Pleural Effusion Classiﬁcation on Chest X-Ray Images with Contrastive

Learning

Felipe Andr

e Zeiser

1 a

, Ismael Garcia Santos

, Henrique Christoph Bohn

Cristiano Andr

e da Costa

1 b

, Gabriel de Oliveira Ramos

1 c

, Rodrigo da Rosa Righi

1 d

Andreas Maier

2 e

, Jos

e Rodrigo Mendes Andrade

and Alexandre Bacelar

Software Innovation Laboratory - SOFTWARELAB,

Universidade do Vale do Rio dos Sinos - Unisinos, S

ao Leopoldo, Brazil

Department of Computer Science, Friedrich-Alexander-Universit

at, Erlangen-N

urnberg, Erlangen, Germany

Medical Physics and Radioprotection Service, Hospital de Clinicas de Porto Alegre, Porto Alegre, Brazil

Keywords:

Contrastive Learning, X-Ray, Pleural Effusion.

Abstract:

Diagnosing pleural effusion is important to recognize the disease’s etiology and reduce the length of hos-

pital stay for patients after ﬂuid content analysis. In this context, machine learning techniques have been

increasingly used to help physicians identify radiological ﬁndings. In this work, we propose using contrastive

learning to classify chest X-rays with and without pleural effusion. A model based on contrastive learning is

trained to extract discriminative features from the images and reports to maximize the similarity between the

correct image and text pairs. Preliminary results show that the proposed approach is promising, achieving an

AUC of 0.900, an accuracy of 86.28%, and a sensitivity of 88.54% for classifying pleural effusion on chest

X-rays. These results demonstrate that the proposed method achieves comparable or superior to state of the art

results. Using contrastive learning can be a promising alternative to improve the accuracy of medical image

classiﬁcation models, contributing to a more accurate and effective diagnosis.

1 INTRODUCTION

Early pleural effusion detection is crucial for recog-

nizing the etiology of the adjacent disease, choosing

the ideal treatment, and reducing the worsening of the

patient’s health status (Hallifax et al., 2017). Pleural

effusion is a ﬂuid accumulation in the space between

the parietal and visceral pleura. Various diseases

and conditions, including infections, neoplasms, heart

failure, and chest trauma, can cause pleural effusion.

According to the literature, the late pleural effusion

diagnosis may be associated with worsening the pa-

tient’s health status or lead to other complications,

such as respiratory failure, sepsis, and even death

(Aboudara and Maldonado, 2019).

Chest radiography is an easily accessible tool to

https://orcid.org/0000-0002-1102-7722

https://orcid.org/0000-0003-3859-6199

https://orcid.org/0000-0002-6488-7654

https://orcid.org/0000-0001-5080-7660

https://orcid.org/0000-0002-9550-5284

detect pleural effusion. However, this ﬁnding can

be challenging in cases of small ﬂuid volume, as the

exam can detect volumes above 200 milliliters (Jany

and Welte, 2019). In this way, machine learning can

potentially contribute to early pleural effusion detec-

tion on chest X-rays. In addition, machine learn-

ing models can analyze large chest X-rays datasets to

identify patterns and characteristics associated with

the presence of chest X-ray ﬁndings (Bustos et al.,

2020; Liz et al., 2023). Therefore, it is possible

to develop automated classiﬁcation models that can

identify cases of pleural effusion, generating a double

reading tool.

One of the main challenges in developing ma-

chine learning models to detect chest X-ray ﬁndings is

the availability of previously labeled datasets (Cohen

et al., 2020). This scenario involves the arduous pro-

cess of identifying and annotating patient chest im-

ages by experienced radiologists (Bustos et al., 2020).

This absence of labeled data can limit the effective-

ness of machine learning models. Over the past few

years, works in the literature have relied extensively

Zeiser, F., Santos, I., Bohn, H., da Costa, C., Ramos, G., Righi, R., Maier, A., Andrade, J. and Bacelar, A.

Pleural Effusion Classiﬁcation on Chest X-Ray Images with Contrastive Learning.

DOI: 10.5220/0012205900003584

In Proceedings of the 19th International Conference on Web Information Systems and Technologies (WEBIST 2023), pages 399-405

ISBN: 978-989-758-672-9; ISSN: 2184-3252

399

on pre-trained models on natural image datasets, such

as for the ImageNet dataset (Baltruschat et al., 2019;

Zeiser et al., 2021; Guan et al., 2021). However, re-

cent studies have demonstrated problems in capturing

speciﬁc medical domain representations using pre-

trained weights (Azizi et al., 2021).

