Академический Документы
Профессиональный Документы
Культура Документы
Abstract— Detection of pulmonary nodules is a frequent improve the radiologists’ performance of in screening and
circumstance in daily radiology practice. Their extraction diagnosis.
depends on their localization. Sub-pleural nodules are directly As lung nodules are characterized by their densities,
connected to the border of the pleura and have consequently location, size and inner content, segmentation and detection of
open contours. Their extraction is challenging and a computer-
nodules have become a major challenge. Several works have
aided diagnosis system is, hence, indispensable. In this paper, we
propose an automatic segmentation approach of sub-pleural lung been developed from CT images to overcome this challenge
nodules from Computed Tomography (CT) scans based on such as those based on active contour [5] [6] [7], region
morphological operations. This method is divided into three growing [8], watershed algorithm [9] [10]… The majority of
steps: pre-processing, initial detection of sub-pleural lung nodule them rely on highly advanced segmentation and learning
and post-processing. First, a region of interest containing the methods.
nodule is extracted and converted using an adaptive thresholding Among these methods, we find the mathematical
algorithm. Second, morphological operations are used to create a morphology based approach. It is an image processing
mask for the lung lobe and segment sub-plural nodule. Finally, technique based on mathematical rules and computer theories
small structures connected to the border of the segmented image
used to analyze geometric structures and topological shapes.
are removed and final nodule regions are detected. The proposed
method is evaluated on 40 CT scan images (17 in axial acquisition Most researchers apply morphological operators to improve
and 23 in coronal reconstruction) and gives a good rate of the input image quality as a pre-processing step or to touch up
accuracy that proves the effectiveness of our approach. their segmentation results as a post-processing step [11] [12]
[13] [14]. Only a few researchers have thought of using
Keyword—Sub-pleural lung nodule, Detection, Segmentation,
morphological operators as a basic method for lung nodules
Morphological operations.
segmentation from CT images.
I. INTRODUCTION In this paper, we propose an automated segmentation
method for sub-pleural lung nodules detected on 2D CT
The lung nodule is a small growth of soft tissue (less than
images using a morphological operation-based approach. The
30 mm) in the pulmonary parenchyma [1]. It presents an
proposed method consists mainly with three steps that will be
important radiological indication for early diagnosis of lung
described in section three. Section 4 presents the experimental
cancer. Lung cancer accounts for about 20% of all nodules'
results. Finally, conclusion and future work are summarized in
cases [2]. Early identification of cancer in lung nodules is
necessary to increase the 5-year survival rate until 70% [3]. section 5.
According to the Early Lung Cancer Action Project (ELCAP)
II. RELATED WORKS
database [4], the malignancy of lung nodules varies between
1% and 80% for nodules with a diameter varying between S.Biradar et al. [15] proposed a segmentation approach
]5mm, 20mm[. based on mathematical morphology to detect circumscribed
Computed tomography (CT) is the gold standard for lung cancer in the early stages from 2D CT images. After
detecting and characterizing small pulmonary nodules as well enhancing acquired image by median filter, the authors
as for diagnosing and monitoring their progress. However, in applied Otsu method, as a preprocessing step, to convert input
some cases, radiologist can miss small nodules especially grayscale images to binary images. Then, they applied
when their diameters are less than 5 mm. Hence, a computer- morphological opening operation using three different
aided detection (CAD) system is essential. The CAD aims to structuring elements as a segmentation process in order to
solve some difficulties in the detection of lung nodules and to extract the nodular region features such as shape,
boundaries… Finally, they applied inversion and clear border
operations, as a post-processing step, to get the segmented suspicion of nodules' malignancy from CT images. The
image. proposed method was able to integrate nodule characteristics
J.Kuruvilla et al. [16] proposed to develop a CAD system into a hierarchical network without segmentation procedure,
based on mathematical morphology operations for the which simplifies the conventional classification.
detection of lung cancers from 2D CT images. The authors Sun et al. [20] implemented three deep learning methods :
proposed four algorithms to segment four types of nodules. CNN, deep belief network (DBN), and SDAE in order to
The lung nodules were classified according to their location in analyze its ability to extract automatically generated features
pulmonary parenchyma. First, they segmented lung lobes in the diagnosis of CT images of pulmonary nodules, and to
using Otsu thresholding and morphological opening to remove compare its performance with traditional computer aided
small objects from the foreground. Morphological opening diagnosis (CADx) systems. The results showed that deep
was applied using a new structuring element which was algorithms achieve desirable performance in the diagnosis of
created by a three different structuring elements combination. lung nodules with a preference from CNN to DBN and SDAE.
