A Novel Deep Learning Model for Pneumothorax Segmentation in

Chest X-Ray Images

Sonam Khattar

, Sheenam and Tushar Verma

Department of Computer Science and Engineering, Chandigarh University, Mohali, India

Keywords: Pneumothorax Detection, Deep Learning Models, U-Net Architecture, Segmentation Models, Chest X-Ray

(CXR) Analysis.

Abstract: Pneumothorax is a potentially fatal illness that must be identified quickly to prevent serious consequences.

Chest X-rays (CXR) are usually used to diagnose it, but manually interpreting the images takes a lot of time

and effort. Convolutional Neural Networks (CNNs), a type of deep learning approach, have demonstrated

potential in automating this procedure for quicker and more precise diagnoses. Using pre-trained backbone

networks—ResNet50, EfficientNet, and XceptionNet—this study creates a segmentation model based on the

U-Net architecture. Performance is evaluated using Dice Coefficient, IoU, and pixel-wise accuracy,

emphasizing segmentation quality and computational efficiency. Experimental results show that EfficientNet

achieves the optimal trade-off between accuracy and resource usage, ResNet50 delivers stable performance

with moderate computational needs, and XceptionNet provides superior precision at a higher computational

cost. Incorporating Dice Loss and Binary Cross-Entropy enhances stability and robustness against class

imbalance. This work demonstrates how backbone selection and augmentation strategies significantly impact

medical image segmentation, offering practical insights for developing more accurate and efficient diagnostic

models for real-world clinical applications.

1 INTRODUCTION

Pneumothorax, which can be defined as a collapsed

lung, is a severe medical condition that should be

diagnosed promptly and accurately. If the case is not

treated in time, complications become very grave,

and death might ensue. Among the most common

imaging modalities used for the diagnosis of

pneumothorax are chest X-rays because of their

simplicity and effectiveness in diagnosing such

medical conditions. However, the manual reading

process takes time and might go wrong. Technique

Deep learning techniques, specifically deep neural

networks, applied to medical images have proven to

be capable of automatic and diagnostic improvement

of the process, including diagnosing pneumothorax

(Wang, , et al. , 2017).

The natural pressure balance

necessary for lung expansion is upset by

pneumothorax, which is the presence of air in the

pleural cavity. It may arise as a result of trauma or

spontaneously (primary or secondary). Current

research emphasizes the evolution in the challenges

https://orcid.org/0000-0002-5444-4358

of managing spontaneous pneumothorax, like

recurrences and side effects: continued air leakage.

With new evidence-based guidelines for management

and new techniques developed in minimally invasive

procedures including chemical pleurodesis and

video-assisted thoracoscopic surgery (VATS),

treatment strategies especially for underlying

diseases are improved and recurrence prevented.

Pneumothorax disease segmentation is the

process or process of accurately identifying and

defining pneumothorax illness segments in medical

images, usually with the usage of chest X-rays or

other imaging modalities for diagnosis. The purpose

is to isolate areas where air may have accumulated in

the pleural space and caused it to collapse. Accurate

diagnosis, severity of the disease, and development of

appropriate treatment plans depend on segmentation.

The areas that need to be differentiated from the rest

of the lung tissue and other anatomical structures in

order to carry out pneumothorax segmentation are

usually the air-filled space between the lung and the

chest wall. This is often achieved through the

848

Khattar, S., Sheenam, and Verma, T.

A Novel Deep Learning Model for Pneumothorax Segmentation in Chest X-Ray Images.

DOI: 10.5220/0013733900004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 3, pages 848-853

ISBN: 978-989-758-763-4

application of advanced manual and automated

techniques, including artificial intelligence and deep

learning algorithms. These techniques are intended to

emphasize the extent of the pneumothorax, thus

helping radiologists and clinicians to measure the size

of the collapse and decide whether urgent

intervention is required.