Recent advances address the use of self-

supervision methods to reduce model dependencies

on large datasets (Chen et al., 2020; Azizi et al., 2021;

Zhang et al., 2022). Generally, the work methodol-

ogy is based on a pre-training phase on weakly la-

beled datasets, and then a ﬁne-tuning of the weights

is performed on a fully labeled dataset to improve the

model’s capabilities. However, these methodologies

only have the ability to predict explicitly annotated

ﬁndings in the dataset, which still does not solve the

need to perform large-scale manual annotations for

less speciﬁc ﬁndings.

The availability of chest X-ray images and re-

ports represents a rich source of information for ma-

chine learning models. However, these data cannot be

used directly in supervised learning models without

the production of explicit labels by radiologists. In

this context, this article presents a contrastive learn-

ing model to classify chest X-ray images without ex-

plicit annotations during the training phase. The main

objective is to explore the contrastive learning effec-

tiveness in detecting pleural effusion in chest X-rays

in a context with the scarcity of labeled datasets. Our

method uses contrastive learning to identify key fea-

tures associated with pleural effusion. In addition,

this method can capture more complex characteris-

tics by analyzing natural language present in medi-

cal reports and learning through natural supervision.

Adopting contrastive learning provides a more efﬁ-

cient approach and reduces development costs since

it eliminates the need for extensive data labeling. The

results of our approach demonstrate that contrastive

learning can lead to more accurate and robust classi-

ﬁcation models, even when the labeled dataset is lim-

ited or incomplete (Radford et al., 2021). Our main

contributions are listed below:

• We propose a neural network model for weakly

supervised pleural effusion classiﬁcation on chest

X-ray images. The model can identify and asso-

ciate the images and reports characteristics, opti-

mizing the similarity of the characteristics vector

from two encoders based on Transformers.

• Our experiments demonstrate consistent perfor-

mance improvements in pleural effusion detec-

tion. Furthermore, the results outperform meth-

ods based on supervised learning, suggesting that

explicit labels are not necessary for pleural effu-

sion detection in chest X-ray images.

The remainder of this paper is organized as fol-

lows. Section 2 presents the most signiﬁcant related

works to deﬁne the present study. Next, section 3

presents the methodology of the work. The section 4

details the results and discussion. Finally, Section 5

presents the conclusions of the work.

2 RELATED WORK

The detection of ﬁndings in chest X-rays plays a cru-

cial role in the diagnosis and treatment of various dis-

eases. However, this task heavily relies on the ex-

pertise and experience of radiologists, making it sus-

ceptible to errors and human limitations. In recent

years, signiﬁcant progress has been made in the ﬁeld

of artiﬁcial intelligence, particularly in the develop-

ment of deep learning algorithms for medical image

analysis. Supervised models have been extensively

explored to aid in the analysis of chest X-ray images,

with several studies focusing on this area (Alghamdi

et al., 2021; Bustos et al., 2020; Elgendi et al., 2021;

Han et al., 2021; Rajpurkar et al., 2017; Wang et al.,

2017). These models have shown promising results in

assisting radiologists and improving the accuracy and

efﬁciency of chest X-ray interpretation.

Supervised learning approaches for classiﬁcation

based on radiological ﬁndings dependent on large

data sets are frequent in the current literature. For

example, (Rajpurkar et al., 2017) developed a ma-

chine learning model that achieved similar accuracy

to a suite of radiologists using a labeled dataset of

over 100,000 images. (Cohen et al., 2020) investi-

gates the performance of models in different domains

and the agreement between them in the X-ray images.

The authors perform experiments to analyze concept

similarity, using a regularization approach in a net-

work to group tasks on different datasets, and observe

the variation between tasks. One of the alternatives

to increase the number of samples for training is data

augmentation (Elgendi et al., 2021). However, data

augmentations commonly applied to tasks in the nat-

ural imagery domain cannot be fully extended to the

health domain. Therefore, labeled datasets are essen-

tial for training machine learning models for X-ray

images (Goceri, 2023).