Then they reversed the image and performed a clear border
operation to remove small structures connected to the border. III. MATERIAL AND METHODS
For juxta-pleural nodule segmentation, the authors proposed to The sub-pleural nodules are attached to the pleura's
split the segmented image from the first algorithm and flipped surface. Unlike the isolated nodule, which is easy to segment,
the right half of the lung into the left side. Then, in each row sub-pleural nodule presents a challenge for researchers due to
of the image, they noted the first '1', calculate and store the the need to find a boundary between nodule and chest wall
difference between its adjacent values as a vector. The starting because of their similar pixel intensity. The methodology of
and the ending point is connected and filled. Finally, they the proposed work is shown in the block diagram in Fig.1.
combined the two halves again to obtain the cancerous part. In the following, we present the details of each step.
To evaluate the proposed method's performance, the obtained
sensitivity, specificity and accuracy for the segmented images A. CT lung acquisition
are respectively 88.24%, 93.33% and 90.63%. . For the evaluation, we used a dataset of lung CT scans from
Mokhled S. et al. [17] proposed a morphological Radiology department of Abderrahmen Mami Hospital –
segmentation and Gabor filter approaches for lung cancer Ariana - Tunisa. The database contained 40 CT images in sharp
detection from 2D CT images. First, the authors used Gabor filer and lung windowing of a sub-pleural nodule (17 in axial
filter in order to enhance the input image and to improve its acquisition and 23 in coronal reconstruction). We implemented
information’s interpretability. Second, they extracted the lung our algorithm in MATLAB (R2017a) programming language
area as ROI using diffusion and morphological erosion. based on its Mathworks bibliography.
Thirdly, after filling holes and removing small objects from
the extracted ROI, the authors applied four step to locate lung B. Pre-processing
nodule and its boundaries. (1) They calculated an overall
threshold level. (2) They converted image to binary based on As input image, we considered a region of interest (ROI)
threshold level. (3) They eroded twice the binary image by the containing the nodule. Then, we converted the grayscale
same structure elements. (4) They found perimeter of nodule image into a binary image by adaptive thresholding based on
and lung lobes in eroded image. Finally, they imposed Otsu method (Fig.2 (b)).
perimeter into ROI image and the selected objects are C. Nodule localization
detected.
Recently, deep learning methods have gained a lot of attention In this step, mathematical morphology is used to isolate
by researchers because of their promising results in solving lung nodules. . First, we filled the lung lobe holes to remove
complex learning problems. These methods have widely small structures corresponding to blood vessels and
proven their performance in various disciplines especially in bronchioles. A hole is a background region surrounded by a
medical image analysis field. Several deep learning connected border of foreground elements.
architectures such as artificial massive neural networks Let A denotes a binary image that enclosed a background
(MTANN) [18], convolutional neural networks (CNN) region (holes), which have to be filled and bm denotes a
[18,20], stacked denoising autoencoder (DSAE) [20,21], marker image which is settled 1-A in the image border, and 0
Multi-crop Convolutional Neural Networks [19], ... were used otherwise:
to detect and classify lung nodules from CT images. 1− A (x, y) if (x,y) is on the border of A½
bm ( x, y) = ® ¾ (1)
In [18], the authors used two end-to-end learning ¯0 otherwise ¿
architectures that are massive formation artificial neural Then
networks (MTANN) and convolutional neural networks c
(CNN) to detect and distinguish benign and malignant lung B = ª¬ R A c ( bm ) º¼ (2)
nodules from CT images. Their results indicated that
MTANN's performance was significantly higher than CNN's. B has the effect to filling all the holes in A. These holes
In [19], an end-to-end deep learning architecture of Multi- correspond to blood vessels and pulmonary bronchioles.
