A Web-based Computer Aided Detection System for Automated

Search of Lung Nodules in Thoracic Computed Tomography Scans

M. E. Fantacci

1,2

, S. Bagnasco

, N. Camarlinghi

, E. Fiorina

3,4

, E. Lopez Torres

3,5

, F. Pennazio

3,4

C. Peroni

3,4

, A. Retico

, M. Saletta

, C. Sottocornola

1,2

, A. Traverso

3,6

and P. Cerello

Physics Department, Pisa University, Largo Pontecorvo 3, Pisa, Italy

Pisa Section of INFN, Pisa, Italy

Torino Section of INFN, Torino, Italy

Physics Department, Torino University, Torino, Italy

CEADEN, Havana, Cuba

Politecnico di Torino, Torino, Italy

Keywords: Computer Aided Detection, Lung Nodules, Thoracic Computed Tomography.

Abstract: M5L, a Web-based fully automated Computer-Aided Detection (CAD) system for the automated detection

of lung nodules in thoracic Computed Tomography (CT), is based on a multi-thread analysis with two

independent CAD subsystems, the lung Channeler Ant Model (lungCAM) and the Voxel-Based Neural

Analysis (VBNA), and on the combination of their results. The lungCAM subsystem is based on a model of

the capabilities that ants show in nature in finding structures, defining shapes and acting according with

local information. The VBNA subsystem is based on a multi-scale filter for spherical structures in searching

internal nodules and on the analysis of the intersections of surface normals in searching pleural nodules. The

M5L performance, extensively validated on 1043 CT scans from 3 independent datasets, including the full

LIDC/IDRI database, is homogeneous across the databases: the sensitivity is about 0.8 at 6-8 False Positive

findings per scan, despite the different annotation criteria and acquisition and reconstruction conditions. A

prototype service based on M5L is hosted on a server operated by INFN in Torino. Preliminary validation

tests of the system have recently started in several Italian radiological institutes.

1 INTRODUCTION

Lung cancer is one of the main public health issues

in developed countries, accounting for about 19%

and 28% of cancer-related deaths in Europe (Parkin,

2010) and the United States of America (American

Cancer Society, 2009), respectively, with a 5-year

survival rate of only 10–16% (Jemal, 2010). Lung

cancer most commonly manifests itself as non-

calcified pulmonary nodules. Computed

Tomography (CT) has been shown to be the most

sensitive imaging modality for the detection of small

pulmonary nodules: therefore low dose high

resolution CT-based screening trials are regarded as

a promising technique for detecting early-stage lung

cancers (Henschke, 1999). Recent results obtained

by the National Lung Screening Trial (NLST),

involving 53454 high-risk patients, show a 20%

reduction of mortality when the screening program

was carried out with the helical CT, rather than with

a conventional chest X-ray (NLST, 2011). The

design and operation of large scale lung cancer

screening programs is now being considered, with

the goal of maximizing their effectiveness and

minimizing their cost. The identification of early-

stage pathological objects in low dose high

resolution CT scans is a very difficult task for

radiologists, taking into account also the big (300-

400) number of noisy slices to be analyzed. To

support radiologists, researchers started the

development of CAD methods to be applied to CT

examinations (Camarlinghi, 2012; van Ginneken,

2010; Golosio, 2009; Gori, 2007; Li, 2003; Messai,

2010; Retico, 2009; Li, 2008). Several studies (Das,

2006; Brochu, 2007; Matsumoto, 2008) reported an

improvement in the sensitivity of radiologists when

assisted by CAD systems, in addition to a relevant

time saving. Other studies (Brown, 2005; Sahiner,

2009) observe that the increase in detection rate is

associated to an increase in the number of false-

213

E. Fantacci M., Bagnasco S., Camarlinghi N., Fiorina E., Lopez Torres E., Pennanzio F., Peroni C., Retico A., Saletta M., Sottocornola C., Traverso A.

and Cerello P..

A Web-based Computer Aided Detection System for Automated Search of Lung Nodules in Thoracic Computed Tomography Scans.