Obtaining labeled datasets can be complex and la-

borious, especially in resource-poor contexts (Bustos

et al., 2020). In addition, the scarcity of labeled data

can lead to less accurate models and impair the effec-

tiveness of detecting radiological ﬁndings in the eval-

uated exams (Alghamdi et al., 2021). Some recent

work has explored the use of unlabeled or partially

labeled datasets to improve the detection of consol-

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

400

idation on chest X-ray images. For example, (Wang

et al., 2017) used semi-supervised learning to improve

exam consolidation detection.

Another promising approach is contrastive learn-

ing, which can be applied to unlabeled or partially

labeled datasets (Zhang et al., 2022). This feature

uses the similarity and differences between pairs of

images to learn discriminative characteristics (Chen

et al., 2020). Therefore, the method is beneﬁcial when

small-scale labeled data is available. For example,

(Han et al., 2021) used contrastive learning in con-

junction with transfer learning to improve consolida-

tion detection on chest X-rays.

However, works in the current literature are pri-

marily based on fully labeled datasets or require par-

tially labeled datasets in the training process. There-

fore, this work proposes to investigate the effective-

ness of contrastive learning to improve the detection

of pleural effusion in chest X-rays in the context of

training models with weak-labeled datasets.

3 MATERIALS AND METHODS

An overview of the methodology used in this work

is presented in Fig. 1. Our methodology can be di-

vided into four main steps: pre-processing, data aug-

mentation, training, and testing. Pre-processing con-

sists of resizing the image, normalizing the contrast,

and processing the report (Section 3.2). The data aug-

mentation step describes the methods used to generate

synthetic images (Section 3.3). Finally, in the train-

ing stage, the models and parameters used are deﬁned

(Section 3.4).

3.1 Dataset

We used three datasets to develop the model. For the

training step, we retrospective collected data from pa-

tients during the COVID-19 pandemic from two hos-

pitals in Porto Alegre/RS. The information collected

comprises clinical data, X-ray images and reports of

hospitalizations from 2020 to 2022. The ethics com-

mittee of each hospital approved the present study un-

der the Certiﬁcate of Submission for Ethical Consid-

eration (CAAE number 33540520.6.3004.5327) . In

addition, this document follows the General Data Pro-

tection Law (LGPD) recommendations. In Tab. 1, the

information relating to Private datasets 01 and 02 is

provided.

We used the PADCHEST public dataset for the

testing step. PADCHEST consists of chest X-rays

with reports. The dataset includes more than 160,000

images of 67,000 patients that were interpreted and

reported by radiologists at Hospital San Juan (Spain)

from 2009 to 2017, encompassing six views of dif-

ferent incidences, additional information on image

acquisition, and patient demographics. The dataset

comprises 19 differential diagnoses, with 27% of re-

ports manually annotated by physicians. In this sense,

we used only manually annotated X-ray images. In

Tab. 1, we present the information for each dataset.

3.2 Pre-Processing

We processed the images from the three datasets using

the same methodology. Given the tenuous differences

between healthy tissues and those affected by any al-

teration, contrast enhancement is commonly used in

the literature (Diniz et al., 2018). Furthermore, con-

trast enhancement can help increase the performance

of deep learning architectures (Pooch et al., 2020).

Therefore, we applied the Contrast-Limited Adaptive

Histogram Equalization (CLAHE) in the chest X-ray

images. CLAHE subdivides the image into subareas

using interpolation between edges. To avoid noise

build-up, use a gray threshold level, redistributing

pixels above that threshold in the image. The CLAHE

can be deﬁned by:

p = [p

max

− p

min

] ∗ G( f ) + p

min

(1)

where p is the pixel’s new gray level value, the values

max

and p

min

are the pixels with the lowest and high-

est values in the neighborhood, and G( f ) corresponds

to the cumulative distribution function (Zuiderveld,

1994).

Human anatomy is unique to each individual. This

aspect is reﬂected in the different anatomical sizes of

patients’ chests. In this way, the X-ray images have

different sizes. However, to process the X-ray images

in deep learning models, the images need a standard

width and height. Therefore, we resized the images

to 512X512 pixels. The reduction was proportional

to the width and height, with a zero padding for the

axis with the smallest size. The pre-trained architec-

ture delimited the image resolution, which will be dis-

cussed in Section 3.4. In Fig. 2, we show an example

with the original and preprocessed X-ray images. We

can observe (Fig. 2) that the image suffered a down-

sampling changing from approximately 2000X 3000

pixels to 512X512 pixels. In addition, the area related

to the lungs is highlighted about the other structures

represented in the image.

