fuzzy knn lung cancer identification by an electronic nosefuzzy knn lung cancer identification by an...
TRANSCRIPT
Fuzzy kNN Lung Cancer Identification by an Electronic Nose
Rossella Blatt1, Andrea Bonarini1, Elisa Calabrò2 Matteo Della Torre3, Matteo Matteucci1 and Ugo Pastorino2
1 Politecnico di Milano, Department of Electronics and Information, Milan, Italy1 Istituto Nazionale Tumori of Milan, Toracic Surgery Department, Milan, Italy
1 SACMI Imola S.C., Automation & Inspection Systems, Imola (BO), Italy
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 2
Outline Objective: Lung Cancer diagnosis classifying the
Olfactory Signal acquired by an Electronic Nose Motivation Functioning of the Electronic Nose
• Signal Acquisition• Signalpreprocessing • Dimensionality Reduction• Pattern Classification and Validation
Comparison with other classifiers, state of the art and current diagnostic techniques
Discussion the importance of obtained results and further directions of research
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 3
Outline Objective: Lung Cancer diagnosis classifying the
Olfactory Signal acquired by an Electronic Nose Motivation Functioning of the Electronic Nose
• Signal Acquisition• Signalpreprocessing • Dimensionality Reduction• Pattern Classification and Validation
Comparison with other classifiers, state of the art and current diagnostic techniques
Discussion the importance of obtained results and further directions of research
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 4
Outline Objective: Lung Cancer diagnosis classifying the
Olfactory Signal acquired by an Electronic Nose Motivation Functioning of the Electronic Nose
• Signal Acquisition• Signalpreprocessing • Dimensionality Reduction• Pattern Classification and Validation
Comparison with other classifiers, state of the art and current diagnostic techniques
Discussion the importance of obtained results and further directions of research
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 5
Outline Objective: Lung Cancer diagnosis classifying the
Olfactory Signal acquired by an Electronic Nose Motivation Functioning of the Electronic Nose
• Signal Acquisition• Signalpreprocessing • Dimensionality Reduction• Pattern Classification and Validation
Comparison with other classifiers, state of the art and current diagnostic techniques
Discussion the importance of obtained results and further directions of research
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 6
Outline Objective: Lung Cancer diagnosis classifying the
Olfactory Signal acquired by an Electronic Nose Motivation Functioning of the Electronic Nose
• Signal Acquisition• Signalpreprocessing • Dimensionality Reduction• Pattern Classification and Validation
Comparison with other classifiers, state of the art and current diagnostic techniques
Discussion the importance of obtained results and further directions of research
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 7
Outline Objective: Lung Cancer diagnosis classifying the
Olfactory Signal acquired by an Electronic Nose Motivation Functioning of the Electronic Nose
• Signal Acquisition• Signalpreprocessing • Dimensionality Reduction• Pattern Classification and Validation
Comparison with other classifiers, state of the art and current diagnostic techniques
Discussion the importance of obtained results and further directions of research
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 8
Motivation Lung cancer causes more than 160,000 deaths a year in the
United Statesmore than any other cancer Once the cancer has been detected, only 14% survives after 5
years of therapy; the survival rate increases to 48% if the cancer is discovered in its earliest stage
Current diagnostic techniques are invasive, very expensive, have a high risk of complications and a not so good performance
Efforts at early detection and treatment have been frustrating to date and hence the overall prognosis remains poor
NECESSITY OF A NON INVASIVE, MORE ACCURATE AND CHEAPER DIAGNOSTIC TECHNIQUE, ABLE TO IDENTIFY THE PRESENCE OF LUNG CANCER IN ITS EARLY STAGE
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 9
Motivation Lung cancer causes more than 160,000 deaths a year in the
United Statesmore than any other cancer Once the cancer has been detected, only 14% survives after 5
years of therapy; the survival rate increases to 48% if the cancer is discovered in its earliest stage
Current diagnostic techniques are invasive, very expensive, have a high risk of complications and a not so good performance
Efforts at early detection and treatment have been frustrating to date and hence the overall prognosis remains poor
NECESSITY OF A NON INVASIVE, MORE ACCURATE AND CHEAPER DIAGNOSTIC TECHNIQUE, ABLE TO IDENTIFY THE PRESENCE OF LUNG CANCER IN ITS EARLY STAGE
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 10
Motivation Lung cancer causes more than 160,000 deaths a year in the
United Statesmore than any other cancer Once the cancer has been detected, only 14% survives after 5
years of therapy; the survival rate increases to 48% if the cancer is discovered in its earliest stage
Current diagnostic techniques are invasive, very expensive, have a high risk of complications and a not so good performance
Efforts at early detection and treatment have been frustrating to date and hence the overall prognosis remains poor
NECESSITY OF A NON INVASIVE, MORE ACCURATE AND CHEAPER DIAGNOSTIC TECHNIQUE, ABLE TO IDENTIFY THE PRESENCE OF LUNG CANCER IN ITS EARLY STAGE
