effectively extract discriminative features from CT images while maintaining generalization capability. Transfer learning is adopted to initialize the model parameters using weights pre-trained on large-scale image datasets, thereby reducing training time and improving con- vergence stability. The ultimate goal is to develop a model that achieves high sensitivity to pulmonary brosis while min- imizing false positives, enabling reliable automated screening in real-world clinical settings.

8. Experimental Results

   1. Quantitative Results

      The quantitative evaluation of the proposed pulmonary brosis detection framework was performed using standard classication metrics including accuracy, precision, recall, F1- score, and ROC-AUC. These metrics provide a comprehen- sive assessment of the models ability to correctly identify pulmonary brosis cases while minimizing false predictions.

      Table IV presents the performance comparison of the eval- uated deep learning architectures. Among the tested models,

      20 25

      Fig. 2. Distribution of datasets used in pulmonary brosis detection studies.

   2. Qualitative Results

   The qualitative evaluation of the proposed system demon- strates its usability and effectiveness in real-world scenarios. A web-based user interface is developed to enable real-time pulmonary brosis prediction from uploaded chest CT images, as shown in Fig. 3. The interface displays the predicted class label along with the corresponding condence score.

   Sample CT scan images used for model evaluation are presented in Fig. 4. These images illustrate representative examples of normal lung structures and pulmonary brosis patterns observed in HRCT scans.

   Visual inspection of prediction outputs conrms that the system consistently differentiates between Normal and Pul- monary Fibrosis cases. The average inference time per image is less than one second, enabling rapid screening and sup- porting clinical workow integration. The qualitative results indicate that the deployed framework provides fast, reliable, and interpretable diagnostic outputs, making it suitable for practical pulmonary brosis screening applications.

   Fig. 3. Web-based interface for pulmonary brosis prediction.

   Fig. 4. Sample CT scan images of (a) normal lung and (b) pulmonary brosis cases.

   To further evaluate the effectiveness of the proposed frame- work, a comparison with existing deep learningbased pul- monary brosis detection methods is presented in Table V.

   TABLE V

   Model	Acc.	Prec.	Rec.	F1
   CNN-Based Method	0.94	0.93	0.92	0.92
   Transfer Learning (ResNet50)	0.98	0.98	0.97	0.97
   Proposed Method (DenseNet121)	0.9995	0.9995	0.9996	0.9995

   Model Performance Comparison with Existing Methods

   The results demonstrate that the proposed DenseNet121- based approach achieves superior performance compared to conventional CNN models and other transfer learning archi- tectures. The improvement is mainly attributed to the dense connectivity structure of DenseNet, which enhances feature reuse and improves gradient propagation during training.

9. Limitations

   The proposed pulmonary brosis detection framework demonstrates strong classication performance; however, sev- eral limitations must be acknowledged. First, the current sys- tem operates on two-dimensional CT image slices and does not explicitly capture volumetric information across consecutive slices. This may limit the ability to model spatial continuity of brotic patterns in three-dimensional lung structures.

   Second, the dataset used in this study represents a specic imaging domain and acquisition protocol. Model performance may vary when applied to CT images obtained from different scanners, hospitals, or patient populations. Domain shift and variations in imaging quality may affect generalization capa- bility.

   Third, the proposed framework performs binary classica- tion and does not differentiate between disease severity levels or stages of pulmonary brosis. As a result, the system cannot be directly used for progression analysis or treatment planning. Finally, although high quantitative performance is achieved, the deep learning model remains partially black-box in nature. While prediction condence is provided, explicit visual expla- nations of decision-making are not yet integrated, which may

   affect clinical trust and interpretability.

   These limitations highlight the need for further validation on multi-center datasets, volumetric modeling, and explainable articial intelligence integration in future work.

10. Conclusion and Future Work

This paper presented a deep learningbased framework for automated pulmonary brosis detection from chest CT images using transfer learning. By leveraging pre-trained convolu- tional neural network architectures, including EfcientNetB0, ResNet50, and DenseNet121, the proposed system achieved high classication performance while maintaining computa- tional efciency. Among the evaluated models, DenseNet121 demonstrated superior accuracy and robustness, achieving a test accuracy of 99.95% with consistently high precision, recall, and F1-score values. These results conrm the ef- fectiveness of transfer learning for capturing discriminative pulmonary patterns in CT images.

In addition to model development, an end-to-end deploy- ment pipeline was implemented using a Flask-based API and a web-based user interface, enabling eal-time pulmonary brosis prediction from uploaded CT images. This practical integration transforms the research model into a deployable diagnostic tool and demonstrates its potential for clinical screening and decision support.

Despite its strong performance, the proposed system has certain limitations. The current framework operates on two- dimensional CT slices and does not explicitly model three- dimensional volumetric information across successive slices. Furthermore, the system performs binary classication and does not estimate disease severity or progression stages.