DOI: 10.5220/0005280102130218

In Proceedings of the International Conference on Bioinformatics Models, Methods and Algorithms (BIOINFORMATICS-2015), pages 213-218

ISBN: 978-989-758-070-3

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

positive findings. In addition, CAD systems act as

detection rates equalizers between observers of

different levels of experience (Brown, 2005). This

paper aims at validating the M5L CAD, which

combines the lungCAM and VBNA subsystems, on

the largest and most heterogenous dataset available,

so as to evaluate its readiness for application as a

support for screening programs and clinical practice.

2 MATERIALS AND METHODS

2.1 The Datasets

Among the required features of a system for clinical

and screening applications is the capability to

provide a performance independent of the dataset

source: for that reason, two public research datasets

were analyzed, collected both from screening

programs and from clinical practice.

The Lung Image Database Consortium (LIDC)

and Image Database Resource Initiative (IDRI)

(Armato, 2011) provide the largest publicly

available collection of annotated CTs: 1018 CT

scans in the LIDC/IDRI database are publicly

available since 2011. LIDC/IDRI is a multi-center

and multi-manufacturer database, which includes a

heterogeneous set of cases, with data taken at

different collimation, voltage, tube current and

reconstructed slice thickness. It therefore provides a

general sample which is likely to realistically

represent the input from a large scale multi-center

screening program as well as clinical practice. In

order to capture the inter-reader variability the

LIDC/IDRI consortium provides, for each CT scan,

four annotations made by different expert

radiologists, obtained with a two phase reading

modality. The LIDC/IDRI annotations contain

nodules with diameter between 3 and 30 mm. The

contours of nodules were marked and each nodule

was classified by every reader on a 1−5 scale and

with nine subjective characteristics: subtlety,

internal structure, calcification, sphericity, margin,

lobulation, spiculation, texture, malignancy. The

central position of nodules with diameter <3 mm and

non-nodules/anomalies with diameter > 3mm was

also recorded.

The ANODE09 (van Ginneken, 2010) dataset

consists of 55 anonymized CT scans provided by the

Utrecht University Medical Center and originates

from the NELSON study, the largest lung cancer

screening trial in Europe. 5 CT scans are made

available together with the radiologist annotations

and can be used for training a CAD system; 50 scans

can only be used for a blind validation. Most of the

database was randomly selected; however some CTs

with a large number of nodules were deliberately

included. Data were acquired with low-dose

exposure settings: 30 mA at 120 (140) kV for patient

weighting less (more) than 80 kg. Axial images were

reconstructed as a set of 2D 512x512 matrix images

with an average thickness of about 0.7 mm. The

ANODE09 annotation protocol foresees the labeling

of relevant nodules for structures with a diameter

larger than 4 mm.

2.2 The LungCAM CAD

The lungCAM structure is a standard approach: the

preprocessing stage (equalization and lung volume

segmentation) is followed by a search for Regions

Of Interest (ROIs), an analytical filter and a neural

classifier. Before starting the actual analysis, CT

scans in DICOM standard format are preprocessed

to reduce the noise contribution: each 2D slice is

analyzed with a Savitzky-Golay filter (Rajagolopan,

2003) that provides noise reduction without loss of

resolution. From then on, every step of the

lungCAM algorithm is intrinsically 3-dimensional.

2.2.1 Lung Segmentation

The lung segmentation (De Nunzio, 2011) proceeds

according to four main steps: analysis of the CT

Hounsfield Unit level distribution and evaluation of

the intensity threshold to be applied in the following

stages; 3D region growing of the lung volume with

the detected threshold; wavefront algorithm for the

definition of the lung surface on the inner side and

the removal of the trachea and the main bronchi;

morphological closing with a cylinder from the out-

side in order to include pleural nodules and close the

holes left by vessels. A check on the training/testing

and validation datasets confirmed that none of the

radiological findings were rejected at this stage.

