Computer Methods and Programs in Biomedicine 177 (2019) 285–296

Contents lists available at ScienceDirect

Computer Methods and Programs in Biomedicine

journal homepage: www.elsevier.com/locate/cmpb

An automatic method for lung segmentation and reconstruction in

chest X-ray using deep neural networks
Johnatan Carvalho Souza∗, João Otávio Bandeira Diniz, Jonnison Lima Ferreira,
Giovanni Lucca França da Silva, Aristófanes Corrêa Silva, Anselmo Cardoso de Paiva
Applied Computing Group (NCA - UFMA), Federal University of Maranhao, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: Background and Objective: Chest X-ray (CXR) is one of the most used imaging techniques for detection and
Received 11 March 2019 diagnosis of pulmonary diseases. A critical component in any computer-aided system, for either detection
Revised 24 May 2019
or diagnosis in digital CXR, is the automatic segmentation of the lung field. One of the main challenges
Accepted 5 June 2019
inherent to this task is to include in the segmentation the lung regions overlapped by dense abnormali-
ties, also known as opacities, which can be caused by diseases such as tuberculosis and pneumonia. This
Keywords: specific task is difficult because opacities frequently reach high intensity values which can be incorrectly
Lung segmentation interpreted by an automatic method as the lung boundary, and as a consequence, this creates a challenge
Lung reconstruction in the segmentation process, because the chances of incomplete segmentations are increased consider-
Chest x-ray
ably. The purpose of this work is to propose a method for automatic segmentation of lungs in CXR that
Convolutional neural networks
addresses this problem by reconstructing the lung regions “lost” due to pulmonary abnormalities.
Methods: The proposed method, which features two deep convolutional neural network models, consists
of four steps main steps: (1) image acquisition, (2) initial segmentation, (3) reconstruction and (4) final
segmentation.
Results: The proposed method was experimented on 138 Chest X-ray images from Montgomery County’s
Tuberculosis Control Program, and has achieved as best result an average sensitivity of 97.54%, an average
specificity of 96.79%, an average accuracy of 96.97%, an average Dice coefficient of 94%, and an average
Jaccard index of 88.07%.
Conclusions: We demonstrate in our lung segmentation method that the problem of dense abnormali-
ties in Chest X-rays can be efficiently addressed by performing a reconstruction step based on a deep
convolutional neural network model.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction The main advantage of CXR is the economical viability. Even in

underdeveloped countries, modern digital radiography machines
In present day, the world has available a variety of important are affordable [6]. In this context, CXR is certainly a very important
three-dimensional examination techniques, such as computer diagnosis tool that helps identity a large variety of lung diseases
tomography (CT) and magnetic resonance imaging (MRI). For chest around the world [7].
screening and detection of lung diseases, for example, CT has Due to these factors, millions of CXRs are generated annually.
been recognized as an successful approach to reduce lung cancer According to the National Health Service (UK), in 2017/18 over
mortality [1]. Several works in the literature have performed lung 22.9 million X-ray images were requested in the United Kingdom,
segmentation and nodule detection in CT scans [2,3]. These exams, which represents 55.63% of all imaging tests, including magnetic
however, have limited availability in most parts of the world, resonance (MRI) and computed tomography (CT) [8]. Among these
mainly due to their high cost [4]. X-Ray exams, 2.2 millions are CXRs. This represents a consider-
For the above mentioned reason, chest X-ray (CXR) is still able diagnosis workload, considering the shortage of radiologists
among the most used imaging tests worldwide, representing at worldwide [4,9].
least one third of all exams in a typical radiology department [5]. A CXR provides a large amount of information about a patient.
The correct interpretation of this information is a major challenge
to radiologists, and requires a high degree of skill, experience,
∗
Corresponding author. and concentration [10]. The distinction of abnormal structures
E-mail address: [email protected] (J.C. Souza). such as infiltrates from normal blood vessels, for example, is a

https://doi.org/10.1016/j.cmpb.2019.06.005
0169-2607/© 2019 Elsevier B.V. All rights reserved.
286 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296

challenging task even for experienced radiologists [10]. Frequently, In this section, we present some of the most relevant works in
when radiologists rate the severity of abnormal findings, large this research area. According to Van Ginneken et al. [12], these
interobserver and even intraobserver differences occur [11]. methods can be divided in rule-based methods, pixel classification
Due to the clinical importance and the complicated nature of schemes, deformable models, and hybrid methods.
CXR, researchers continue to explore the use of image processing Zheng et al. [25] proposed a semi-automatic rule-based method
and machine learning techniques to develop computational meth- that uses the anatomical information provided by the exam to
ods to assist radiologists in reading chest images [12]. Typically, determine initial borders for the lung fields. The method has
these techniques are combined to develop computer-aided detec- simple steps such as binarization and morphological operations,
tion (CAD) systems, which provide support to radiologists in the and uses a contour smoothing technique and projection curve to
challenging task of identifying abnormalities on patients. Two of perform the segmentation. Using a private dataset of 40 exams,
the key steps that are required to compose CAD systems are seg- this method obtained an average accuracy of 95%.
mentation [13,14,16,15] and classification [17] into medical images. Pixel classification methods explore the intensity differences of
An important step of CAD systems for CXRs is lung segmenta- the inner and outer regions of the lung. Annangi et al. [26] pro-
tion [6]. It provides structural information of shape irregularities posed a method that uses the contrast difference between lung
and size measurements of the lungs, which can be used to analyze fields and borders to orientate an active contour technique. The
severe clinical conditions such as pleural effusion, emphysema, and method obtained an average Dice coefficient of 0.88 in a dataset
pneumothorax [18]. Since a lung segmentation mask specifically of 1130 CXR images obtained from several hospitals in China.
determines the lung region, it also determines the non-lung region Deformable models are popular sets of methods that have been
by exclusion, which minimizes the effects of imaging artifacts in extensively studied and applied in medical image segmentation
the CAD system [19]. because they can be applied to segment a large set of biological
One of the major challenges of lung segmentation in CXR structures [27]. Xu et al. [28] proposed an optimized active shape
is to include in the segmentation mask the regions of the lung model (ASM) for lung field segmentation. The method, called
field overlapped by abnormal structures. In good scenarios, where ERF-ASM, was proposed to address one of the main limitations
patients are healthy, or in cases which they have only a small of typical ASMs, which is the need for an initialization suffi-
nodule, for example, CAD methods are usually able to provide ciently close to the target. Experimental results demonstrated
reliable segmentations, because the contrast between the lung a significant enhance of performance compared to other ASMs.
fields and their boundaries is maintained overall. However, in Tests were performed on a combination of two datasets, to-
more complicated (and also frequent) scenarios, patients may have talling 143 images, in which was reported an average accuracy
diseases that affect their lungs with dense abnormalities, such as of 95%, an average sensitivity of 91%, and an average specificity
opacities and consolidations [20,21]. In general, these abnormali- of 97%.
ties are ill-defined structures that overlap the lung field with high Another recent method based on deformable models is pre-
intensity pixels, which results in lower contrast between the lungs sented by Candemir et al. [22]. In this work, they present a
and their boundaries. For this reason, the segmentation task in nonrigid registration-driven lung segmentation method using
CXR becomes significantly more complex in these cases [22]. image retrieval-based patient specific adaptive lung models that
Given the presented scenario, the purpose of this work is to detects lung boundaries. The method consists of three main
present an automatic lung segmentation method that addresses stages: 1) a content-based image retrieval approach for identifying
this problem by performing a reconstruction step on the affected training images (with masks) most similar to the patient CXR
lung regions. The proposed method is basically divided into four using a partial radon transform and Bhattacharyya shape simi-
steps: (a) image acquisition, where we detail the materials used in larity measure, 2) creating the initial patient-specific anatomical
our method; (b) initial segmentation, which is based on an AlexNet model of lung shape using SIFT-flow for deformable registration
deep convolutional network (CNN) model [23]; (c) reconstruction, of training masks to the patient CXR, and 3) extracting refined
which is based on a ResNet18 CNN model [24]; and (d) final lung boundaries using a graph cuts optimization approach with
segmentation, obtained by combining the outputs of both CNNs. a customized energy function. According to the authors, this
The proposed method encompasses a series of contributions, method has surpassed the state of art in several previous works.
in which we highlight: (1) advances in AlexNet based on pixel The methodology was tested in two datasets: Japanese Society of
classification for initial segmentation of lung in CXR; (2) use of Radiological Technology (JSRT) Database, where it achieved a Dice
a reconstruction step based on ResNet18 that refines the initial coefficient index of 96,7%; and Montgomery County Dataset (MC),
segmentation; (3) a method based on lungs with dense anomalies, in which they achieved 96% of the same index.
making the method robust in segmenting even when it difficult According to Ngo and Carneiro [29], computer-aided diagnosis
to differentiate from healthy tissue; (4) a robust methodology of digital chest X-ray (CXR) images critically depends on the auto-
capable of fully automated segmentation of lung in CXR. For these mated segmentation of the lungs, which is a challenging problem
reasons, we believe the proposed method provides an efficient due to the presence of strong edges at the rib cage and clavicle,
segmentation step to any CAD system. the lack of a consistent lung shape among different individuals,
This paper is organized as follows. In Section 2 we present and the appearance of the lung apex. It has resulted in the publica-
a review of the related works in the literature. In Section 3 we tion in the last decades of robust hybrid methods, which are gener-
present and detail the materials and the proposed method used for ally based on the combination of different techniques, for example,
lung segmentation in CXRs. In Section 4 we detail the experiments techniques based on pixel classification and deformable models.
and the obtained results. In Section 5, we present a discussion Van Ginneken and ter Haar Romeny [30] proposed a robust
regarding the advantages and limitations of our method. Finally, hybrid method that combines a rule-based segmentation with a
in Section 6, we show our conclusions and perspectives for future pixel classification approach. The use of these two complementary
work. techniques achieved an average accuracy of 94% on a dataset of
230 images. In addition, the authors highlight that their method is
2. Related works very fast and can be implemented in any standard computer.
The fast advances of hardware power, particularly in GPUs, en-
Over the last decades, the scientific community have been abled deep learning approaches to become very popular in the last
publishing a number of methods for lung segmentation in CXR. years. Methods based on convolutional neural networks (CNN) and
 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296 287