Finally, we pre-processed the Private datasets 01

and 02 X-ray reports. We keep only the alpha charac-

ters, removing special characters, numbers, and punc-

tuation from the reports. For the PADCHEST dataset,

it was only necessary to translate the pleural effusion

Pleural Effusion Classiﬁcation on Chest X-Ray Images with Contrastive Learning

401

Text Encoder

Image

Encoder

I1 * T1

I1 * T2

I1 * T3

I1 * T4

I2 * T1

I2 * T2

I2 * T3

I2 * T4

I3 * T1

I3 * T2

I3 * T3

I3 * T4

I4 * T1

I4 * T2

I4 * T3

I4 * T4

Training

tubo endotraqueal com extremidade no bronquio

principal direito cateter venoso central direita com

extremidade na projecao da veia cava superior demais

aspectos sem modificacao em relacao ao exame

anterior da vespera

faixa atelectasica em terco medio inferior do pulmao

direito infiltrado reticular difuso bilateral elevacao de

hemicupula diafragmatica direita aumento da area

cardiaca aorta ateromatosa seios costofrenicos laterais

livres

area de reducao da transparencia do parenquima

pulmonar na metade inferior do pulmao esquerdo

podendo representar area de consolidacao com

obliteracao do recesso costofrenico relacionada

moderado derrame pleural mediastino centrado indice

cardiotoracico dentro da normalidade recesso

costofrenico direito livre

presenca de volumoso derrame pleural esquerda com

discreto desvio do mediastino para direita nao ha

evidencia de lesao consolidativa ou tumescente

grosseira em pulmao direito nao foi possivel

determinar area cardiaca nao ha evidencia de derrame

pleural no seio costofrenico lateral direito

Testing

Text Encoder

Image

Encoder

I1 * T1

I1 * T2

sem presenca derrame pleural

presenca derrame pleural

Pre-processing

Original X-ray Pre-processed X-ray

Data Augmentation

Pre-processed X-ray Flipped X-ray

STEP

Figure 1: Summary of the proposed methodology. The text and image encoders are trained together to predict the correct pairs

from a training batch. The text encoder is fed with generic text, and the image encoder with the chest X-ray for prediction.

Table 1: Chest X-ray datasets for training and testing the architecture. Total.: Total; Image: Image; Man.: Manual; and Aut.:

Automatic.

Dataset Patients Annotation Label Tot. Img. Img. Used Used in

Private 01 697 Man. Report 6.650 6.650 Training

Private 02 1.754 Man. Report 7.084 7.084 Training

PADCHEST 69.882 Man./Aut. Report and Findings 160.000 17.513 Teste

Total 72.333 - Report and Findings - 173.734 -

ﬁnding from Spanish into Portuguese, which was per-

formed by automatic substitution.

3.3 Data Augmentation

We used a data augmentation methodology to in-

crease the cases in the training set. In this context, we

applied a horizontal ﬂip to each training set image.

The choice of only one data augmentation method

is related to the image characteristics. X-ray exams

represent anatomical structures that can be missed or

distorted by other data augmentation methods such as

distortion, shearing, or cropping.

3.4 Contrastive Learning Architecture

Current models used in the ﬁeld of computer vi-

sion require large datasets labeled (Dosovitskiy et al.,

2020). Speciﬁcally, in the context of imaging stud-

ies, obtaining large labeled datasets is a signiﬁcant

challenge. In this way, methods that can learn from

already established information may be able to scale

with greater generalization for the day-to-day use

of radiologists. In this sense, recent advances in

contrastive learning methods present a possibility of

learning representations in images using natural lan-

guage (Radford et al., 2021; Zhang et al., 2022).

In this way, our architecture shown in Fig. 1 is

based on Connecting text and images (CLIP) (Rad-

ford et al., 2021). The CLIP architecture is based on

the idea of zero-shot learning, seeking to ﬁnd relation-

ships between image and text pairs. This association

is built using two encoders, one responsible for ex-

tracting features representing the image and another

for the text. As a result, the model’s objective func-

tion is that both encoders produce feature vectors for

similar images and texts.