2.2.2 ROI Hunting

The segmentation algorithm is performed with the

Channeler Ant Model (CAM) (Cerello, 2010), based

on Virtual Ant Colonies and conceived for the

segmentation of complex structures with different

shapes and intensity range in a noisy 3D

environment. The CAM exploits the natural

capabilities of Virtual Ant Colonies to modify the

environment and communicate with each other by

pheromone deposition. The ant life cycle is a

sequence of atomic time steps, during which the

BIOINFORMATICS2015-InternationalConferenceonBioinformaticsModels,MethodsandAlgorithms

214

behavior is determined by a set of rules that control

the pheromone release, the movements and the

variations of the ant energy, a parameter related to

breeding and death. The lung internal structures are

segmented by iteratively deploying ant colonies in

voxels with intensity above a pre-defined threshold

(anthills). Ants live according to the model rules

until the colony extinction: the pheromone

deposition generates pheromone maps. Each voxel

visited by an ant during the life of a colony is

removed from the allowed volume for future ant

colonies. New ant colonies are iteratively deployed

in unvisited voxels that meet the anthill requirement.

By an iterative thresholding of pheromone maps a

list of ROI candidates is obtained. ROIs with a

radius larger than 10 mm are post-processed in order

to disentangle nodules attached to internal lung

structures like vessels and bronchi. The CAM is

iteratively deployed in the right and left lungs,

separately, as a segmentation method for the vessel

tree and the nodule candidates. The first ant colony

segments the vessel tree, starting from an anthill in

the vicinity of its root. The segmented object is then

removed from the original image and the coordinates

of all its voxels are stored as a single Region Of

Interest (ROI). In the remaining image, iteratively,

any voxel with intensity above a predefined

threshold (-700 HU) is a new anthill and a colony

deployed from there generates a pheromone image.

When no more voxels meet the condition to become

an anthill, the information provided by the global

pheromone map is analyzed. The pheromone map

analysis is also iterative: each voxel with a

pheromone content above a minimum accepted

value is used as a seed for a region growing with an

adaptive threshold which is iteratively lowered until

a minimum growth rate of the region is reached.

Every grown region with a radius in the 0.8 − 25

mm range is considered as a nodule candidate.

About 20% of relevant pulmonary nodules are seg-

mented together with a vascular structure they are

connected to. If features were evaluated for the

whole ROI, these nodules would typically be

rejected by further filtering and classification. In

order to address the problem a dedicated algorithm

module was developed. All the structures obtained

from the pheromone map analysis with radius larger

than 10 mm are further analyzed in order to identify

and disentangle spherical-like sub-structures. The 10

mm value was empirically set based on the

minimum size for attached structures that causes a

relevant change in the ROI feature values. Each

voxel that belongs to the structure being analyzed is

averaged with the neighbors inside a sphere of

radius R. Then, the average map is thresholded

again, resulting in a thinner object. Structures with a

diameter smaller than R disappear (e.g., thin vessels

attached to the nodules). However also the nodules

shrink. In order to recover the nodule original size,

the neighbors of each remaining voxel in the average

inside a sphere of radius R/2 with value above 4/3 of

the threshold in the original map are restored as part

of the structure. The procedure is repeated three

times, with spheres of increasing radius (R = 1.5,

2.5, 3.5 mm) that generate sub-structures of

increasing size. The output voxels of the three

iterations are combined in logical OR to generate a

final nodule candidate output mask, which is then

treated as a ROI for further analysis.

2.2.3 Filtering and Classification

The choice of a suitable set of ROI features is a key

to the success of the filtering and classification

stages. Ideally, any computable quantity which is

expected to show a different pattern for true nodules

and false candidates would be a useful feature.

However, the use of a large number of features on a

small training dataset could bias the classifier and

cause a loss of generality. The choice to select a

small number of features for the neural classifier

training aims at optimizing the generality and

keeping the performance stable as the validation

dataset size increases. A set of features was selected

for the nodule candidate analysis, according to the

following criteria: 3D spatial features which are

invariant to rotation and translation and can

disentangle spherical-like structures from ROIs

originating from vessel parts or lung walls; features

based on the voxel HU intensity, so as to capture

density patterns; the fraction of ROI voxels attached

to the walls of the lung volume is crucial in

distinguishing internal and juxta-pleural nodules,

which are characterized by a different shape;

therefore, its use allows the classification of both the

subsamples with the same neural network. The list

of features is reported in Table 1. The average

number of ROIs after the nodule hunting, depending

on the number of slices, ranges between several

hundreds to few thousands per CT scan, a number

far too large to be used as input for a neural network

classifier. The vast majority of findings is easily

rejected with an analytical filter based on

correlations between the radius, the sphericity and

the fraction of voxels connected to the lung mask. In

addition to the sphericity-related selection, two other

filtering conditions were applied to the nodule

candidates: the fraction of voxels connected to lung

AWeb-basedComputerAidedDetectionSystemforAutomatedSearchofLungNodulesinThoracicComputed