recurrent neural networks, for example, achieved unprecedented expressive results in different domains of images, which is no
results in many research areas such as computer vision, natural different in CXR images. Recent work of Kalinovsky and Kovalev
language processing and voice recognition. These advances also [35], Saidy and Lee [36] use fully-connected networks for lung
reflected in methods of image processing for medical images, in segmentation. The advantage of using these networks is that there
which promising results were obtained in a wide range of tasks, is no need for a feature engineering to determine the best ones
such as segmentation of lungs [31,32], brain [33], breast [33], for a classifier to be able to differentiate classes, when using
spinal cord [34], among others. For instance, Kalinovsky and convolutional layers the network will be in charge of finding the
Kovalev [35] proposed a method in which an encoder-decoder most representative features for the problem.
convolutional neural network (ED-CNN) is used to perform lung Based on these advantages, disadvantages and limitations in
segmentation. Experiments were performed in a dataset of 354 related works, we propose a fully automatic and hybrid method
chest x-rays, in which was achieved an average accuracy of 96.2%. capable of efficiently segmenting lung into CXR. As we have seen
Another recent work published by Saidy and Lee [36], proposed in the literature, the deep learning methods have surpassed the
the use of a semantic segmentation based on a fully connected state of the art in several image domains. In proposing our method,
network for lung segmentation in the JSRT dataset. The network is we attempted to encompass two problems: lung segmentation,
called SegNet and similar to the work of Kalinovsky and Kovalev and reconstruction of segmentation in lungs with tuberculosis,
[35] uses an encoder-decoder approach. It is designed to map since CXR exams in tuberculous patients present regions of high
low-resolution features to input resolution for pixel-wise classi- intensity, often considered regions that are non-lungs, and we can
fication in order to produce features that are useful for accurate segment these lungs in a robust way. Therefore, we propose the
boundary localization. The authors achieved average dice index of use of two deep learning networks, the first capable of initially
96%, sensitivity of 95% and specificity of 99%. segmenting the lungs and the second capable of reconstructing
According to Hooda et al. [37], automatic analysis of chest regions with tuberculosis.
radiographs using computer-aided diagnosis (CAD) systems is
pivotal to perform mass screening and detect early signs of various 3. Materials and method
abnormalities in patients. In a chest radiographic CAD system,
segmentation of lung fields is a pre-requisite step to precisely The proposed method for lung segmentation in CXRs consists
define region-of-interest and is subsequently used by other stages of four steps, as illustrated in Fig. 1. In the first step we gather the
of the CAD system. They present a variant of fully-convolutional image dataset and the necessary information for experimentation.
network that performs segmentation of lung in CXRs. The archi- The second step is the initial segmentation, which is, in short, the
tecture achieves the testing accuracy of 98.92% and testing Jaccard classification of CXR patches performed by a deep convolutional
of 95.88% on the JSRT dataset. neural network (CNN). Step number three is the “reconstruction” of
Mittal et al. [38] proposed a method also based in deep learn- the segmentation mask, performed by a second deep CNN model.
ing. The network is called LF-SegNet, according to the authors a The final segmentation is obtained in step five, in which we com-
fully convolutional encoder’decoder network for segmenting lung bine the initial segmentation and the reconstructed masks into one
Fields from CXRs. The authors list the difficulty of the manual final mask. Each of these steps are detailed in the next subsections.
segmentation task and how automatic segmentation may be im-
portant for assisting specialist physicians. Also, the method has as 3.1. Dataset information
contribution a new unique architecture that can do an automatic
segmentation of the lungs. The method was tested in the JSRT and In this work, we use the Montgomery County (MC) Dataset,
MC datasets, and the results were 98.73% of accuracy and 95.10% which is a publicly available chest x-ray database from the De-
of Jaccard index. partment of Health and Human Services of Montgomery County,
Knowing that hybrid methods are producing the most accurate Maryland. This dataset was collected by the Montgomery County’s
results in lung segmentation and that deep learning methods have Tuberculosis Control program and consists of 138 CXR images, of
achieved state of art in several image domains, Ngo and Carneiro which 80 are from normal patients and 58 are from patients with
[29] proposed a method that combined the advances brought by some manifestation of tuberculosis. The CXR images are available
the deep learning techniques and the deformable models. The in 12-bit gray-scale with resolutions of either 4020 × 4892 or
authors used a Deep belief network and level-set for elaboration 4892 × 4020 and the pixel spacing in both horizontal and vertical
of the method and tested in the JSRT dataset. According to the directions is 0.0875 mm. In addition, the MC dataset has gold
authors, the results reached an average accuracy of 96.5% standard lung segmentation masks (Fig. 2), which were marked
As we can see in these works, there is a great search in com- under the supervision of a radiologist and made available by
putational methodologies for lung detection in CXR. Also, this task Candemir et al. [22].
is not something recent, over the years several methods have been
constructed for lung segmentation. We also note that several types 3.2. Initial segmentation
of approaches are used to achieve increasingly expressive results.
Over the years, methods have become more complete and with After acquisition of the image dataset, we train a deep CNN
higher validation metrics. model to obtain an initial segmentation of the lung regions. This
It is worth mentioning that rule-based approaches are the task can be divided into four substeps: division of CXRs into
simplest ones, which usually use morphological operations and patches; classification of patches using the CNN model; generating
binarization for lung segmentation, this type of approach does the segmentation masks by plotting the patches classified as lung;
not add any pattern recognition characteristics to its methods. and post-processing.
In turn, deformable methods are mostly knowlegde-based or use
atlases. Disadvantage in this type of approach is that it usually has 3.2.1. Division into patches
a certain size to solve the problem in a set of images, when new Initially, we scale all CXR images to 512 × 512, which is a
images of different size of the atlas are tested, not the guarantee resolution that satisfactorily retains the structural details of the
that the method will work. exam, while significantly reducing the model’s computational
With the advent of more sophisticated hardware, deep cost. Since images in the MC dataset may have two different
learning-based works are beginning to emerge, which present dimensions (4020 × 4892 and 4892 × 4020), we scale images in
288 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296