3.4.1 Training

The parameters for the encoders followed the CLIP

architecture (Radford et al., 2021). The image en-

coder is based on the ViT-B/32 architecture pre-

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

402

(a) Original X-ray (b) Pre-processed X-ray

Figure 2: Example of an original chest X-ray image (a) and pre-processed (b).

trained with CLIP model weights. A transformer ar-

chitecture with 12 layers of size 512 and 8 attention

heads was developed for the text encoder. The maxi-

mum length of the text string was 76 characters. The

implementation of the deep learning model followed

the hyper-parameters of the CLIP model.

The model was trained for 100 epochs using the

Adam optimizer with an adaptive learning rate of

0.00001. The batch size was 32 images. The model

was trained with the loss function of cosine similarity,

maximizing the similarity between the correct image

and text pairs.

3.4.2 Evaluation

For model evaluation, each image is processed with

two synthetic texts in Portuguese: (a) presence of

pleural effusion; and (b) without presence of pleural

effusion. First, the respective encoders calculate the

feature vector for the image and a feature vector for

each synthetic text. Then, the cosine similarity is cal-

culated with the normalization of the result performed

using a softmax function. The output of the softmax

function then represents the binary classiﬁcation for

the presence or absence of pleural effusion on the

chest X-ray. We used the area under the receiver oper-

ating characteristic (AUC), precision/recall curve, ac-

curacy, sensitivity, speciﬁcity, and F1-score as perfor-

mance metrics for evaluating the model. In addition,

we performed a bootstrap to generate the performance

metrics conﬁdence intervals.

4 RESULT AND DISCUSSION

This section presents the results of the proposed

method and the comparison with the current literature

for the detection of pleural effusion. The best weights

were chosen automatically based on the validation set

error. Tab. 2 presents the performance obtained for

the evaluation metrics in the PADCHEST dataset and

the respective comparison with the current literature.

Few studies in the current literature speciﬁcally

focus on the ﬁnding of pleural effusion. Therefore,

direct comparison and identifying what could be con-

sidered state o the art for this speciﬁc ﬁnding is com-

plex. Therefore, the survey of related works was

mainly based on work that detected different ﬁndings

in chest X-ray images.

Despite the different objectives, the results pre-

sented in Tab. 2 show that even without labels during

training, the methodology based on the CLIP (Rad-

ford et al., 2021) architecture is capable of achiev-

ing results very close or superior to supervised learn-

ing methods. At this point, it is essential to highlight

the sensitivity (88.54%) for detecting pleural effusion,

demonstrating that the model is capable of identifying

the ﬁndings.

Regarding the ROC curve (Fig. 3), the proposed

method obtained a value of 0.900. Compared with

other classiﬁcation methods using the same dataset

(Liz et al., 2023), our approach achieved superior per-

formance in terms of AUC-ROC. Furthermore, the

method proposed by (Liz et al., 2023) used part of the

PADCHEST dataset in the training process. While in

our method, PADCHEST was used only for method

testing. Therefore, it is possible that our method has

a superior ability to generalize multi-center ﬁndings

and may be more susceptible to use in clinical settings

to aid in diagnosing pleural effusion.

The results of our study demonstrate that the con-

trastive learning method can be a practical approach

to classifying pleural effusions on X-ray images. Al-

though there is space for improvement in the classiﬁ-

cation performance metrics and extension for a radio-

Pleural Effusion Classiﬁcation on Chest X-Ray Images with Contrastive Learning

403

Table 2: Comparison with related works. AUC: area under the receiver operating characteristic. Acc: Accuracy. Sen:

Sensibility. Spe: Speciﬁcity. F1: F1-score.

Study Dataset AUC Acc Sen Spe F1

(Cohen et al., 2020) Multiple 0.890 - - - -

(Guan et al., 2021) NIH ChestX-ray14 0.835 - - - -

(Ibrahim et al., 2022) NIH ChestX-ray14 - 88.86 - - -

(Liz et al., 2023) PADCHEST 0.656 - - - 56.70

(Rajpurkar et al., 2017) NIH ChestX-ray14 0.863 - - - -

(Serte and Serener, 2021) ChestX-ray8 0.780 75.00 100.00 67.00 -

(Zaidi et al., 2021) NIH ChestX-ray14 0.874 98.20 - - -

Ours PADCHEST 0.900 86.28 88.54 76.91 68.41

ROC Curve

True Positive Rate

False Positive Rate

Figure 3: ROC curve for the PADCHEST dataset.

logical multi-ﬁnding classiﬁcation model, the current

results are promising and open possibilities for stud-

ies of our model in clinical settings.