TomographyScans

215

surface is required to be less than 0.6 and the Radius

must be larger than 1.2 mm. Irregular structures are

filtered with these criteria. The CT equalization and

filtering procedure dramatically reduces the average

number of FP findings per scan, from about 1000 to

about 50, a value which is appropriate as input for

training and running a neural classifier. The filtering

process also reduces the pre-classification sensitivity

to about 75 − 90%, depending on the input dataset.

Table 1: List of features extracted from the nodule output

mask. Features labeled with the asterisk were not used in

the classification stage.

Geometrical features Intensity-related features

Center of gravity Xi=x,y,z(*) Average

Radius (mm) Average outside mask

Sphericity Std. Deviation

Skewness of distance from Xi Std. Deviation outside mask

Kurtosis of distance from Xi Maximum

Volume (mm3) (*) Entropy

Fraction of voxels connected to

the pleura

Entropy outside mask

A feed forward neural network (FFNN) was

selected as nodule candidate classification method.

The training sample was made of 5 and 69 CTs from

the ANODE09 and LIDC/IDRI databases,

respectively. The training was carried on in cross-

validation mode. The FFNN configuration was

defined as follows: 13 input neurons, 1 hidden layer

with 25 neurons and 1 neuron in the output layer,

representing the probability of the finding to be

relevant.

2.3 VBNA CAD

The VBNA CAD system deals differently with

internal and juxtapleural nodules, by means of two

dedicated procedures: CADI for internal and CADJP

for juxtapleural nodules (Camarlinghi, 2012; Retico,

2008; Retico, 2009; Camarlinghi, 2011). Both are

three-step procedures. The first step consists in the

lung segmentation; the second step consists in the

ROI (Region Of Interest) hunter and performs the

candidate nodule selection; the third step consists in

the FP reduction. For the last step, an original

procedure, the Voxel-Based Neural Approach is

implemented to reduce the number of FPs in the lists

of internal and juxtapleural candidate nodules.

2.3.1 Segmentation

The aim of the segmentation algorithm implemented

in our analysis is to allow a conservative

identification of the internal region of the lung

parenchyma. In this region we apply the algorithm

for internal nodule detection. The 3-dimensional

segmentation algorithm is based on four main steps.

Once the scans have been isotropically resampled, to

separate the low-intensity lung parenchyma from the

high-intensity surrounding tissue (fat tissue and

bones), the voxel intensities are thresholded at a

fixed value; then, in order to discard all the regions

not belonging to the lungs, the biggest connected

component not crossing the boundary of the volume

is considered. Vessels and airways are not in

included in the segmented lung at this stage since

their volume is outside the segmented lung volume.

To include them without modifying the pleura

surface morphology, i.e. without modifying the

shape of pleura irregularities (including juxtapleural

nodules), a combination of morphological operators

is applied. In particular, a sequence of the dilation

and the erosion operators with spherical kernels rd

and re, with re > rd, is implemented. Finally, the

logical OR operation between the so obtained mask

and the original lung mask provides the final mask

P, where the vessels and the airway walls are filled

in, while maintaining the original shape of the lung

border. The identified lung mask is used for CADI,

whereas its boundary is used for CADJP.

2.3.2 ROI Hunting for Internal Nodules

In the CADI, the internal nodules are modeled as

spherical objects with a Gaussian profile, following

the approach proposed in (Li, 2003). To detect this

kind of objects (Retico, 2008), a dedicated dot-

enhancement (DE) filter is implemented. The filter

determines the local geometrical characteristics of

each voxel by using the eigenvalues of the Hessian

matrix. To enhance the sensitivity of this filter to

nodules of different sizes, a multi-scale approach has

to be followed. This approach combines the DE

function with Gaussian smoothing at several scales

with the prescriptions given in (Li, 2003). Local

maxima of the matrix filtered by the dot-

enhancement are the internal candidate nodule

locations.