Recently, extensive interest has been directed in

applying DNNs to classify pneumothorax using CXR

images. DNNs, specially CNNs, have shown

wonderful promise for various tasks in medical

imaging, which includes lung disease detection,

cardiac anomaly recognition, and tumor

classification. In these tasks, deep models can reach

very high accuracy and sensitivity as well as

specificity in these tasks; sometimes, they even

outperform traditional approaches by machine

learning or even human experts(Rajpurkar, et al. ,

2017). For instance, it has been proven that the 121-

layer DenseNet model developed by CheXNet has

enabled its application to match the performance of a

human radiologist in the detection of pneumonia

through CXR images (Irvin, et al. , 2017).

Despite all this, pneumothorax classification by

such techniques raises several challenges. This

includes problems of class imbalance, the

interpretability of the models, and large annotated

datasets (Liu, et al. , 2019). This article highlights the

key contributions in the detection of pneumothorax

by deep neural networks. The discussions highlight

the ongoing challenges as well as future directions

2 LITERATURE SURVEY

For the last couple of years, there has been intensive

exploration of using deep learning for the

classification of pneumothorax, for the main reason

that now it is possible to take advantage of large-scale

CXR datasets annotated by hundreds of thousands of

images. The development of a benchmark for deep

learning models was powered by ChestX-ray14, a

dataset of over 112 000 images annotated for 14

different thoracic conditions (Cao, et al. , 2020).

Specifically, pneumothorax was one of the 14

conditions that have been fine-tuned using CNN

architectures such as ResNet and DenseNet, known

to outperform most models in extracting the relevant

features from medical images(Krizhevsky,

Sutskever, et al. , 2012).

Pre-trained models have also fast-tracked

pneumothorax detection. A DenseNet-121 model

pre-trained on ImageNet and fine-tuned on the

ChestX-ray14 dataset attained a good area under the

curve of 0.93 for pneumothorax classification. This is

the principle of transfer learning wherein models pre-

trained on large general datasets are adapted to

specific medical tasks, which considerably improves

performance even with just quite minimal labeled

data (He, et al. , 2016).

The challenge offered a particular pneumothorax

dataset with more than 10,000 labeled CXR images

for further acceleration of deep learning models. This

challenge introduced several segmentation models,

such as U-Net and Mask R-CNN, and now not only

classify pneumothorax but also provide pixel-wise

segmentation, and therefore those models are

valuable for detection and localization

alike(Girshick, et al. , 2017). These models thus

demonstrate that segmentation does improve the

classification by establishing context related to

localizing the affected region.

Apart from classification in recent years, multi-

task learning has garnered much attention to improve

the performance of a model. Liu et al. (Liu, et al. ,

2019) recently proposed a multi-task learning

architecture that integrates both classification and

segmentation into its function. Such an architecture

enables end-to-end learning from two tasks at one

time, thereby improving the performance of

pneumothorax detection by leveraging

complementary information arising from the inherent

task of the segmentation mechanism.

Class imbalance in this problem is one of the most

significant problems encountered in pneumothorax

classification where the number of positive cases, that

is, pneumothorax is much limited as compared with

negative cases. Many strategies have been applied to

handle the problem with the help of weighted loss

functions and data augmentation techniques (Zhuang,

et al. , 2018). The weighted loss functions give more

penalties to the cases of misclassified cases that are

pneumothorax thereby reducing the class imbalance

effect. Another form of applying data augmentation

is rotating, flipping, and adjusting the contrast of

images in an artificial manner to create more diverse

positive samples so that the model can be made more

robust (Ronneberger, et al. , 2015).

Synthetic creation of training data To increase

data diversity, some researchers applied synthetic

data generation techniques. These comprise

Generative Adversarial Networks, GANs. GANs

allow the production of realistic images of

pneumothorax as additional examples that could

prove useful to overcome small imbalanced datasets.

In fact, Rajpurkar et al. (Goodfellow, et al. , 2014)

demonstrated that images of GAN-generated were

A Novel Deep Learning Model for Pneumothorax Segmentation in Chest X-Ray Images

849

useful for augmentation of images of small datasets

and improving model performance.

Another challenge of deploying deep learning

models in medical imaging involves explainability.