crop Convolutional Neural Networks was proposed to classify
85
2018 IEEE 4th Middle East Conference on Biomedical Engineering (MECBME)
Input CT image
Selection of
ROI
Nodule Otsu
detection Thresholding
Morphological closing
operation
Mathematical
subtraction operation
Fig.1. Block diagram of proposed method: (a) Pre-processing steps (b) Localization steps (c) Post-processing step
86
2018 IEEE 4th Middle East Conference on Biomedical Engineering (MECBME)
Where:
• True positive (TP): Predicts an area as a real part of
the nodule.
• True negative (TN): Predicts an extra-nodular area
as a real extra nodular one.
• False negative (FN): Predicts a nodular area as an
extra-nodular one.
Fig.2. Localization steps: (a) creating a lung mask by morphological • False negative (FP): Predicts an extra-nodular area
closing; (b) pre-processed image; (c) lung nodule detected after subtraction
as a nodular one.
operation
87
2018 IEEE 4th Middle East Conference on Biomedical Engineering (MECBME)
The proposed method showed a significant gain of run mathematical morphology to detect nodules [14] [15] [16],
time of 9.53 seconds as well as its stability and ability to our proposed method can retain the actual shape of the
properly detect sub-pleural nodules obtaining an average Fm nodule (Fig.3 (b)) which provides information about the
over 82 %. cancer evolution, its malignancy degree and its stage. In
most works, currently detected nodules are spherical or
The ROC curve for the proposed system is shown in ellipsoidal which reflects the structuring element effect
Fig.4 and represents an AUC estimated at 0.766. From during the application of the morphological operation.
preliminary results, our algorithm can be implemented in
clinical applications to help physicians in the detection and So we can conclude that our proposed algorithm is much
diagnosis of lung nodules. more efficient to count good perspective compared to others
methods. The proposal of a good method for detection of
Fig.5 and Fig.6 present some examples of the lung nodule from CT image based only on mathematical
segmentation results in axial and in coronal acquisitions morphology is mainly based on a good segmentation
respectively. Here, we show the efficiency of our method for algorithm and that is what we are trying to prove by the
detecting sub-pleural nodules from CT image in axial and proposed approach.
coronal acquisitions. Unlike other methods that have used
Fig.5: Examples of segmentation results of the CT images in axial acquisitions: (a) Original image; (b) Selection of ROI;
(c) Binarization by adaptive thresholding; (d) Filling holes; (e) lung mask creation; (f) nodules detection
Fig.6: Examples of segmentation results of the CT images in coronal acquisitions: (a) Original image; (b) Selection of ROI;
(c) Binarization by adaptive thresholding; (d) Filling holes; (e) lung mask creation; (f) nodules detection
88
2018 IEEE 4th Middle East Conference on Biomedical Engineering (MECBME)
[13] Gupta Anindya, et al. "Methods for increased sensitivity and scope in
V. CONCLUSION automatic segmentation and detection of lung nodules in CT
images.", IEEE- ISSPIT, 2015.
In this paper, a fast automatic segmentation method has [14] Zhao Yu-qian, et al. "Medical images edge detection based on
been proposed to detect sub-pleural lung nodule based on mathematical morphology", IEEE-EMBS, 2006
morphological operators approach. This work aims to solve [15] Biradar, et al., Agalatakatti. "Lung Cancer Identification Using CT
the problem of open contour since sub-pleural nodules are Images." International Journal Of Engineering And Computer
directly connected to the boarder of pleural sheet. Moreover, Science 4 (2015): 13022-13025.
our method retain the concavities and the original shapes of [16] Jinsa Kuruvilla, Dr.K.Gunavathi, "Detection of Lung Cancer using
Morphological Operations", IJSERVolume 4, Issue 8, August-2013.