2.3.3 ROI Hunting for Juxta-Pleural

Nodules

In the CADJP (Retico, 2009), in order to identify

juxtapleural candidate nodules, pleura surface

normals are constructed and each voxel is assigned a

score proportional to the number of normals

intersecting in it. Normals are evaluated using the

triangular mesh representing the pleura surface,

obtained applying the marching cube algorithm on

BIOINFORMATICS2015-InternationalConferenceonBioinformaticsModels,MethodsandAlgorithms

216

the lung mask. In particular, the normal to each

triangle is calculated by using the vector product

between the triangle edges; then, the normals to each

mesh vertex are evaluated averaging all the triangle

normals of the neighboring triangles. Since the

evaluation of the normal intersections in the real 3D

space is a complex and computationally intensive

operation, it is implemented in the voxel space. This

means that each voxel is associated a score

proportional to the number of normals passing

through it. To deal with noise, cylinders with

Gaussian profile are considered instead of segments

(Paik, 2004). This information is collected in the

score matrix S(x,y,z). The local maxima of the 3D

matrix S(x,y,z) are the juxtapleural candidate nodule

locations.

2.3.4 Classification

In order to classify the candidate nodule findings

obtained in the previous step, an original procedure,

the Voxel-Based Neural Approach (Gori, 2007),

performs the reduction of the number of FPs in the

lists of internal and juxtapleural candidate nodules.

First, a ROI including voxels belonging to the

candidate nodule is defined from each location

provided by the previous step. The basic idea of the

VBNA is to associate with each voxel of a ROI a

feature vector defined by the intensity values of its

3D neighbors (in this case 5 x 5 x 5 intensity values)

and the eigenvalues of the gradient matrix and of the

Hessian matrix. In the firsta version of VBNA, che

classification procedure was perfoermed by means

of a FFNN. Now support vector machines (SVM),

by which have been obtained better resukts, are

implemented for the classification procedure. Other

classification methods have not yet already been

tested. The training sample was made of 69 CTs

from the LIDC/IDRI database. At the end of this

step, each ROI is assigned a degree of suspicion

averaging the score of all the voxels belonging to it.

2.4 Subsystems Combination: the M5L

CAD

Each CAD subsystem can be improved in the future,

working on specific weaknesses. However, one

quick and effective way to improve the overall

performance is to combine the results, as

demonstrated in (van Ginneken, 2010) for the

ANODE09 challenge participants. The outputs of

the two CAD subsystems described are evaluated

and combined following the same procedure adopted

for the ANODE09 study (van Ginneken, 2010). The

resulting CAD system is referred to as M5L. The

findings of each CAD subsystem must be considered

in terms of their degree of suspicion p, which is the

final output of the procedure of candidate nodules

classification for the two separate subsystems.

In order to combine findings from different CAD

subsystems, a normalization of the finding

probabilities is needed (Niemeijer, 2011). This

operation is carried out by associating a new value

f(p) to each finding with degree of suspicion p. The

new degree of suspicion f(p) is evaluated according

to the performance obtained by the corresponding

CAD system on the validation set, i.e., evaluating

for each finding with probability p the function

corresponding to TP/(FP+TP+1), where TP(FP) is

the number of true (false) positives obtained by

considering all the CAD findings with pi>=p. Of

course, this procedure requires to know the

annotations and the performance of each CAD

system on a selected set of data.

The f(p) values can

therefore be considered as the score related to the

probability that a finding in the validation set with

likelihood p or higher represents a true nodule. The

function f(p) is computed for every finding from

every subsystem. All findings are then checked

against a “matching condition” defined by a

preselected clustering distance.

3 RESULTS

The results have been evaluated in terms of FROC

(Free-response Receiver Operating Characteristic)

curves. In fact, Receiver Operating Characteristic

(ROC) methodology is widely used in evaluating

medical imaging modalities but has several

drawbacks when the detection task, e.g., nodule

detection, involves localizing the abnormality, while

FROC methodology offers a more natural

framework to describe observer performance in such

studies and has other advantages (Chakraborty,

1989).