Clinical applications have to utilize whatever is put

into a model such that the results can be trusted and

interpreted by the radiologists. Techniques such as

Gradient-weighted Class Activation Mapping, or

Grad-CAM, can even hone in on regions of an image

to which the model responds most in terms of

supporting the prediction the model is generating,

thereby illuminating how the decision was made

(Selvaraju, et al. , 2017).

Despite the promising results, deep learning

models for pneumothorax detection are rarely used in

clinical applications. Before these models can gain

Table 1: Comparative analysis of existing work

Author(

Yea

Method

Employe

Objective Accura

Rajpurk

ar et al.

202

DenseNet

-121 +

Transfer

Learning

Pneumothor

classificatio

n using

CXR

ima

93%

Liu et

al.

202

Multi-task

Learning

Framewor

Simultaneo

classificatio

n and

segmentatio

95%

Singh et

al.

202

Mask

RCNN

with

ResNet10

Pneumothor

ax detection

and

segmentatio

96%

Wang et

al.

202

GAN-

Augmente

d Training

Data

Improved

performanc

e on small,

imbalanced

datasets

90%

Kaur et

al.

202

U-Net

Architectu

Enhanced

pixel-wise

segmentatio

n and

classificatio

94.5%

Smith et

al.

202

ResNet +

Weighted

Loss

Functions

Addressing

class

imbalance

for

pneumothor

ax detection

92%

widespread applicability to clinical settings, model

robustness and generalizability to different

populations must be ensured with effortless

integration into clinical workflows (Langlotz, et al. ,

2017). Clinical trials, coupled with regulatory

approvals, would be also paramount to the

establishment of the safety and efficacy of AI-driven

diagnostic tools.

3 DATASET REQUIREMENT

This study derives its dataset from a pneumothorax

segmentation challenge hosted on Kaggle, especially

targeting medical image segmentation tasks. It

comprises a set of grayscale chest X-ray images in

DICOM format along with the corresponding binary

masks indicating regions of interest. Here, the masks

are obtained in the format of Run-Length Encoding

(RLE) describing the presence and extent of

pneumothorax, which is essentially a collapsed lung.

An image without pneumothorax has been assigned

the mask value of -1. The dataset is divided into two

stages. Stage 1 comprises the training and test sets,

while Stage 2 consists of a training set expanded from

the Stage 1 training and test datasets with added

annotations. Segmentation masks are specified in a

stage_2_train.csv file, and test images are also

available in a zipped archive. The dataset was split

into training, validation, and test subsets to ensure

proper model training and unbiased evaluation.

Figure 1: Sample images of infected lung

4 PROPOSED WORK

This research deals with creating a segmentation

model based on the U-Net architecture, using three

pre-trained backbone architectures: ResNet50,

EfficientNet, and XceptionNet. The dataset consists of

chest X-ray images along with their corresponding

segmentation masks. All preprocessing steps include

scaling pixel values to a range of [0, 1] and resizing

every image and masks to the same resolution, for

INCOFT 2025 - International Conference on Futuristic Technology

850

instance, (256 x256). The dataset is split into training,

validation, and test groups, carefully so that

evaluation about the model's performance should not

be biased. The U-Net design is applied because it

serves fairly well for medical image segmentation. To

increase the effectiveness of feature extraction,

ResNet50, EfficientNet, and XceptionNet are

employed as encoders with weights pre-trained from

ImageNet for achieving transfer learning. For the

decoder component of U-Net, the balanced layout is

retained to achieve effective upsampling and

reconstruction of segmentation masks. This approach

enables closer analysis of how each backbone impacts

segmentation accuracy and computer efficiency.

Actually, there is a relatively small number of labeled

medical imaging datasets available, and data

augmentation is very important for making models

stronger and better at general use. Augmentation

methods are as follows: shape transformations

(rotating, flipping, cropping), brightness and contrast

adjustments, adding noise (such as Gaussian noise or

fake artifacts), and bending images to create more

realistic changes.

Figure 2: Proposed model for disease detection.