these nodules which give an important information of their
[17] Mokhled S. et al.., "Lung Cancer Detection using Morphological
malignity degree. Segmentation and Gabor Filtration Approaches", IJERT, Vol. 3 Issue
The proposed method was tested on a database of 40 CT 7, July – 2014.
scans with sub-pleural nodules, from radiology department [18] N. Tajbakhsh and K. Suzuki, “Comparing two classes of end-to-end
of the Abderrahmen Mami Hospital – Ariana - Tunisia. The machine-learning models in lung nodule detection and classification:
Mtanns vs. cnns,” Pattern Recognit., vol. 63, pp. 476–486, 2017.
proposed method shown many advantages such as low
[19] W. Shen, et al., “Multi-crop convolutional neural networks for lung
computational cost, robustness, easy implementation, nodule malignancy suspiciousness classification,” Pattern Recognit.,
accuracy, stability and its ability to overcome the limitations vol. 61, pp. 663–673, 2017.
of some advanced approaches. It can reliably be used as an [20] W. Sun, et al., “Automatic feature learning using multichannel ROI
additional tool to assist radiologists in the cancer diagnosis based on deep structured algorithms for computerized lung cancer
process. diagnosis”. Computers in biology and medicine, vol. 89, p. 530-539,
2017
In the future, parameters that characterizes the lung
[21] J. Cheng, et al., “Computer-aided diagnosis with deep learning
nodules will be selected to classify them as benign and architecture: Applications to breast lesions in us images and
malignant. In addition, we will correlate these different pulmonary nodules in ct scans,” Sci. Rep., vol. 6, p. 24454, 2016.
parameters to other histological types of lung tumors.
REFERENCES
[1] Mayo, John R., and Stephen Lam. "Computed tomography and the
secrets of lung nodules." Canadian Association of Radiologists
Journal, (2015)
[2] Yuan, et al. "Hybrid-feature-guided lung nodule type classification on
CT images. " Computers & Graphics, (2017).
[3] Camarlinghi, Niccolò. "Automatic detection of lung nodules in
computed tomography images: training and validation of algorithms
using public research databases." The European Physical Journal Plus
128 (2013): 1-21.
[4] Valente Igor Rafael, et al.,"Automatic 3D pulmonary nodule
detection in CT images: a survey." Computer methods and programs
in biomedicine 124 (2016): 91-107.
[5] Farag Amal A., et al., "A novel approach for lung nodules
segmentation in chest CT using level sets", IEEE Trans on Image
Processing 22.12 (2013): 5202-5213.
[6] Sun Shanhui, et al., "Automated 3-D segmentation of lungs with lung
cancer in CT data using a novel robust active shape model
approach." IEEE trans on medical imaging 31.2 (2012): 449-460.
[7] Rebouças Filho, et al. "Novel and powerful 3D adaptive crisp active
contour method applied in the segmentation of CT lung
images." Medical image analysis 35 (2017): 503-516.
[8] Emaminejad, et al. "Fusion of quantitative image and genomic
biomarkers to improve prognosis assessment of early stage lung
cancer patients." IEEE Trans on Biomedical Engineering 63.5 (2016):
1034-1043.
[9] Avinash.S, et al., "An improved image processing analysis for the
detection of lung cancer using Gabor filters and watershed
segmentation technique.", IEEE- ICICT Vol. 3, 2016.
[10] Chaudhary Anita, et al., "Lung cancer detection on CT images by
using image processing.", IEEE- ICCS, 2012.
[11] Gangireddy, et al., "A Novel Approach for Lung Cancer Detection
Using Morphological Operations", International Journal of Innovative
Research in Computer and Communication Engineering, Vol. 4, Issue
12, 2016.
[12] C. Karthikeyan and B. Ramadoss, "Segmentation Algorithm for CT
Images Using Morphological Operation and Artificial Neural
Network", International Journal of Computer Theory and
Engineering, Vol. 3, No. 4, August 2011.
89