Figure 1 shows the results obtained for the

lungCAM and VBNA separate subsystems and for

the combined M5L on the 949 scans of the LIDC

test dataset (949 scans). To obtain the combined

M5L result the following matching criterion has

been used: a CAD finding is considered a true

positive if its Euclidean distance from the center of

the lesion annotated by the radiologists is less than

1.5 times the radius of the annotated lesion. The

M5L sensitivity at 8 FP/scan reaches 80% which,

given the size and heterogeneity of the dataset, is

AWeb-basedComputerAidedDetectionSystemforAutomatedSearchofLungNodulesinThoracicComputed

TomographyScans

217

quite remarkable. In the case of ANODE09 the

FROC curves are shown in Figure 2.

Figure 1: FROC curves of the lungCAM and VBNA

subsystems ad of their M5L combination on the LIDC test

validation dataset (949 CT scans).

Figure 2: FROC curves of the lungCAM and VBNA

subsystems ad of their M5L combination on the

ANODE09 test validation dataset (50 CT scans).

4 CONCLUSIONS

The results, obtained on a database so large and

heterogeneous, are very satisfactory. One of the

main purposes of this work was to show that, even

without changing parameters and making

optimizations, the performance is satisfactory. In

fact we applied a previous training configuration to a

much larger and heterogeneous dataset (the full

LIDC/IDRI). In view of a future application of the

M5L CAD in screening programs or clinical

practice, the optimization can be achieved by

iteratively using training samples of increasing size.

Furthermore, demonstrating a generalization

capability is, at the present development stage, even

more important than optimizing the sensitivity on a

selected dataset. The M5L CAD has been already

implemented in a cloud computing environment

(Berzano, 2012) and is now available for the

radiologists of our collaboration as experimental

web service for clinical tests.

REFERENCES

Parkin, D. et al., 2010. Int J Cancer, 127(12), 2893.

American Cancer Society, 2009. Cancer Facts and

Figures.http://www.cancer.org/Research/CancerFacts

Figures.

Jemal, A. et al., 2010. CA Cancer J. Clin. 60, 277.

Henschke, C. et al., 1999. Lancet, 354(9173), 99.

The NLST (National Lubg Screening Trial) Research

Team, 2011. N. Engl. J. Med. 365, 395.

Camarlinghi, N. et al., 2012. Int. J. Comput. Assist.

Radiol. Surg. 7, 455.

Van Ginneken B. et al., 2010. Med. Image Anal. 14, 707.

Golosio, B. et al., 2009. Med. Phys. 36, 3607.

Gori, I. et al., 2007. Proceedings of the SPIE Medical

Imaging Conference 6514, 6514R.

Li, Q. et al., 2003. Med. Phys. 30, 2040.

Messay, T. et al., 2010. Med. Image Anal. 14, 390.

Retico, A. et al., 2009. SPIE Medical Imaging 2009:

Computer-Aided Diagnosis, 7260, 72601S.

Li, Q. et al., 2008. Acad. Radiol. 15, 165.

Das, M. et al., 2006. Radiology 241, 564.

Brochu, B. et al., 2007. Journal de Radiologie 88, 573.

Matsumoto, S. et al., 2008. Radiation Medicine 26, 562.

Brown, M.S. et al., 2005. Acad. Radiol. 12, 681.

Sahiner, B. et al., 2009. Acad. Radiol. 16, 1518.

Armato III, S.G. et al., 2011. Med. Phys. 38, 915.

Rajagolopan S. et al., 2003. Proc. SPIE Medical Imaging

5029, 773.

De Nunzio, G. et al., 2011. Journal of Digital Imaging 24,

11.

Cerello, P. et al., 2010. Pattern Recognition 43, 1476.

Retico, A. et al., 2008. Comput. Biol. Med., 38(4), 525.

Retico, A. et al., 2009. Comput Biol Med 39(12), 1137.

Camarlinghi, N. et al., 2011. Il Nuovo Cimento, 1, 65.

Paik, S.D. et al., 2004. IEEE Trans Med Imaging 23(6),

661.

Niemeijer, M. et al., 2011. IEEE Trans Med Imaging

30(2), 215.

Chakraborty, D., 1989. Med. Phys. 16, 561.

Berzano, D. et al., 2012. IEEE Nuclear Science

Symposium and Medical Imaging Conference Record

(NSS/MIC), 968.

BIOINFORMATICS2015-InternationalConferenceonBioinformaticsModels,MethodsandAlgorithms

218