Fig. 1. Steps of the proposed method.

Otherwise, they are labeled as non-lung. The parameters patch

size of 32 × 32 and the threshold for being lung patch of 20% were
defined empirically after several experiments. Fig. 3 illustrates the
processes of dividing an CXR image in patches and labeling.

3.2.2. Patch classification

In this work, the classification of CXR patches is performed by
a deep CNN model, which is a biologically inspired model capable
of learning features in a hierarchical manner [23]. In these models,
the learning process occurs during training, through an implicit
feature extraction that has the advantage of eliminating the need
for manually determining and extracting the set of features to
Fig. 2. A sample from Montgomery County dataset. (A) A chest x-ray image and (B)
its ground truth lung segmentation mask. solve the addressed problem. Typically, a CNN model consists of
a number of convolutional, pooling and activation layers, each
having several hyperparameters, as detailed in [39,40].
a way that their proportions are not deformed. Scaling is done The CNN model used in our work for initial segmentation is
by, initially, placing the original CXR image in the center of a illustrated in Fig. 4. This model is based on AlexNet [23], which
background image of size 4892 × 4892 and pixels equal to zero, is a popular CNN architecture that gained notoriety after winning
and then resizing it to the final resolution of 512 × 512. the ImageNet Large Scale Visual Recognition Challenge in 2012.
After scaling, images are divided into patches of size 32 × 32, It consists of five convolutional layers with max-pooling after
which will be used as input to the CNN model in the next substep. the first, second, and fifth convolutional layers, followed by two
But first, we label these patches in two classes: lung and non-lung. fully connected layers with dropout regularization of 0.5, and an
The criterion used for labeling is based on the proportion of lung output layer with softmax activation for binary classification. The
pixels and non-lung pixels. Patches that contain 20% or more kernel sizes for all convolutional and pooling layers are, respec-
pixels of lung, according to the ground truth, are labeled as lung. tively, 3 × 3 and 2 × 2, and the number of neurons in each fully

Fig. 3. Patch division and labeling: The ground truth mask (A) is used as reference to label patches in a CXR image (B).
 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296 289

Fig. 4. The CNN architecture used for initial segmentation.

Fig. 5. Initial segmentation: A chest x-ray sample (A), the generated mask with Fig. 6. Compromised initial segmentation: A chest x-ray image (A) showing patient
false positives (B), and the post-processed segmentation mask (C). with severe lung lesions, the initial segmentation generated (B), and the ground
truth mask (C).

connected layer is 4096. Finally, the number of filters in the con-

volutional layers 1 to 5 are 96, 256, 384, 384 and 256, respectively. size 15 × 15 in the dilation, and an ellipse of size 19 × 19 in
the closing. Other sets of parameters have been tested for the
3.2.3. Plotting of lung patches purpose of post-processing. However, this was the set of param-
As mentioned in the beginning of this section, the initial eters that achieved the best results. Fig. 5 shows the result the
segmentation is obtained by the classification of CXR patches. post-processing in a CXR image.
Explicitly, our purpose is to build a model that is able to dis-
tinguish lung patches from non-lung patches. For this reason, 3.3. Reconstruction
during the training phase of our CNN model, patches can be
extracted without overlap, that is, side by side. However, in order In good scenarios, where patients are healthy and the CXRs
to generate a finer lung segmentation mask, we need to extract show lungs in good conditions, our initial segmentation alone
and classify the CXR patches with overlap. In other words, patches is able to achieve great results. However, in problematic cases
have to be extracted with a small stride, such as 1 or 2. This such as the ones mentioned in Section 1, our first CNN may not
stride value is directly related to the smoothness of the output include in the initial segmentation mask the lung regions severely
segmentation masks, and also to the number of model predictions, affected by opacities or consolidations. This is due to the fact
which translates to computational cost. that the model uses texture information to perform segmenta-
In this work we classify all patches in the CXR image using a tion, and therefore contrast difference is an extremely important
stride of 2. If a CXR patch is classified as lung, then its respective factor. Fig. 6 illustrates a case where the initial segmentation is
patch in the segmentation mask will have its four central pixels compromised by the above mentioned situation.
painted white, that is, set to 255. On the other hand, CXR patches In order to provide a solution to these particular cases, and
classified as non-lung will have their respective patches in the therefore develop a more general segmentation model, we train
segmentation mask painted black, that is, set to 0. In this way, we a second CNN to perform a reconstruction on the affected lung
determine our segmentation of the lungs. Fig. 5(B) illustrates the regions. After we train the patch classification network, we use
result of this process. the trained model to generate initial segmentation masks for both
the training and validation sets. At this point, we will also have
3.2.4. Post-processing the initial segmentation masks for the test set, since we have al-
After the plotting of all lung patches, we observed that a con- ready tested our initial segmentation network. Therefore, we have
siderable amount of false positives was generated in the resultant three new image sets, which are the initial segmentation masks
segmentation masks. Since morphological operations have been generated from the training, validation and test sets, respectively.
extensively used for false positive reduction in a large variety of With these three new sets we perform the training and testing of
works including lung segmentation [41–43], we chose to apply the reconstruction network.
some of these techniques in our post-processing step, which con- Inspired by the segmentor network in [45] we modify ResNet-
sists of: (1) an erosion, (2) a dilation, (3) an area filter that keeps 18 [24] to generate a reconstructed mask from the initial segmen-
the two largest objects, and (4) a closing operation. The area filter- tation obtained in the previous step. Fig. 7 details the architecture
ing technique has been used to perform simple lung segmentation of our reconstruction network.
routines in CXR [43,44], and is based mainly on the prior knowl- The training of the reconstruction network is performed
edge that the lungs are the two largest objects in a CXR image. with the initial segmentation masks resized from 512 × 512 to
The structuring elements used in the morphological opera- 128 × 128. This was necessary due to hardware limitations, since
tions are: a cross of size 19 × 19 in the erosion, an ellipse of an architecture this deep requires a considerable amount of mem-
290 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296

Fig. 7. The CNN architecture used for lung reconstruction.