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 11
Motivation Lung cancer causes more than 160,000 deaths a year in the
United Statesmore than any other cancer Once the cancer has been detected, only 14% survives after 5
years of therapy; the survival rate increases to 48% if the cancer is discovered in its earliest stage
Current diagnostic techniques are invasive, very expensive, have a high risk of complications and a not so good performance
Efforts at early detection and treatment have been frustrating to date and hence the overall prognosis remains poor
NECESSITY OF A NON INVASIVE, MORE ACCURATE AND CHEAPER DIAGNOSTIC TECHNIQUE, ABLE TO IDENTIFY THE PRESENCE OF LUNG CANCER IN ITS EARLY STAGE
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 12
Motivation Fundamental Principle of Clinical Chemistry: “Every
pathology changes people chemical composition, modifying the concentration of some chemicals in the human body”• In the medical field, clinicians have always considered
odor as a fundamental information for the diagnosis of several diseases
It has been demonstrated (Gordon et al, 1985) that the presence of lung cancer alters the percentage of some volatile organic compounds (VOCs) present in human breath• These VOCs can be considered as lung cancer markers
and thus used to diagnose it
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 13
Motivation Fundamental Principle of Clinical Chemistry: “Every
pathology changes people chemical composition, modifying the concentration of some chemicals in the human body”• In the medical field, clinicians have always considered
odor as a fundamental information for the diagnosis of several diseases
It has been demonstrated (Gordon et al, 1985) that the presence of lung cancer alters the percentage of some volatile organic compounds (VOCs) present in human breath• These VOCs can be considered as lung cancer markers
and thus used to diagnose it
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 14
Motivation Fundamental Principle of Clinical Chemistry: “Every
pathology changes people chemical composition, modifying the concentration of some chemicals in the human body”• In the medical field, clinicians have always considered
odor as a fundamental information for the diagnosis of several diseases
It has been demonstrated (Gordon et al, 1985) that the presence of lung cancer alters the percentage of some volatile organic compounds (VOCs) present in human breath• These VOCs can be considered as lung cancer markers
and thus used to diagnose it
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 15
Motivation Fundamental Principle of Clinical Chemistry: “Every
pathology changes people chemical composition, modifying the concentration of some chemicals in the human body”• In the medical field, clinicians have always considered
odor as a fundamental information for the diagnosis of several diseases
It has been demonstrated (Gordon et al, 1985) that the presence of lung cancer alters the percentage of some volatile organic compounds (VOCs) present in human breath• These VOCs can be considered as lung cancer markers
and thus used to diagnose it
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 16
Electronic Nose An electronic nose is an instrument able to acquire, detect and
analyse the olfactory signal It is composed of an array of non specific electronic devices able to
convert a physical or chemical information into an electrical signal• It is non specific because it does
not look for particular compounds in the analyzed substance, but for different patterns
• Each sensor reacts in a different way to the analyzed substance providing multidimensional data that can be considered as an olfactory blueprint of the substance itself
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 17
Electronic Nose An electronic nose is an instrument able to acquire, detect and
analyse the olfactory signal It is composed of an array of non specific electronic devices able to
convert a physical or chemical information into an electrical signal• It is non specific because it does
not look for particular compounds in the analyzed substance, but for different patterns
• Each sensor reacts in a different way to the analyzed substance providing multidimensional data that can be considered as an olfactory blueprint of the substance itself
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 18
Electronic Nose An electronic nose is an instrument able to acquire, detect and
analyse the olfactory signal It is composed of an array of non specific electronic devices able to
convert a physical or chemical information into an electrical signal• It is non specific because it does
not look for particular compounds in the analyzed substance, but for different patterns
• Each sensor reacts in a different way to the analyzed substance providing multidimensional data that can be considered as an olfactory blueprint of the substance itself
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 19
Electronic Nose An electronic nose is an instrument able to acquire, detect and
analyse the olfactory signal It is composed of an array of non specific electronic devices able to
convert a physical or chemical information into an electrical signal• It is non specific because it does
not look for particular compounds in the analyzed substance, but for different patterns
• Each sensor reacts in a different way to the analyzed substance providing multidimensional data that can be considered as an olfactory blueprint of the substance itself