The Dice Loss method is especially useful for

handling imbalanced segmentation tasks, and it also

supports the addition of Binary Cross-Entropy to

improve stability. Adam optimizer along with a

learning rate scheduler supports efficient

optimization, while early stopping was used to check

the validation Dice Score, thereby reducing

possibilities of overfitting.Model performance is

verified using the Dice Score, that indicates overlap

between the predicted and actual masks.Other

evaluation metrics include IoU and pixel-wise

accuracy. The comparisons of the backbones are

conducted based on segmentation quality and

computer resource usage like inference time and

memory usage. Researchers conclude by doing

ablation studies to understand how backbones and

augmentation strategies influence results. They use

these results to derive the best setups for a medical

image segmentation task.

5 EXPERIMENTAL RESULTS &

DISCUSSION

The proposed U-Net-based segmentation model has

shown higher accuracy in chest X-ray segmentation

using the three backbones - ResNet50, EfficientNet,

and XceptionNet. The quantitative analysis obtained

shows that both training and validation loss decreased

incessantly throughout the epochs during

optimization. Moreover, the Dice Coefficient and IoU

metrics obtained proved effective performance,

although the best balance between accuracy and

computational efficiency was achieved by

EfficientNet. ResNet50 was yielding stable accuracy

with moderately demanding resources, while

XceptionNet provided better segmentation precision

at the cost of additional computational requirements.

The Dice Coefficient, in particular, showed

improvement by Dice Loss and Binary Cross-Entropy

combination, thus proving the robust nature of

predictions when dealing with the imbalanced

datasets.

(a)

(b)

Figure 3: Training and validation Loss(a) and IoU (b) of

proposed model

A Novel Deep Learning Model for Pneumothorax Segmentation in Chest X-Ray Images

851

(a)

(b)

Figure 4: Dice Coefficient(a) and Training and validation

Loss(b) of proposed model

Visual outputs illustrate the qualitative

performance of the models. The segmentation results

are very similar to the ground truth masks: the models

distinguish regions of anatomy well. Images illustrate

(a) the input chest X-ray, (b) ground truth masks, and

the model is able to capture great detail.

Figure 5: The output's visualizations include the following:

(a) the original image, (b) the ground truth mask, and (c)

the suggested segmentation result.

Data augmentation played an important role in

the improvement of generalization performance. The

effectiveness of rotation, flipping, adjustments to

brightness levels, and adding noise was even

confirmed through ablation studies. These augmented

datasets improved the IoU and Dice Coefficient

scores substantially, which in turn allow the model to

handle changes in real-world scenarios. Evaluation of

computational efficiency highlights EfficientNet as

the optimal backbone for practical applications due to

its low inference time and memory usage without

compromising accuracy. ResNet50 remains a safe bet

on the balanced performance front, with XceptionNet

suitable if accuracy is to be accorded more

importance over efficient use of resources. Taking

everything together-the U-Net architecture, pre-

trained backbones, effective loss functions, and

augmentation techniques-assures robust

segmentation in medical imaging tasks(Verma,

Tushar, et al. , 2024), (Verma, Tushar, et al. , 2024).

What the study has offered was to optimize model

performance based on real-world deployment as

much as it had suggested to balance between

accuracy and computation feasibility(Sheenam,

Sheenam, et al. , 2023).

6 CONCLUSION

This study proposed the U-Net-based segmentation

model that can effectively handle the chest X-ray

image segmentation task using pre-trained backbones

like ResNet50, EfficientNet, and XceptionNet. The

model demonstrated high accuracy in the

segmentation process, while EfficientNet presented

the optimal trade-off between accuracy and

computational costs. ResNet50 achieved relatively

stable accuracy with modest resource utilization,

whereas XceptionNet demonstrated higher precision

in segmentation at a greater cost of higher

computational requirements. The use of Dice Loss

and Binary Cross-Entropy improved the robustness

of the model, especially regarding the imbalanced

dataset challenge. In general, data augmentation, pre-

trained backbones, and careful model design led to

effective and efficient segmentation results.