ory to represent all feature maps in the fully connected layer. We reconstructed mask in a certain level, which deforms the con-
take the output of the last resblock of ResNet-18 and pass it to a tour previously established in the initial segmentation. Since the
fully connected (FC) layer to perform upsampling. Then, the output purpose of the reconstruction is to recover lung regions while
of the FC layer is passed to a final max pooling layer, which results keeping the previously segmented ones, we combine both masks
in the reconstruction of the lung regions. The final reconstructed to provide a final segmentation.
mask is obtained after a fine smoothing using a median filter with
kernel size of 5 × 5. 4. Results
In our reconstruction network we use batch normalization
[46] right after each convolutional layer and before activation. The In this section, we present the results of the proposed method
max pooling layers use kernels of size 3 × 3 with stride 2, and for for lung segmentation. Initially, we briefly present the metrics
all the convolutions in the resblocks we use filters of size 3 × 3. used for performance evaluation, then we detail the experiments
and discuss the results. In addition, we present a set of study cases
3.4. Final segmentation for further analysis, and also discuss and compare our method
with other related works.
The final step of the proposed method is the combination of We highlight that the proposed method is implemented en-
initial segmentation and reconstruction into a single mask. We tirely in Python, in which we used mainly the Keras deep learning
achieve this by performing a binary OR operation on the two library [47] with Tensorflow [48] as back-end. The machine used
segmentation masks. Fig. 8 shows an example of this process. in our experiments consists of an Intel Core i7-7700K CPU, 16GB of
In general, our first CNN model obtains precise contours for RAM, and two Nvidia GeForce GTX 1080-Ti graphic cards, running
most parts of the lungs. The motivation for using the reconstruc- on Windows 10 operating system.
tion step in this work is mostly the cases in which dense abnor-
malities affect the overall shape of the segmented lungs, which 4.1. Evaluation metrics
frequently result in a considerable loss of segmented lung regions.
As explained in Section 3.3, the main benefit of the recon- In order to evaluate the performance of the proposed method,
struction is that it is able to recover a considerable portion of we use a set of metrics widely used and accepted by the scientific
the missing lung regions. However, it also seems to erode the community to the assess medical image segmentation methods
[49]. The used evaluation metrics are the Dice similarity coeffi-
cient, the Jaccard index, also known as Intersection over Union,
Accuracy, Sensitivity, and Specificity.

4.2. Experimental results

This section presents an analysis of results of each step of the

proposed method, organized in the following order: 1) Dataset
preparation; 2) Initial segmentation results; and 3) Final segmen-
tation results. We also provide in this section a detailed analysis
of 4) Case studies, and a 5) Comparison with related works.

4.2.1. Dataset preparation

As described in Section 3, we use the MC dataset in our experi-
Fig. 8. Process of obtaining the final segmentation. ments. As already known, convolutional neural networks require a
 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296 291

Table 1 have different texture patterns, specially in normal patients, as

Number of patches generated for training, validation, and test.
illustrated in Fig. 9. In cases aggravated by dense pulmonary
Dataset Training Validation Test Total lesions, however, this factor becomes less evident.
Montgomery County 80 20 38 138
Lung patches 5136 1316 2804 9256 4.2.2. Initial segmentation results
Non-lung patches 10,279 2656 5121 18,052
As described in Section 3.2, the initial segmentation is obtained
using our AlexNet-based CNN, which performs patch classifica-
tion. Using a batch size of 128, the training of this CNN model
consisted of 500 epochs. We used stochastic gradient descent
as optimizer to minimize the difference between the ground
truth and the network output, and the loss function used was
the binary cross entropy [50]. The learning rate was set to a
fixed valued of 0.01. Weights were initialized using a normalized
Fig. 9. Patch samples: (A) Three lung patches and (B) three non-lung patches. initialization [51] and updated using the standard backpropagation
algorithm [40]. This set of parameters was selected after several
training experiments, in which this was the one that obtained the
more significant number of samples in the training phase [39,40],
best results.
so that the learning of patterns can be successfully achieved.
During validation, our model obtained 95.36% of sensitivity,
Despite our experimentation database having a smaller number
97.85% of specificity, and 97.02% of accuracy. In the test phase,
of samples compared to other deep learning applications, such as
it achieved 94% of sensitivity, 97.44% of specificity, and 96,22%
image recognition, we attempted to separate approximately 70%
of accuracy, which demonstrates that our model was capable of
for training (100 CXR images) and the remainder (30% of the total)
satisfactorily learning the patterns of both lung and non-lung
for testing. Thus, we believe this is a good ratio of images for both
patches in CXR without overfitting.
training and test.
As mentioned in Section 3.2 the initial segmentation consists of
In addition, during the training phase, there is a need for a
several steps. The post-processing is used to refine the output of
validation set, which is used to adjust the training with the best
our first CNN by reducing false positives. Table 2 shows the results
possible parameters, until the validation reaches significant results.
of the initial segmentation before and after post-processing.
For this purpose, we have used 20% of the training images for
It can be observed that despite the sensitivity rate has
validation. This way, from the 138 CXR images of this dataset, we
decreased after post-processing, all the other metrics were sig-
have a good ratio of training (80 images), validation (20 images)
nificantly improved. It is worth mentioning that the increases of
and test (38 images).
Dice and Jaccard indexes were considerably large, which translates
The number of patches generated to train the CNN in the initial
to improved segmentation masks with fewer false positives, since
segmentation step (Section 3.2) is detailed in Table 1.
these indexes indicate the overlap and similarity between objects.
It can be observed in Table 1 that the number of non-lung
It was obtained almost 9% of improvement of Dice coefficient and
patches is approximately two times higher than the number of
over 14% of Jaccard index. Fig. 10 shows the results of our initial
lung patches. This happens because the lung fields occupy less
segmentation for two images.
space in a CXR than the other structures overall, and consequently
this is reflected on the number of generated patches. In addition,
we ignore the patches extracted solely from the background, that Table 2
is, the patches that contain only pixels equals to zero. This is Initial segmentation results before, and after post-processing.
done for two reasons: the number of patches generated in each Initial segmentation Sensitivity Specificity Accuracy Dice Jaccard
class becomes less unbalanced; and we spare processing time by
Before post-processing 97.01% 91.33% 92.70% 85.68% 75.08%
not submitting patches without any relevant information to the After post-processing 93.54% 98.72% 97.57% 94.41% 89.64%
CNN model. In general, patches from lung and non-lung classes

Fig. 10. Initial segmentation examples. (A) Original CXR image, (B) initial segmentation, (C) post-processed segmentation, and (D) ground truth.
292 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296