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 20
Electronic Nose's Output According to the used pattern analysis algorithm, the
output of an electronic nose can be:• the detection of a specific substance• an estimate of the concentration of the odor• some particular characteristic of the odor that allows to
associate it to a particular class
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 21
Electronic Nose's Output According to the used pattern analysis algorithm, the
output of an electronic nose can be:• the detection of a specific substance• an estimate of the concentration of the odor• some particular characteristic of the odor that allows to
associate it to a particular class
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 22
Electronic Nose's Output According to the used pattern analysis algorithm, the
output of an electronic nose can be:• the detection of a specific substance• an estimate of the concentration of the odor• some particular characteristic of the odor that allows to
associate it to a particular class
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 23
Electronic Nose's Output According to the used pattern analysis algorithm, the
output of an electronic nose can be:• the detection of a specific substance• an estimate of the concentration of the odor• some particular characteristic of the odor that allows to
associate it to a particular class
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 24
Electronic Nose1. Signal Acquisition
● Acquisition is done through a sensor array that measures a given physical or chemical quantity and convert it into an electrical signal
2. Signal Processing● Preprocessing: aimed to reduce the impact of noise● Dimensionality Reduction: reduce the dimensionality of the problem, enhancing
classification performance
3. Classification and Validation● Classification between the two classes “healthy” and “sick”
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 25
Electronic Nose1. Signal Acquisition
● Acquisition is done through a sensor array that measures a given physical or chemical quantity and convert it into an electrical signal
2. Signal Processing● Preprocessing: aimed to reduce the impact of noise● Dimensionality Reduction: reduce the dimensionality of the problem, enhancing
classification performance
3. Classification and Validation● Classification between the two classes “healthy” and “sick”
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 26
Electronic Nose1. Signal Acquisition
● Acquisition is done through a sensor array that measures a given physical or chemical quantity and convert it into an electrical signal
2. Signal Processing● Preprocessing: aimed to reduce the impact of noise● Dimensionality Reduction: reduce the dimensionality of the problem, enhancing
classification performance
3. Classification and Validation● Classification between the two classes “healthy” and “sick”
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 27
Electronic Nose1. Signal Acquisition
● Acquisition is done through a sensor array that measures a given physical or chemical quantity and convert it into an electrical signal
2. Signal Processing● Preprocessing: aimed to reduce the impact of noise● Dimensionality Reduction: reduce the dimensionality of the problem, enhancing
classification performance
3. Classification and Validation● Classification between the two classes “healthy” and “sick”
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 28
Signal Acquisition
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 29
Sensors' Technology Current types of electronic noses are different according to the used
sensor technology and, therefore, according to the measured physical or chemical quantity• Conductivity sensors
− Metal Oxide sensor (MOS)− Polymer sensors
• Piezoelectric sensors (Mass change)− Quartz Crystal Microbalance (QCM)− Surface Acousticwave (SAW)
• MOSFET sensors• OPTICALfiber sensor
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 30
Sensors' Technology Current types of electronic noses are different according to the used
sensor technology and, therefore, according to the measured physical or chemical quantity• Conductivity sensors
− Metal Oxide sensor (MOS)− Polymer sensors
• Piezoelectric sensors (Mass change)− Quartz Crystal Microbalance (QCM)− Surface Acousticwave (SAW)
• MOSFET sensors• OPTICALfiber sensor
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 31
Sensors' Technology Current types of electronic noses are different according to the used
sensor technology and, therefore, according to the measured physical or chemical quantity• Conductivity sensors
− Metal Oxide sensor (MOS)− Polymer sensors
• Piezoelectric sensors (Mass change)− Quartz Crystal Microbalance (QCM)− Surface Acousticwave (SAW)
• MOSFET sensors• OPTICALfiber sensor
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 32
Sensors' Technology Current types of electronic noses are different according to the used
sensor technology and, therefore, according to the measured physical or chemical quantity• Conductivity sensors
− Metal Oxide sensor (MOS)− Polymer sensors
• Piezoelectric sensors (Mass change)− Quartz Crystal Microbalance (QCM)− Surface Acousticwave (SAW)
• MOSFET sensors• OPTICALfiber sensor
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 33
Sensors' Technology Current types of electronic noses are different according to the used
sensor technology and, therefore, according to the measured physical or chemical quantity• Conductivity sensors
− Metal Oxide sensor (MOS)− Polymer sensors