Future work will focus on extending the framework to incorporate three-dimensional convolutional neural networks for volumetric CT analysis, enabling more comprehensive spa- tial context modeling. Additionally, multi-class classication will be explored to differentiate between pulmonary brosis stages and other interstitial lung diseases. Integration of clin- ical parameters, such as pulmonary function test results and patient demographics, is also planned to enhance diagnostic reliability. Finally, explainable articial intelligence techniques will be investigated to provide visual interpretation of model

predictions, thereby improving clinician trust and facilitating clinical adoption.

Author Contributions

Byrisetty Aasreeja Suryaprakash: Conceptualization, dataset preparation, model development, experimentation, and manuscript drafting.

Dr. J. Jayaprakash: Research supervision, methodology validation, and technical guidance.

Dr. P. Dhivya: Review, editing, and academic supervision of the research work.

1. A. Dosovitskiy et al., An image is worth 16×16 words: Transformers for image recognition at scale, in Proc. ICLR, 2021.

2. Z. Zhou et al., UNet++: A nested U-Net architecture for medical image segmentation, IEEE Trans. Med. Imaging, vol. 39, no. 6, pp. 1856 1867, 2020.

References

1. A. Walsh et al., Deep learning-based classication of brotic lung disease from chest CT images, Radiology, 2022.

2. S. Humphries et al., CT-based deep learning model for automated diagnosis of idiopathic pulmonary brosis, Chest, 2021.

3. Y. Kim et al., Pulmonary brosis detection using CNN on HRCT images, Sensors, 2022.

4. R. Li et al., DenseNet-based classication of interstitial lung disease,

   Frontiers in Medicine, 2022.

5. M. Walsh et al., AI for diagnosis of brotic lung disease, European Respiratory Journal, 2022.

6. P. Goyal et al., Transfer learning for pulmonary brosis detection,

   Biomedical Signal Processing, 2023.

7. H. Choi et al., Automated classication using deep CNN, Scientic Reports, 2022.

8. T. Wang et al., Deep learning for interstitial lung disease classication,

   Computers in Biology and Medicine, 2022.

9. M. Raghu et al., Explainable deep learning for medical diagnosis,

   Nature Medicine, 2021.

10. L. Zhang et al., Attention-based deep learning network, IEEE Access, 2022.

11. R. Brown et al., Outcome prediction in brotic lung disease, Am. J. Respir. Crit. Care Med., 2022.

12. K. Suzuki, Overview of deep learning in medical imaging, 2021.

13. J. Ma et al., Transfer learning survey, Neural Computing and Appli- cations, 2023.

14. A. Esteva et al., Guide to deep learning in healthcare, 2019.

15. D. Litjens et al., Deep learning in medical image analysis, 2017.

16. G. Huang et al., Densely connected convolutional networks, CVPR 2017.

17. K. He et al., Deep residual learning, CVPR 2016.

18. Y. Gao et al., 3D CNN-based pulmonary disease detection, 2020.

19. A. Madani et al., View classication using deep learning, 2018.

20. T. Zhou et al., Weakly supervised learning for medical imaging, 2019.

21. J. Long et al., Fully convolutional networks, CVPR 2015.

22. P. Rajpurkar et al., CheXNet, arXiv 2017.

23. A. Krizhevsky et al., ImageNet classication with deep CNNs, NIPS 2012.

24. J. Deng et al., ImageNet database, CVPR 2009.

25. F. Chollet, Xception, CVPR 2017.

26. S. Minaee et al., Deep learning-based lung disease detection, 2021.

27. H. Shin et al., Deep CNNs for computer-aided detection, 2016.

28. J. Yao et al., Interpretable CNN, 2020.

29. B. van Ginneken et al., Computer-aided diagnosis, 2006.

30. T. Litjens et al., Medical image analysis survey, 2017.

31. A. Wong et al., COVID-Net, 2020.

32. S. Ioffe and C. Szegedy, Batch normalization, ICML 2015.

33. N. Srivastava et al., Dropout, JMLR 2014.

34. J. Bergstra et al., Random search optimization, JMLR 2012.

35. T. Hastie et al., Elements of Statistical Learning, 2009.

36. O. Ronneberger, P. Fischer, and T. Brox, U-Net: Convolutional net- works for biomedical image segmentation, in Proc. MICCAI, 2015, pp. 234241.

37. S. Ioffe and C. Szegedy, Batch normalization: Accelerating deep network training by reducing internal covariate shift, in Proc. ICML, 2015.

38. N. Srivastava et al., Dropout: A simple way to prevent neural networks from overtting, JMLR, vol. 15, pp. 19291958, 2014.