Future work will involve more sophisticated data

augmentation like generative models and synthetic

data generation. The architecture of the U-Net can

also include an attention mechanism so that the model

gives major attention to the regions which are

important in the X-ray images. Lightweight models

and pruning techniques might be helpful in reducing

the computational overhead and making the model

usable in real-world clinical settings. This is also

INCOFT 2025 - International Conference on Futuristic Technology

852

further work in a larger scope dataset and in other

testing tasks to check its generalizability for any

medical imaging task.

REFERENCES

X. Wang et al., "ChestX-ray8: Hospital-scale chest X-ray

database and benchmarks on weakly-supervised

classification and localization of common thorax

diseases," in Proc. IEEE Conf. Comput. Vis. Pattern

Recognit., 2017, pp. 2097-2106.

P. Rajpurkar et al., "CheXNet: Radiologist-level

pneumonia detection on chest X-rays with deep

learning," arXiv preprint arXiv:1711.05225, 2017.

J. Irvin et al., "CheXNet: Radiologist-level pneumonia

detection on chest X-rays with deep learning," arXiv

preprint arXiv:1711.05225, 2017.

F. Liu et al., "Multi-task learning for segmentation and

classification of pneumothorax in chest X-rays," in

Proc. IEEE Int. Symp. Biomed. Imag., 2019, pp. 1286-

1290.

J. Cao et al., "Addressing class imbalance in medical image

classification," in Proc. IEEE Conf. Comput. Vis.

Pattern Recognit. Workshops, 2020, pp. 837-845.

A. Krizhevsky, I. Sutskever, and G. Hinton, "ImageNet

classification with deep convolutional neural

networks," in Adv. Neural Inf. Process. Syst., 2012, pp.

1097-1105.

K. He et al., "Deep residual learning for image

recognition," in Proc. IEEE Conf. Comput. Vis. Pattern

Recognit., 2016, pp. 770-778.

SIIM-ACR Pneumothorax Segmentation Challenge,

[Online]. Available: https://www.kaggle.com/c/siim-

acr-pneumothorax-segmentation

R. Girshick et al., "Mask R-CNN," in Proc. IEEE Int. Conf.

Comput. Vis., 2017, pp. 2961-2969.

F. Liu et al., "Multi-task learning for segmentation and

classification of pneumothorax in chest X-rays," in

Proc. IEEE Int. Symp. Biomed. Imag., 2019, pp. 1286-

1290.

L. Zhuang et al., "Lung disease detection in chest X-rays

using deep learning," in Proc. IEEE Conf. Comput. Vis.

Pattern Recognit. Workshops, 2018, pp. 121-130.

O. Ronneberger et al., "U-Net: Convolutional networks for

biomedical image segmentation," in Proc. Int. Conf.

Med. Image Comput. Comput.-Assist. Intervent., 2015,

pp. 234-241.

T. Goodfellow et al., "Generative adversarial nets," in Adv.

Neural Inf. Process. Syst., 2014, pp. 2672-2680.

R. Selvaraju et al., "Grad-CAM: Visual explanations from

deep networks via gradient-based localization," in

Proc. IEEE Int. Conf. Comput. Vis., 2017, pp. 618-626.

R. Langlotz et al., "A roadmap for foundational research on

artificial intelligence in medical imaging," Radiology,

vol. 291, no. 3, pp. 781-791, 2019

Verma, Tushar, et al. "Bone fracture detection using

machine learning." AIP Conference Proceedings. Vol.

3072. No. 1. AIP Publishing, 2024.

Verma, Tushar, et al. "Leaf Disease Detection Using

OpenCV And Deep Learning." 2024 Sixth

International Conference on Computational

Intelligence and Communication Technologies

(CCICT). IEEE, 2024.

Sheenam, Sheenam, Sonam Khattar, and Tushar Verma.

"Automated Wheat Plant Disease Detection using Deep

Learning: A Multi-Class Classification

Approach." 2023 3rd International Conference on

Intelligent Technologies (CONIT). IEEE, 2023.

A Novel Deep Learning Model for Pneumothorax Segmentation in Chest X-Ray Images

853