• Piezoelectric sensors (Mass change)− Quartz Crystal Microbalance (QCM)− Surface Acousticwave (SAW)
• MOSFET sensors• OPTICALfiber sensor
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 34
MOS Sensors VOCs markers of lung cancer are present in diseased
people's breath in very small quantities (varying from parts per million ppm to parts per billion ppb)
The most important MOS sensors' peculiarity is their high sensitivity (in the order of ppb)• Moreover they result to be characterized by low cost, high
speed response and relatively simple electronics We used an array of six MOS sensors
that react to gases with a variation of resistance• The registered signal corresponds to the
change of resistance through time produced by gas flow
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 35
MOS Sensors VOCs markers of lung cancer are present in diseased
people's breath in very small quantities (varying from parts per million ppm to parts per billion ppb)
The most important MOS sensors' peculiarity is their high sensitivity (in the order of ppb)• Moreover they result to be characterized by low cost, high
speed response and relatively simple electronics We used an array of six MOS sensors
that react to gases with a variation of resistance• The registered signal corresponds to the
change of resistance through time produced by gas flow
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 36
MOS Sensors VOCs markers of lung cancer are present in diseased
people's breath in very small quantities (varying from parts per million ppm to parts per billion ppb)
The most important MOS sensors' peculiarity is their high sensitivity (in the order of ppb)• Moreover they result to be characterized by low cost, high
speed response and relatively simple electronics We used an array of six MOS sensors
that react to gases with a variation of resistance• The registered signal corresponds to the
change of resistance through time produced by gas flow
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 37
MOS Sensors' Response Each measure consists of three main phases:
• Before: the instrument inhales the reference air, showing in its graph a relatively constant curve
• During: the electronic nose inhales the analyzed gas, producing a change in the sensors' response
− It contains informations about how sensors react to a particular substance
• After: during this phase the instrument returns to the reference line
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 38
MOS Sensors' Response Each measure consists of three main phases:
• Before: the instrument inhales the reference air, showing in its graph a relatively constant curve
• During: the electronic nose inhales the analyzed gas, producing a change in the sensors' response
− It contains informations about how sensors react to a particular substance
• After: during this phase the instrument returns to the reference line
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 39
MOS Sensors' Response Each measure consists of three main phases:
• Before: the instrument inhales the reference air, showing in its graph a relatively constant curve
• During: the electronic nose inhales the analyzed gas, producing a change in the sensors' response
− It contains informations about how sensors react to a particular substance
• After: during this phase the instrument returns to the reference line
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 40
MOS Sensors' Response Each measure consists of three main phases:
• Before: the instrument inhales the reference air, showing in its graph a relatively constant curve
• During: the electronic nose inhales the analyzed gas, producing a change in the sensors' response
− It contains informations about how sensors react to a particular substance
• After: during this phase the instrument returns to the reference line
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 41
Involved Population We analyzed the breath of 101 volunteers
Total population101 volunteers
Healthy58
Lung Cancer43
Primary Lung Cancer
23
Pulmonary Metastasis
20
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 42
Involved Population We analyzed the breath of 101 volunteers
Total population101 volunteers
Healthy58
Lung Cancer43
Primary Lung Cancer
23
Pulmonary Metastasis
20
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 43
Involved Population We analyzed the breath of 101 volunteers
Total population101 volunteers
Healthy58
Lung Cancer43
Primary Lung Cancer
23
Pulmonary Metastasis
20
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 44
Signal Acquisition
The breath acquisition has been made inviting all volunteers to blow into a nalophan bag of approximately 400cm3
We considered only the portion of the breath exhaled directly form lung
Finally, the air contained in the bag was input into the electronic nose
For each person we took two measures for a total of 202 measurements (116 healthy, 86 diseased)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 45
Signal Acquisition
The breath acquisition has been made inviting all volunteers to blow into a nalophan bag of approximately 400cm3
We considered only the portion of the breath exhaled directly form lung
Finally, the air contained in the bag was input into the electronic nose
For each person we took two measures for a total of 202 measurements (116 healthy, 86 diseased)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 46
Signal Acquisition
The breath acquisition has been made inviting all volunteers to blow into a nalophan bag of approximately 400cm3