Table 3 Despite having achieved lower quantitative results than the ini-
Results of the reconstruction performed by the ResNet18-based CNN on the test
tial segmentation, the final segmentation is considered successful
set.
for most cases. Our reconstruction was able to include in the seg-
Segmentation Sensitivity Specificity Accuracy Dice Jaccard mentation mask several lung regions that were overlap by abnor-
Initial 93.54% 98.72% 97.57% 94.41% 89.64% mal structures in the original CXR images. However, as described
Final 97.54% 96.79% 96.97% 93.56% 88.07% in Section 3.4, the disadvantage of this step is that it erodes the
segmentation contour in a certain level. As a result, we lose part of
the true positive pixels on the borders. In other words, part of the
4.2.3. Final segmentation results lung regions correctly established by the initial segmentation is
As detailed in Section 3.3, our second CNN model is used to lost on the reconstruction step, typically the regions of the border.
reconstruct the lung regions affected by dense abnormalities that In addition, there are many cases that the initial segmentation
were not included in the initial segmentation. These abnormalities provides extremely precise contours, as illustrated in Fig. 10.
are frequently caused by diseases such as tuberculosis, that overlap In such cases, the reconstruction can not provide any further
the lung fields with non-lung texture patterns. Differently from the enhancement to the initial segmentation, and therefore only de-
first CNN model, which takes as input a set of CXR patches, our creases the performance metrics due to the eroding factor. These
second CNN model takes as input a binary image that consists of are the main reasons for the results of the final segmentation
an initial segmentation mask. Therefore, the training and validation being lower than the initial segmentation for some metrics.
sets used to train our second CNN consists of the initial segmen-
tation masks generated from the training and validation sets. 4.3. Case studies
Using a batch size of 3, our ResNet18-based CNN is trained for
100 epochs, in which we use stochastic gradient descent as opti- In this section we present a set of case studies with the purpose
mizer with learning rate of 0.5, momentum of 0.9, and decay of of analyzing the influence of the reconstruction step in the pro-
0.01. The loss function consists of two binary cross entropy terms. posed method. Such cases illustrate the most common situations
One is used to analyze the background, and the other to analyze of the application of our lung segmentation method on the MC
the actual lung field region, obtained from the ground truth. dataset. We show and discuss the following cases: (1) reconstruc-
Based on approach used in [45], for each loss term is assigned a tion was not necessary, but was able to keep a good result; (2)
weight, with the purpose of making distinction of segmentation reconstruction was not necessary and notably affected the quality
mistakes inside and outside the lung field region. The weights of of the initial segmentation; and (3) reconstruction was necessary
the background and lung field loss functions are, respectively, 1.5 and was able to improve the quality of the segmentation.
and 1. The final loss function consists of the summation of the
two weighted loss terms. After performing several training exper-
iments, this was set of parameters that obtained the best results. 4.3.1. Case study 1 - MCUCXR_0100_0
As detailed in Section 3.4 the final output of our method is the The first case study shows the application of the proposed
combination of the initial segmentation and the reconstruction method on the MCUCXR_0100_0 exam. In this case, the initial
masks. Table 3 presents the performance overview of our proposed segmentation alone was capable of determining very precise
method, showing the results for each of the segmentation steps. lung contours, as illustrated in Fig. 12A, where the contour in
Analyzing the results on Table 3, we can observe that the final red indicates the initial segmentation and the contour in green
segmentation reached 97.54% of sensitivity, which represents a 4% indicates the ground truth. The initial segmentation obtained
increase, compared to the initial segmentation, to the detriment 99.12% of sensitivity, 99.18% of specificity and 98.10% of Dice, while
of an average decrease of 1% on the other metrics. This significant the final segmentation 99.46% of sensitivity, 98.28% of specificity
increase of sensitivity is due to the many lung regions that the and 96.73% of Dice. In spite of the reconstruction step being
reconstruction was able to include in the final segmentation. It is considered unnecessary in this exam, its application on the initial
worth mentioning that the sensitivity rate is extremely important segmentation mask did not worsened the results in any significant
for our method because it represents the rate of lung pixels aspect. As a matter of fact, this case is considered a success, since
correctly segmented, which is directly related to the objective of it still maintained a great correspondence with the marking of
our work. Fig. 11 shows the results of the final segmentation for the specialist, as showed in Fig. 12B (red contour indicates final
two cases. segmentation and green contour indicates ground truth).

Fig. 11. Final segmentation examples. (A) Original CXR image, (B) initial segmentation, (C) reconstruction, (D) final segmentation, and (E) ground truth.
 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296 293

Fig. 12. Case study 1 - MCUCXR_0100_0. (A) Initial segmentation (red) vs ground Fig. 14. Case study 3 - MCUCXR_0108_1. (A) Initial segmentation (red) vs ground
truth (green), and (B) final segmentation (red) vs ground truth (green). truth (green), and (B) final segmentation (red) vs ground truth (green).

of Dice. This illustrates another very common situation in the ap-

plication of our proposed method, which is the reconstruction step
being able to improve the lung contours of the initial segmenta-
tion. For this reason, overall qualitative results were improved and
sensitivity rates on the full test set were considerably increased.
As mentioned in Section 4.2.3, while the initial segmentation
obtained an average sensitivity of 93.54%, the final reconstruction
reached 97.54%, representing an increase of 4%. This shows the
importance of the reconstruction step in the proposed method.

4.4. Comparison with related works

In this section, we present a comparative analysis of the

Fig. 13. Case study 2 - MCUCXR_0102_0. (A) Initial segmentation (red) vs ground
truth (green), and (B) final segmentation (red) vs ground truth (green).
proposed method and the related works mentioned in Section 2.
Zheng et al. [25] proposed a semi-automatic method that uses
anatomical information provided by exam to determine initial bor-
4.3.2. Case study 2 - MCUCXR_0102_0 ders for the lung fields. The method was tested in a private dataset
The second case study shows the application of our lung of 40 exams, and obtained an accuracy of 95%. Despite achieving
segmentation method on the MCUCXR_0102_0 exam. In Fig. 13A a high accuracy rate, this method is semi-automatic. Our method
we can observe that the initial segmentation once again was able is completely automatic, and there is no need for anatomical
to determine very precise lung boundaries (red contour indicates information, in addition the accuracy achieved was almost 97%.
initial segmentation and green contour indicates ground truth). A method based on pixel classification was proposed by An-
However, after the final segmentation step, one can observe that nangi et al. [26]. The method uses edge information to guide an
the result was compromised with several false positives, as illus- active contouring technique. The method obtained an Dice index
trated in Fig. 13B (red contour indicates final segmentation and of 88%. It is observed that this method used techniques based on
green contour indicates ground truth). The initial segmentation active contours, this type of technique requires a starting point
obtained 98.68% of sensitivity, 99.12% of specificity and 98.23% of from where the algorithm will shape the contours until finding
Dice, while the final segmentation obtained 99.73% of sensitivity, the ideal contour based on a stop condition. In our method, we
94.77% of specificity and 93.64% of Dice. Despite occurring most propose the use of two convolutional neural networks to achieve
frequently in smaller scales, it still represents one of the most optimal segmentation for the lungs, thus, we insert classifiers
common cases of the application of the proposed method in the capable of segmentation without needing a starting point. In
MC dataset. It is also the main reason for the final segmentation addition, our method reached a Dice of 93.56%.
not being able to improve the performance rates for most metrics. Deformable models are proposed by Xu et al. [28] for seg-
We consider this case a fail, since it significantly affected the qual- mentation of lungs in CXR. The authors proposed a new active
ity of the initial segmentation, which had previously determined shape model called ERF-ASM. Even bypassing the limitation of
lung contours very similar to the ground truth. addressing an initiation criterion close to the ideal, this method
still needs an initial segmentation to model the final contour of
4.3.3. Case study 3 - MCUCXR_0108_1 the lungs. This method achieved an accuracy of 95%. The method
The last case study details the application of the proposed proposed by our work does not require an initialization criterion
method on the MCUCXR_0108_1 exam. From Fig. 14A we can see because it uses two convolutional neural networks, the first one
that the initial segmentation is missing a considerable part of the responsible for the initial segmentation and the second to refine
lung field in the right side of the exam (red contour indicates the mask obtained initially. Even so, our method overcomes the
initial segmentation and green contour indicates ground truth). work of Xu et al. [28] in all validation metrics.
However, the reconstruction step is able to improve the segmen- Another work that uses deformable models is proposed in
tation by recovering most of the missing lung regions, as shown Candemir et al. [22]. This work uses non-rigid registration meth-
in Fig. 14B (red contour indicates final segmentation and green ods followed by adaptive models for detection of lung edges. By
contour indicates ground truth). In this particular case, it was ob- using non-rigid registration techniques, the method ultimately
tained in the initial segmentation 88.99% of sensitivity, 99.32% of requires an initial image to which the entire base will be regis-
specificity and 93.06% of Dice, and after the final segmentation the tered, conditioning the entire performance of the method to this
results were 97.97% of sensitivity, 98.01% of specificity and 95.72% registration. For this reason, we chose not to use any registrations
294 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296