5 CONCLUSION

In conclusion, this study explored the use of con-

trastive learning to classify pleural effusion on chest

X-rays. The results showed that the proposed

approach could outperform reference models and

achieve signiﬁcantly higher accuracy. In addition, the

use of contrastive learning can be helpful in scenarios

with restricted information about the clinical behav-

ior of the disease, such as outbreaks and pandemics

of lesser-known or emerging diseases.

This work has a few limitations, as described next.

The ﬁrst is related to the technical limitation of be-

ing a model capable of ﬁnding only one ﬁnding per

image. Thus, studying methods that can identify dif-

ferent image ﬁndings is a prospect for future work.

Another limiting aspect is the data set for training.

In this way, using sets with a greater variety of ﬁnd-

ings is foreseen, allowing more excellent reliability

for clinical application. Finally, interpretability stud-

ies identify how the model creates the representations

and how it relates to them during the prediction pro-

cess. These studies can be extended to propose inter-

pretability methods for radiologists.

The beneﬁts of using contrastive learning to accel-

erate the creation of business models for real use are

signiﬁcant. The proposed method can help save time

and valuable resources in creating machine learning

models for healthcare applications, enabling faster

and more accurate diagnoses, and improving the qual-

ity of the patient’s hospital course. These advantages

make the proposed approach a promising tool to en-

hance the analysis of chest X-ray images in real clin-

ical scenarios.

ACKNOWLEDGEMENTS

The authors would like to thank the Coordination for

the Improvement of Higher Education Personnel -

CAPES (Financial Code 001), the National Council

for Scientiﬁc and Technological Development - CNPq

(Grant numbers 309537/2020-7 e 40572/2021-9), and

the Research Support Foundation of the State of Rio

Grande do Sul - FAPERGS (Grant number 21/2551-

0000118-6) for your support in this work.

REFERENCES

Aboudara, M. and Maldonado, F. (2019). Update in the

management of pleural effusions. Medical Clinics,

103(3):475–485.

Alghamdi, H. S., Amoudi, G., Elhag, S., Saeedi, K., and

Nasser, J. (2021). Deep learning approaches for de-

tecting covid-19 from chest x-ray images: A survey.

Ieee Access, 9:20235–20254.

Azizi, S., Mustafa, B., Ryan, F., Beaver, Z., Freyberg, J.,

Deaton, J., Loh, A., Karthikesalingam, A., Kornblith,

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

404

S., Chen, T., et al. (2021). Big self-supervised models

advance medical image classiﬁcation. In Proceedings

of the IEEE/CVF International Conference on Com-

puter Vision, pages 3478–3488.

Baltruschat, I. M., Nickisch, H., Grass, M., Knopp, T., and

Saalbach, A. (2019). Comparison of deep learning

approaches for multi-label chest x-ray classiﬁcation.

Scientiﬁc reports, 9(1):1–10.

Bustos, A., Pertusa, A., Salinas, J.-M., and de la Iglesia-

Vay

a, M. (2020). Padchest: A large chest x-ray image

dataset with multi-label annotated reports. Medical

image analysis, 66:101797.

Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. (2020).

A simple framework for contrastive learning of visual

representations. In International conference on ma-

chine learning, pages 1597–1607. PMLR.

Cohen, J. P., Hashir, M., Brooks, R., and Bertrand, H.

(2020). On the limits of cross-domain generalization

in automated x-ray prediction. In Medical Imaging

with Deep Learning, pages 136–155. PMLR.

Diniz, J. O. B., Diniz, P. H. B., Valente, T. L. A., Silva,

A. C., de Paiva, A. C., and Gattass, M. (2018). De-

tection of mass regions in mammograms by bilateral

analysis adapted to breast density using similarity in-

dexes and convolutional neural networks. Computer

methods and programs in biomedicine, 156:191–207.

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn,

D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer,

M., Heigold, G., Gelly, S., et al. (2020). An image is

worth 16x16 words: Transformers for image recogni-

tion at scale. arXiv preprint arXiv:2010.11929.

Elgendi, M., Nasir, M. U., Tang, Q., Smith, D., Grenier,

J.-P., Batte, C., Spieler, B., Leslie, W. D., Menon,

C., Fletcher, R. R., et al. (2021). The effectiveness

of image augmentation in deep learning networks for

detecting covid-19: A geometric transformation per-

spective. Frontiers in Medicine, 8:629134.