We considered only the portion of the breath exhaled directly form lung
Finally, the air contained in the bag was input into the electronic nose
For each person we took two measures for a total of 202 measurements (116 healthy, 86 diseased)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 47
Signal Acquisition
The breath acquisition has been made inviting all volunteers to blow into a nalophan bag of approximately 400cm3
We considered only the portion of the breath exhaled directly form lung
Finally, the air contained in the bag was input into the electronic nose
For each person we took two measures for a total of 202 measurements (116 healthy, 86 diseased)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 48
Preprocessing & Dimensionality Reduction
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 49
Signal preprocessing
Preprocessing:• Manipulation of the baseline: transformation of the sensor
response w.r.t. its baseline for the purpose of contrast enhancement and drift compensation
• Reduction of humidity effects• Normalization: compensation for the scale difference
among the six sensors− Each sensor has been forced to have zero mean and variance
equal to 1
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 50
Signal preprocessing
Preprocessing:• Manipulation of the baseline: transformation of the sensor
response w.r.t. its baseline for the purpose of contrast enhancement and drift compensation
• Reduction of humidity effects• Normalization: compensation for the scale difference
among the six sensors− Each sensor has been forced to have zero mean and variance
equal to 1
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 51
Signal preprocessing
Preprocessing:• Manipulation of the baseline: transformation of the sensor
response w.r.t. its baseline for the purpose of contrast enhancement and drift compensation
• Reduction of humidity effects• Normalization: compensation for the scale difference
among the six sensors− Each sensor has been forced to have zero mean and variance
equal to 1
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 52
Signal preprocessing
Preprocessing:• Manipulation of the baseline: transformation of the sensor
response w.r.t. its baseline for the purpose of contrast enhancement and drift compensation
• Reduction of humidity effects• Normalization: compensation for the scale difference
among the six sensors− Each sensor has been forced to have zero mean and variance
equal to 1
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 53
Dimensionality Reduction
Objective: reduce the dimensionality of the problem, improving the classification ability 1. Extraction of a number of descriptors able to represent data
characteristics in the most efficient way2.Feature Selection: maximization of the informative content3.Feature Extraction: projection in the lower dimensional space
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 54
Dimensionality Reduction
Objective: reduce the dimensionality of the problem, improving the classification ability 1. Extraction of a number of descriptors able to represent data
characteristics in the most efficient way2.Feature Selection: maximization of the informative content3.Feature Extraction: projection in the lower dimensional space
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 55
Dimensionality Reduction
Objective: reduce the dimensionality of the problem, improving the classification ability 1. Extraction of a number of descriptors able to represent data
characteristics in the most efficient way2.Feature Selection: maximization of the informative content3.Feature Extraction: projection in the lower dimensional space
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 56
Dimensionality Reduction
Objective: reduce the dimensionality of the problem, improving the classification ability 1. Extraction of a number of descriptors able to represent data
characteristics in the most efficient way2.Feature Selection: maximization of the informative content3.Feature Extraction: projection in the lower dimensional space
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 57
Descriptors Extraction
Extraction of a number of descriptors able to represent data characteristics in the most efficient way
We considered 10 descriptors, calculated according to:
• resistance's variation• course of the curve• derivative• integral• FFT• resistance's minimum
Some of them were vectors, for a total of 39 descriptors Given 6 sensors, we obtained 234 dimensions
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 58
Descriptors Extraction
Extraction of a number of descriptors able to represent data characteristics in the most efficient way
We considered 10 descriptors, calculated according to:
• resistance's variation• course of the curve• derivative• integral• FFT• resistance's minimum
Some of them were vectors, for a total of 39 descriptors Given 6 sensors, we obtained 234 dimensions
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 59
Descriptors Extraction
Extraction of a number of descriptors able to represent data characteristics in the most efficient way
We considered 10 descriptors, calculated according to:
• resistance's variation• course of the curve• derivative• integral• FFT• resistance's minimum
Some of them were vectors, for a total of 39 descriptors Given 6 sensors, we obtained 234 dimensions
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 60
Descriptors Extraction
Extraction of a number of descriptors able to represent data characteristics in the most efficient way