in our method, making it fully automatic and independent of that can be compromised by regions of anomalies such as turber-
a fixed image. While the registration-based work of Candemir culosis. We also emphasize that our method reached validation
et al. [22] reported a Dice of 96% using the MC dataset, our fully metrics superior to most of the works of the literature, showing
automatic method achieved 93.56% of the same index. how promising and robust our method is. With this, we can
The method proposed by Van Ginneken et al. [12] proposed a demonstrate the feasibility of using our method for the task of
hybrid technique using rule-based segmentation and pixel classi- lung segmentation in CXRs.
fication. As we talk about the other works, these techniques are
directed, and use information of intensity of pixels. When inserting 5. Discussion
a classifier based on CNN, we inserted a certain intelligence to our
method, making it more robust in the segmentation of lung. The proposed method has shown great robustness in the task
The method proposed by Hooda et al. [37] presents a fully- of lung segmentation in CXR. The main steps of a CAD system
convolutional network that performs segmentation of lungs. The could be observed in this work, namely image acquisition, image
work presents a high value of accuracy and Jaccard on the JSRT segmentation, and false positive reduction (post-processing). By
dataset. The JSRT dataset, unlike MC, does not have dense ab- analyzing the development and the evaluation of our work, we
normalities such as tuberculosis, only nodules. Our method when can highlight the following aspects:
applied to the MC base using a reconstruction step to treat the
anomaly problem presents very promising results. 1. It consists of a fully automated method that deals with the
Some works that use encoder-decoder networks are proposed problem of dense abnormalities in lung segmentation. The
by Kalinovsky and Kovalev [35], Saidy and Lee [36], and Mittal difficulty of this task is widely recognized in our research
et al. [38] reaching very promising metrics in the task of lung area, and for that reason most works achieve worse results
segmentation. Both works present the images with marking to in this type of dataset.
the network and it is in charge to learn the features and in the 2. The use of a public dataset is another positive factor in
future to classify new images in regions of lung and non-lung. this work, since it enables other authors to reproduce the
An advantage of our method in relation to these is that we insert experiments and establish comparative analysis. In our work
a reconstruction step, which is concerned with the existence of we used the Montgomery County Dataset, which is a public
possible anomalies that make the segmentation task more difficult. database of chest X-rays that contains some patients with
Concerned about this, we ensure that our method is more robust manifestations of tuberculosis. It is worth mentioning that
in the task of CXR segmentation of patients with abnormalities, our results in this database are very promising, and for
such as tuberculosis. this reason the proposed method has achieved a prominent
Finally, a hybrid method is presented by Ngo and Carneiro place among other important works.
[29], which used deep learning techniques along with deformable 3. Although our work focuses on reconstruction for the seg-
models. By doing this, the author inserts two relevant techniques mentation of abnormal lungs, it is worth mentioning that
for lung segmentation in his method, the first capable of prior our initial segmentation alone is capable to provide excellent
segmentation and the second capable of refining such segmenta- results comparable to other relevant works in the literature.
tion. The authors achieved 96.5% accuracy as a result. It is believed 4. In the proposed method lung segmentation is a fully au-
that by using a reconstruction network to refine an initial segmen- tomated process performed entirely by two CNN models,
tation, we make the method more robust and invariant to initial which are very robust tools that perform implicit feature
parameters. Also, our method was able to achieve 96.97% accuracy. extraction and selection. This is a positive aspect because
Table 4 shows the relationship between these works and the it eliminates the need for empirically determining the set
proposed method. of features to be used in the learning process, and the
As we can see in the Table 4, there are several methods in the techniques to be used in feature selection.
literature that investigate lung segmentation in CXR. Also, it is 5. Despite the fact that our initial segmentation achieve ex-
seen that over the years the techniques have become increasingly pressive results, we used another CNN model to enhance
robust and have always achieved results higher than 90% in all the segmentation for more complex cases. We address the
validation metrics. This shows how fundamental is the search for problem of lung segmentation with dense abnormalities by
increasingly automatic computational methodologies in the task of performing a reconstruction step, and to the best of our
lung segmentation. knowledge, this is the first method that addresses these
However, we emphasize that our method is fully automatic, cases explicitly.
which uses two convolutional neural networks, capable of not 6. The reconstruction step is performed by a ResNet-18-based
only segmenting the lungs, but also reconstructing the marking CNN, which is a deep neural network architecture that uses

Table 4
Comparison of results of the proposed method and related works.

Work Dataset Disease Technique Sen Acc Dice Jaccard

Zheng et al. [25] Private No Semi-automatic – 95% – –

Annangi et al. [26] Private No Semi-automatic – – 88% –
Xu et al. [28] Private No Semi-automatic 91% 95% – –
Candemir et al. [22] JSRT No Semi-automatic – – 96.7% 95.4%
MC Yes – – 96% 94.1%
Van Ginneken et al. [12] Private No Semi-automatic – 94% – –
Kalinovsky and Kovalev [35] Private No Automatic – 96.2% – –
Saidy and Lee [36] JSRT No Automatic 95% – 96% –
Ngo and Carneiro [29] JSRT No Semi-automatic – 96.5% – –
Hooda et al. [37] JSRT No Automatic – 98.92% – 95.88%
Mittal et al. [38] JSRT+MC No Automatic – 98.73% – 95.10%
Proposed Method MC Yes Automatic 97.54% 96.97% 93.56% 88.07%
 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296 295