Goceri, E. (2023). Medical image data augmentation: tech-

niques, comparisons and interpretations. Artiﬁcial In-

telligence Review, pages 1–45.

Guan, Q., Huang, Y., Luo, Y., Liu, P., Xu, M., and Yang, Y.

(2021). Discriminative feature learning for thorax dis-

ease classiﬁcation in chest x-ray images. IEEE Trans-

actions on Image Processing, 30:2476–2487.

Hallifax, R., Talwar, A., Wrightson, J., Edey, A., and Glee-

son, F. (2017). State-of-the-art: Radiological inves-

tigation of pleural disease. Respiratory medicine,

124:88–99.

Han, Y., Chen, C., Tewﬁk, A., Ding, Y., and Peng, Y.

(2021). Pneumonia detection on chest x-ray using

radiomic features and contrastive learning. In 2021

IEEE 18th International Symposium on Biomedical

Imaging (ISBI), pages 247–251. IEEE.

Ibrahim, R. F., Yhiea, N., Mohammed, A. M., and Mo-

hamed, A. M. (2022). Pleural effusion detection us-

ing machine learning and deep learning based on com-

puter vision. In Proceedings of the 8th International

Conference on Advanced Intelligent Systems and In-

formatics 2022, pages 199–210. Springer.

Jany, B. and Welte, T. (2019). Pleural effusion in

adults—etiology, diagnosis, and treatment. Deutsches

Arzteblatt International, 116(21):377.

Liz, H., Huertas-Tato, J., S

anchez-Monta

es, M., Del Ser,

J., and Camacho, D. (2023). Deep learning for under-

standing multilabel imbalanced chest x-ray datasets.

Future Generation Computer Systems.

Pooch, E. H. P., Alva, T. A. P., and Becker, C. D. L. (2020).

A deep learning approach for pulmonary lesion identi-

ﬁcation in chest radiographs. In Brazilian Conference

on Intelligent Systems, pages 197–211. Springer.

Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G.,

Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark,

J., et al. (2021). Learning transferable visual models

from natural language supervision. In International

conference on machine learning, pages 8748–8763.

PMLR.

Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan,

T., Ding, D., Bagul, A., Langlotz, C., Shpanskaya, K.,

et al. (2017). Chexnet: Radiologist-level pneumonia

detection on chest x-rays with deep learning. arXiv

preprint arXiv:1711.05225.

Serte, S. and Serener, A. (2021). Classiﬁcation of covid-19

and pleural effusion on chest radiographs using cnn

fusion. In 2021 International Conference on INno-

vations in Intelligent SysTems and Applications (IN-

ISTA), pages 1–6. IEEE.

Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., and Sum-

mers, R. M. (2017). Chestx-ray8: Hospital-scale chest

x-ray database and benchmarks on weakly-supervised

classiﬁcation and localization of common thorax dis-

eases. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 2097–

2106.

Zaidi, S. Z. Y., Akram, M. U., Jameel, A., and Al-

ghamdi, N. S. (2021). Lung segmentation-based pul-

monary disease classiﬁcation using deep neural net-

works. IEEE Access, 9:125202–125214.

Zeiser, F. A., Costa, C. A. d., Ramos, G. d. O., Bohn, H.,

Santos, I., and Righi, R. d. R. (2021). Evaluation of

convolutional neural networks for covid-19 classiﬁ-

cation on chest x-rays. In Intelligent Systems: 10th

Brazilian Conference, BRACIS 2021, Virtual Event,

November 29–December 3, 2021, Proceedings, Part

II 10, pages 121–132. Springer.

Zhang, Y., Jiang, H., Miura, Y., Manning, C. D., and Lan-

glotz, C. P. (2022). Contrastive learning of medical

visual representations from paired images and text. In

Machine Learning for Healthcare Conference, pages

2–25. PMLR.

Zuiderveld, K. (1994). Graphics gems iv. In Heckbert,

P. S., editor, Graphics Gems, chapter Contrast Lim-

ited Adaptive Histogram Equalization, pages 474–

485. Academic Press Professional, Inc., San Diego,

CA, USA.

Pleural Effusion Classiﬁcation on Chest X-Ray Images with Contrastive Learning

405