We considered 10 descriptors, calculated according to:
• resistance's variation• course of the curve• derivative• integral• FFT• resistance's minimum
Some of them were vectors, for a total of 39 descriptors Given 6 sensors, we obtained 234 dimensions
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 61
Descriptors Extraction
Extraction of a number of descriptors able to represent data characteristics in the most efficient way
We considered 10 descriptors, calculated according to:
• resistance's variation• course of the curve• derivative• integral• FFT• resistance's minimum
Some of them were vectors, for a total of 39 descriptors Given 6 sensors, we obtained 234 dimensions
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 62
Descriptors Extraction
Extraction of a number of descriptors able to represent data characteristics in the most efficient way
We considered 10 descriptors, calculated according to:
• resistance's variation• course of the curve• derivative• integral• FFT• resistance's minimum
Some of them were vectors, for a total of 39 descriptors Given 6 sensors, we obtained 234 dimensions
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 63
Descriptors Extraction
Extraction of a number of descriptors able to represent data characteristics in the most efficient way
We considered 10 descriptors, calculated according to:
• resistance's variation• course of the curve• derivative• integral• FFT• resistance's minimum
Some of them were vectors, for a total of 39 descriptors Given 6 sensors, we obtained 234 dimensions
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 64
Feature Selection Objective: maximization of the informative content
• Nonparametric test of MannWhitneyWilcoxon• Scatter Plot• MANOVA
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 65
Feature Selection Objective: maximization of the informative content
• Nonparametric test of MannWhitneyWilcoxon• Scatter Plot• MANOVA
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 66
Feature Selection
We found that the most discriminative features were:● Delta: sensors' resistance change during the measurement● Classic: the ratio between the reference line and the minimum value
reached during the measurement● Phase Integral: the closed area determined by the plot of the state of
the graph of the measurement
Number of directions: 5 features x 6 sensors = 30 directions
● Integral● Single Point: the
minimum value of the resistance reached during the measurement
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 67
Feature Selection
We found that the most discriminative features were:● Delta: sensors' resistance change during the measurement● Classic: the ratio between the reference line and the minimum value
reached during the measurement● Phase Integral: the closed area determined by the plot of the state of
the graph of the measurement
Number of directions: 5 features x 6 sensors = 30 directions
● Integral● Single Point: the
minimum value of the resistance reached during the measurement
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 68
Feature Extraction Objective: project data in a lower dimensional space maximizing the
classification accuracy ● Principal Component Analysis PCA
● PCA transforms data in a linear way, projecting features into the direction with maximum variance
● It does not take in account sample labels● Non Parametric Linear Discriminant Analysis NPLDA (Fukunaga, 1983)
● A generalization of Fisher's linear discriminant analysis: it looks for the projections able to maximize differences between classes and minimize those intraclass
● It removes the unimodal gaussian assumption by computing the between scatter matrix S
b using local information and the knearest
neighbors rule● S
b is full rank, allowing to extract more than c1 features
(c=number of classes)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 69
Feature Extraction Objective: project data in a lower dimensional space maximizing the
classification accuracy ● Principal Component Analysis PCA
● PCA transforms data in a linear way, projecting features into the direction with maximum variance
● It does not take in account sample labels● Non Parametric Linear Discriminant Analysis NPLDA (Fukunaga, 1983)
● A generalization of Fisher's linear discriminant analysis: it looks for the projections able to maximize differences between classes and minimize those intraclass
● It removes the unimodal gaussian assumption by computing the between scatter matrix S
b using local information and the knearest
neighbors rule● S
b is full rank, allowing to extract more than c1 features
(c=number of classes)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 70
Feature Extraction Objective: project data in a lower dimensional space maximizing the
classification accuracy ● Principal Component Analysis PCA
● PCA transforms data in a linear way, projecting features into the direction with maximum variance
● It does not take in account sample labels● Non Parametric Linear Discriminant Analysis NPLDA (Fukunaga, 1983)
● A generalization of Fisher's linear discriminant analysis: it looks for the projections able to maximize differences between classes and minimize those intraclass
● It removes the unimodal gaussian assumption by computing the between scatter matrix S
b using local information and the knearest
neighbors rule● S
b is full rank, allowing to extract more than c1 features
(c=number of classes)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 71