the concept of residual learning to accelerate and improve CNN model, which is based on AlexNet, to perform patch classi-
the learning process, providing better segmentation results. fication. Then, a second CNN model, which is based on ResNet18,
7. We propose a fully automatic method that performs not only is used to perform a reconstruction of missing parts of the lung
the regular lung segmentation, but also that provides a solu- field. Finally, the output of the proposed method is obtained by
tion to the more complex cases with dense abnormalities in the combination of both initial segmentation and reconstruction.
the lung fields, consequently generating expressive results. The proposed method is validated on a public database,
8. Finally, the combination of all techniques in this work namely Montgomery County Chest X-ray dataset, which con-
provided better segmentation of lungs with dense abnor- tains CXR exams from normal patients but also from several
malities, achieving an accuracy of 96.97%, a sensitivity of patients with manifestations of tuberculosis. For this reason,
97.54%, a specificity 96.79%, a Dice index of 93.56%, and a the MC dataset is very heterogeneous and complex, which re-
Jaccard of 88.07%. sults in a challenging task that most likely illustrates realistic
scenarios.
Despite having several positive factors, the proposed method
also has some limitations, in which we highlight:
6.1. Work evaluation
1. The MC Dataset is a very complex and heterogeneous
database because it contains exams with a large variety of The use of an AlexNet-based CNN for patch classification was
lung abnormalities, which challenges our proposed method quite robust. In general, it provided outstanding results for an
with more realistic scenarios. However, this dataset contains initial segmentation step, since it was able to determine very
only 138 images, of which 80 are used for training. This precise lung boundaries for most of the normal lungs. Due to the
can be considered a small number of samples for training extreme texture difference, one limitation of this step is that it
deep learning models if compared to the number of images was not possible to segment lung regions overlapped by dense
typically used in image recognition models. In the widely lesions. However, this problem is addressed in the further step of
known ImageNet Large Scale Visual Recognition Challenges reconstruction.
(ILSVRC), for example, deep learning models have been Despite not being able to improve some of the performance
trained with over 1.2 million images since 2012. Therefore, metrics, the reconstruction step was of great importance in the
we believe we may not have fully explored the potential of proposed method. It helped overcome the limitation of the pre-
the techniques described in this work, due to limitations of vious step, which resulted in better qualitative results. Though, it
data. is worth mentioning that the sensitivity rates were considerably
2. Another limitation is the fact that the reconstruction does improved by this step.
not ensure a quantitative improvement of the segmentation. The combination of the techniques mentioned in this paper
Despite being capable of providing a segmentation that was able to achieve great results. We believe that, in general, this
includes the regions affected by dense abnormalities, the work represents a great contribution to the scientific community.
reconstruction can also include false positives, and remove In spite of all its limitations, it has shown great robustness for
true positives, and therefore decrease the quantitative the fact that even on such a diverse and complex dataset it
results. has achieved a prominent place among the best works in the
3. The results obtained by the proposed method can be con- literature.
sidered excellent, being among the bests in the literature. In addition, the way techniques were used and adapted to per-
However, the reconstruction step decreases some of the form lung segmentation is very innovative and may inspire other
performance metrics. We believe this limitation could be authors in new coming works. It is worth mentioning that most
mitigated or even overcame by exploring the use of contour computer-aided detection systems rely on a robust segmentation
enhancement techniques. method. We therefore believe that the proposed method is very
4. Finally, it is worth mentioning that deep learning models, promising and relevant.
such as the CNNs used in our work, typically have a great
number of parameters. Frequently, this set of parameters is 6.2. Future work
empirically determined by the user, which may not provide
optimal results. This specific situation is another limitation Despite having obtained promising results, the proposed
of the proposed method. method can be improved in some of aspects. Following, we
present some of the current limitations of our work that could be
All the above-mentioned observations aggregate value to this overcome on future work.
work. The many positive aspects listed in this section have en- The main limitation of this work that should be addressed is
abled the proposed method to achieve excellent results. It is worth the fact that the reconstruction does not always guarantee a quan-
mentioning that, as far as we know, the literature does not have titative improvement of the segmentation. Since false positives can
any lung segmentation methodology that provides a solution to be incorrectly included, and true positives removed, we believe
the specific problem of dense abnormalities. We highlight that, that by using an active contour based technique, such as level set,
even with some limitations, the proposed method was capable in order to enhance the reconstructed segmentation.
to achieve outstanding results, and for these reasons, we believe Regardless of the MC dataset being a very complex and het-
we our work provides a very important contribution, for being an erogeneous, it contains only 138 exams, which can be considered
innovative, robust and promising automated segmentation method. a small number of samples, specially for training deep learning
models. Therefore, in order to validate our work, provide a more
6. Conclusion robust lung segmentation method, and enable a better comparison
with the literature, as future work we suggest that other public or
This paper presented a fully automatic method for lung field private CXR databases are used in the experiments.
segmentation in CXR. For this purpose, several techniques are Finally, since CNN models typically have a large number of
used, which include two convolutional neural networks, and other empirically determined parameters, we suggest that optimization
image processing techniques such as morphological operations and methods, such as evolutionary algorithms, should be used to
filtering. In order to generate an initial segmentation, we use one obtain improved CNN architectures for both models.
296 J.C. Souza, J.O. Bandeira Diniz and J.L. Ferreira et al. / Computer Methods and Programs in Biomedicine 177 (2019) 285–296