Feature Extraction Objective: project data in a lower dimensional space maximizing the
classification accuracy ● Principal Component Analysis PCA
● PCA transforms data in a linear way, projecting features into the direction with maximum variance
● It does not take in account sample labels● Non Parametric Linear Discriminant Analysis NPLDA (Fukunaga, 1983)
● A generalization of Fisher's linear discriminant analysis: it looks for the projections able to maximize differences between classes and minimize those intraclass
● It removes the unimodal gaussian assumption by computing the between scatter matrix S
b using local information and the knearest
neighbors rule● S
b is full rank, allowing to extract more than c1 features
(c=number of classes)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 72
Feature Extraction Objective: project data in a lower dimensional space maximizing the
classification accuracy ● Principal Component Analysis PCA
● PCA transforms data in a linear way, projecting features into the direction with maximum variance
● It does not take in account sample labels● Non Parametric Linear Discriminant Analysis NPLDA (Fukunaga, 1983)
● A generalization of Fisher's linear discriminant analysis: it looks for the projections able to maximize differences between classes and minimize those intraclass
● It removes the unimodal gaussian assumption by computing the between scatter matrix S
b using local information and the knearest
neighbors rule● S
b is full rank, allowing to extract more than c1 features
(c=number of classes)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 73
Feature Extraction Objective: project data in a lower dimensional space maximizing the
classification accuracy ● Principal Component Analysis PCA
● PCA transforms data in a linear way, projecting features into the direction with maximum variance
● It does not take in account sample labels● Non Parametric Linear Discriminant Analysis NPLDA (Fukunaga, 1983)
● A generalization of Fisher's linear discriminant analysis: it looks for the projections able to maximize differences between classes and minimize those intraclass
● It removes the unimodal gaussian assumption by computing the between scatter matrix S
b using local information and the knearest
neighbors rule● S
b is full rank, allowing to extract more than c1 features
(c=number of classes)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 74
Feature Extraction: PCA vs NPLDA
No discrimination on the yaxis Poor discrimination with very
strong overlap on the xaxis
No discrimination on the yaxis Good discrimination on the xaxis
Best projection: the first NPLDA direction
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 75
Feature Extraction: PCA vs NPLDA
No discrimination on the yaxis Poor discrimination with very
strong overlap on the xaxis
No discrimination on the yaxis Good discrimination on the xaxis
Best projection: the first NPLDA direction
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 76
Feature Extraction: PCA vs NPLDA
No discrimination on the yaxis Poor discrimination with very
strong overlap on the xaxis
No discrimination on the yaxis Good discrimination on the xaxis
Best projection: the first NPLDA direction
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 77
Classification
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 78
Classification: Fuzzy kNN Fuzzy kNearest Neighbors (Keller, 1985)
A variation of the classic kNN kNN assigns a sample to the class of the k closest samples in the
training set It has been demonstrated that the kNN is formally a non
parametric approximation of the Maximum a Posteriori MAP criterion
The asymptotic performance is never higher than the double of the Bayesian error (with k=1)
Critical aspect: choice of value k (with a limited number of samples)
too large: the problem is too much simplified and the local information loses its relevance
too small: density estimation too sensitive to outliers
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 79
Classification: Fuzzy kNN Fuzzy kNearest Neighbors (Keller, 1985)
A variation of the classic kNN kNN assigns a sample to the class of the k closest samples in the
training set It has been demonstrated that the kNN is formally a non
parametric approximation of the Maximum a Posteriori MAP criterion
The asymptotic performance is never higher than the double of the Bayesian error (with k=1)
Critical aspect: choice of value k (with a limited number of samples)
too large: the problem is too much simplified and the local information loses its relevance
too small: density estimation too sensitive to outliers
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 80
Classification: Fuzzy kNN Fuzzy kNearest Neighbors (Keller, 1985)
A variation of the classic kNN kNN assigns a sample to the class of the k closest samples in the
training set It has been demonstrated that the kNN is formally a non
parametric approximation of the Maximum a Posteriori MAP criterion
The asymptotic performance is never higher than the double of the Bayesian error (with k=1)
Critical aspect: choice of value k (with a limited number of samples)
too large: the problem is too much simplified and the local information loses its relevance
too small: density estimation too sensitive to outliers
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 81
Classification: Fuzzy kNN Fuzzy kNN assigns a fuzzy class membership
function to each sample Neighbors are weighted according to their membership
function to each class The provided output will be a fuzzy vector that will