Conflict of interest [25] Y. Zheng, L. Li, M. Kallergi, W. Qian, R.A. Clark, Improved method for automatic
identification of lung regions in chest radiographs, in: Medical Imaging 20 0 0:
Image Processing, 3979, International Society for Optics and Photonics, 20 0 0,
None. pp. 1138–1147.
[26] P. Annangi, S. Thiruvenkadam, A. Raja, H. Xu, X. Sun, L. Mao, A region based
Acknowledgements active contour method for x-ray lung segmentation using prior shape and low
level features, in: 2010 IEEE international symposium on biomedical imaging:
from nano to macro, IEEE, 2010, pp. 892–895.
The authors acknowledge FAPEMA for the financial support. [27] T. McInerney, D. Terzopoulos, Deformable models in medical image analysis: a
survey, Med. Image Anal. 1 (2) (1996) 91–108.
References [28] T. Xu, M. Mandal, R. Long, I. Cheng, A. Basu, An edge-region force guided ac-
tive shape approach for automatic lung field detection in chest radiographs,
Comput. Med. Imaging Graph. 36 (6) (2012) 452–463.
[1] N.L.S.T.R. Team, Reduced lung-cancer mortality with low-dose computed tomo-
[29] T.A. Ngo, G. Carneiro, Lung segmentation in chest radiographs using dis-
graphic screening, New Engl. J. Med. 365 (5) (2011) 395–409.
tance regularized level set and deep-structured learning and inference, in:
[2] P.P. Rebouças Filho, P.C. Cortez, A.C. da Silva Barros, V.H.C. Albuquerque,
Image Processing (ICIP), 2015 IEEE International Conference on, IEEE, 2015,
J.M.R. Tavares, Novel and powerful 3d adaptive crisp active contour method
pp. 2140–2143.
applied in the segmentation of ct lung images, Med. Image Anal. 35 (2017)
[30] B. Van Ginneken, B.M. ter Haar Romeny, Automatic segmentation of lung fields
503–516.
in chest radiographs, Med. Phys. 27 (10) (20 0 0) 2445–2455.
[3] I.R.S. Valente, P.C. Cortez, E.C. Neto, J.M. Soares, V.H.C. de Albuquerque,
[31] A. Mittal, R. Hooda, S. Sofat, Lung field segmentation in chest radiographs: a
J.M.R. Tavares, Automatic 3d pulmonary nodule detection in ct images: a sur-
historical review, current status, and expectations from deep learning, IET Im-
vey, Comput. Methods Programs Biomed. 124 (2016) 91–107.
age Process. 11 (11) (2017) 937–952.
[4] Y. Gordienko, P. Gang, J. Hui, W. Zeng, Y. Kochura, O. Alienin, O. Rokovyi,
[32] G.L.F. da Silva, T.L.A. Valente, A.C. Silva, A.C. de Paiva, M. Gattass, Convolutional
S. Stirenko, Deep learning with lung segmentation and bone shadow exclusion
neural network-based pso for lung nodule false positive reduction on ct im-
techniques for chest x-ray analysis of lung cancer, in: International Conference
ages, Comput. Methods Programs Biomed. 162 (2018) 109–118.
on Theory and Applications of Fuzzy Systems and Soft Computing, Springer,
[33] P.H.B. Diniz, T.L.A. Valente, J.O.B. Diniz, A.C. Silva, M. Gattass, N. Ventura,
2018, pp. 638–647.
B.C. Muniz, E.L. Gasparetto, Detection of white matter lesion regions in mri
[5] R.H. Daffner, M. Hartman, Clinical Radiology: the Essentials, Lippincott
using slic0 and convolutional neural network, Comput. Methods Programs
Williams & Wilkins, 2013.
Biomed. 167 (2018) 49–63.
[6] C. Qin, D. Yao, Y. Shi, Z. Song, Computer-aided detection in chest radiography
[34] J.O.B. Diniz, P.H.B. Diniz, T.L.A. Valente, A.C. Silva, A.C. Paiva, Spinal cord detec-
based on artificial intelligence: a survey, Biomed. Eng. Online 17 (1) (2018) 113.
tion in planning ct for radiotherapy through adaptive template matching, im-
[7] X-ray (radiography) - chest, 2018, Accessed: 2018-12-16. (https://www.
slic and convolutional neural networks, Comput. Methods Programs Biomed.
radiologyinfo.org/en/info.cfm?pg=chestrad).
170 (2019) 53–67.
[8] N.H.S. England, Diagnostic imaging dataset annual statistical release 2017/18,
[35] A. Kalinovsky, V. Kovalev, Lung image segmentation using deep learning meth-
2018.
ods and convolutional neural networks (2016).
[9] T.R.C. of Radiologists, Clinical radiology uk workforce census 2017 report, 2018.
[36] L. Saidy, C.-C. Lee, Chest x-ray image segmentation using encoder-decoder con-
[10] L.G. Quekel, A.G. Kessels, R. Goei, J.M. van Engelshoven, Miss rate of lung can-
volutional network, in: 2018 IEEE International Conference on Consumer Elec-
cer on the chest radiograph in clinical practice, Chest 115 (3) (1999) 720–724.
tronics-Taiwan (ICCE-TW), IEEE, 2018, pp. 1–2.
[11] L.H. Garland, On the scientific evaluation of diagnostic procedures: presiden-
[37] R. Hooda, A. Mittal, S. Sofat, An efficient variant of fully-convolutional network
tial address thirty-fourth annual meeting of the radiological society of north
for segmenting lung fields from chest radiographs, Wireless Personal Commun.
america, Radiology 52 (3) (1949) 309–328.
101 (3) (2018) 1559–1579.
[12] B. Van Ginneken, B.M.T.H. Romeny, M.A. Viergever, Computer-aided diagnosis
[38] A. Mittal, R. Hooda, S. Sofat, Lf-segnet: a fully convolutional encoder–decoder
in chest radiography: a survey, IEEE Trans. Medical Imaging 20 (12) (2001)
network for segmenting lung fields from chest radiographs, Wireless Personal
1228–1241.
Commun. 101 (1) (2018) 511–529.
[13] Z. Ma, J.M.R. Tavares, R.N. Jorge, A review on the current segmentation algo-
[39] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, nature 521 (7553) (2015) 436.
rithms for medical images, in: Proceedings of the 1st International Conference
[40] Y. LeCun, K. Kavukcuoglu, C. Farabet, et al., Convolutional networks and appli-
on Imaging Theory and Applications (IMAGAPP), 2009.
cations in vision., in: ISCAS, 2010, 2010, pp. 253–256.
[14] Z. Ma, J.M.R. Tavares, R.N. Jorge, T. Mascarenhas, A review of algorithms for
[41] D. Yamamoto, H. Arimura, S. Kakeda, T. Magome, Y. Yamashita, F. Toy-
medical image segmentation and their applications to the female pelvic cavity,
ofuku, M. Ohki, Y. Higashida, Y. Korogi, Computer-aided detection of mul-
Comput. Methods Biomech. Biomed. Eng. 13 (2) (2010) 235–246.
tiple sclerosis lesions in brain magnetic resonance images: false posi-
[15] P.C. Gonçalves, J.M.R. Tavares, R.N. Jorge, Segmentation and simulation of ob-
tive reduction scheme consisted of rule-based, level set method, and sup-
jects represented in images using physical principles (2008).
port vector machine, Comput. Med. Imaging Graph. 34 (5) (2010) 404–
[16] R.B. Oliveira, E. Mercedes Filho, Z. Ma, J.P. Papa, A.S. Pereira, J.M.R. Tavares,
413.
Computational methods for the image segmentation of pigmented skin le-
[42] M.G.F. Eadgahi, H. Pourreza, Localization of hard exudates in retinal fundus
sions: a review, Comput. Methods Programs Biomed. 131 (2016) 127–141.
image by mathematical morphology operations, in: Computer and Knowl-
[17] Z. Ma, J.M.R. Tavares, Effective features to classify skin lesions in dermoscopic
edge Engineering (ICCKE), 2012 2nd International eConference on, IEEE, 2012,
images, Expert Syst. Appl. 84 (2017) 92–101.
pp. 185–189.
[18] A.H. Dallal, C. Agarwal, M.R. Arbabshirani, A. Patel, G. Moore, Automatic esti-
[43] E. Soleymanpour, H.R. Pourreza, et al., Fully automatic lung segmentation and
mation of heart boundaries and cardiothoracic ratio from chest x-ray images,
rib suppression methods to improve nodule detection in chest radiographs, J.
in: Medical Imaging 2017: Computer-Aided Diagnosis, 10134, International So-
Med. Signals Sensors 1 (3) (2011) 191.
ciety for Optics and Photonics, 2017, p. 101340K.
[44] S.G. Armato III, M.L. Giger, H. MacMahon, Automated lung segmentation
[19] X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, R.M. Summers, Chestx-ray8:
in digitized posteroanterior chest radiographs, Acad. Radiol. 5 (4) (1998)
hospital-scale chest x-ray database and benchmarks on weakly-supervised
245–255.
classification and localization of common thorax diseases, in: Computer Vi-
[45] K. Ehsani, R. Mottaghi, A. Farhadi, Segan: segmenting and generating the in-
sion and Pattern Recognition (CVPR), 2017 IEEE Conference on, IEEE, 2017,
visible, arXiv preprint (2017).
pp. 3462–3471.
[46] S. Ioffe, C. Szegedy, Batch normalization: accelerating deep network training
[20] E. Skoura, A. Zumla, J. Bomanji, Imaging in tuberculosis, Int. J. Infect. Dis. 32
by reducing internal covariate shift, arXiv:1502.03167 (2015).
(2015) 87–93.
[47] F. Chollet, et al., Keras, 2015, (https://keras.io).
[21] The radiology assistant: chest x-ray - lung disease, 2018, Accessed: 2019-02-
[48] M. Abadi, et al., TensorFlow: Large-scale machine learning on heterogeneous
06. (http://www.radiologyassistant.nl/en/p50d95b0ab4b90/chest- x- ray- lung-
systems, 2015, (https://www.tensorflow.org/). Software available from tensor-
disease.html).
flow.org.
[22] S. Candemir, S. Jaeger, K. Palaniappan, J.P. Musco, R.K. Singh, Z. Xue, A. Karar-
[49] M. Bland, An Introduction to Medical Statistics, Oxford University Press, Ox-
gyris, S. Antani, G. Thoma, C.J. McDonald, Lung segmentation in chest radio-
ford, UK, 2015.
graphs using anatomical atlases with nonrigid registration, IEEE Trans. Med.
[50] L.-Y. Deng, The cross-entropy method: a unified approach to combinatorial op-
Imaging 33 (2) (2014) 577–590.
timization, Monte-Carlo simulation, and machine learning, 2006.
[23] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep con-
[51] X. Glorot, Y. Bengio, Understanding the difficulty of training deep feedforward
volutional neural networks, in: Advances in neural information processing sys-
neural networks, in: Proceedings of the thirteenth international conference on
tems, 2012, pp. 1097–1105.
artificial intelligence and statistics, 2010, pp. 249–256.
[24] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition,
in: Proceedings of the IEEE conference on computer vision and pattern recog-
nition, 2016, pp. 770–778.