express the membership value of the unknown sample to each class
The fuzzy output can be used to investigate the possibility of early detection and to understand the gravity of the disease
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 82
Classification: Fuzzy kNN Fuzzy kNN assigns a fuzzy class membership
function to each sample Neighbors are weighted according to their membership
function to each class The provided output will be a fuzzy vector that will
express the membership value of the unknown sample to each class
The fuzzy output can be used to investigate the possibility of early detection and to understand the gravity of the disease
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 83
Classification: Fuzzy kNN Fuzzy kNN assigns a fuzzy class membership
function to each sample Neighbors are weighted according to their membership
function to each class The provided output will be a fuzzy vector that will
express the membership value of the unknown sample to each class
The fuzzy output can be used to investigate the possibility of early detection and to understand the gravity of the disease
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 84
Classification: Fuzzy kNN Fuzzy kNN assigns a fuzzy class membership
function to each sample Neighbors are weighted according to their membership
function to each class The provided output will be a fuzzy vector that will
express the membership value of the unknown sample to each class
The fuzzy output can be used to investigate the possibility of early detection and to understand the gravity of the disease
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 85
Classification: Fuzzy kNN The membership value of unlabeled sample x to ith class is
given by the following formula:
where μij represents the membership of labeled x
j to the ith class
The parameter m determines how heavily the distance is weighted when calculating each neighbor's contribution to membership value
μij can be crisp or calculated according to a fuzzy rule
We defined a fuzzy triangular membership function with maximum value at the average of the class and null outside the minimum and maximum value (multiplied by a constant)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 86
Classification: Fuzzy kNN The membership value of unlabeled sample x to ith class is
given by the following formula:
where μij represents the membership of labeled x
j to the ith class
The parameter m determines how heavily the distance is weighted when calculating each neighbor's contribution to membership value
μij can be crisp or calculated according to a fuzzy rule
We defined a fuzzy triangular membership function with maximum value at the average of the class and null outside the minimum and maximum value (multiplied by a constant)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 87
Classification: Fuzzy kNN The membership value of unlabeled sample x to ith class is
given by the following formula:
where μij represents the membership of labeled x
j to the ith class
The parameter m determines how heavily the distance is weighted when calculating each neighbor's contribution to membership value
μij can be crisp or calculated according to a fuzzy rule
We defined a fuzzy triangular membership function with maximum value at the average of the class and null outside the minimum and maximum value (multiplied by a constant)
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 88
CrossValidation Considering the not so big dimension of population and that for each
person we had two measures, we opted for a modified LeaveOneOutapproach: Each test set is composed by the pair of measurements
corresponding to the same person, instead of a single measure as would have been in the normal leaveoneout method
Doing this way we avoided that one of these two measures could belong to the training set, while using the other one in the test set
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 89
CrossValidation Considering the not so big dimension of population and that for each
person we had two measures, we opted for a modified LeaveOneOutapproach: Each test set is composed by the pair of measurements
corresponding to the same person, instead of a single measure as would have been in the normal leaveoneout method
Doing this way we avoided that one of these two measures could belong to the training set, while using the other one in the test set
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 90
Results: Confusion Matrix Performance have been evaluated through confusion matrix and
corresponding performance index We considered different values for k (k=1,3,5,9,101)
Fuzzy kNN demonstrated to be robust to these changes, keeping its results invariant
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 91
Results: Confusion Matrix Performance have been evaluated through confusion matrix and
corresponding performance index We considered different values for k (k=1,3,5,9,101)
Fuzzy kNN demonstrated to be robust to these changes, keeping its results invariant
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 92
Results: Confusion Matrix Performance have been evaluated through confusion matrix and
corresponding performance index We considered different values for k (k=1,3,5,9,101)
Fuzzy kNN demonstrated to be robust to these changes, keeping its results invariant
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 93
Results: Confusion Matrix Performance have been evaluated through confusion matrix and
corresponding performance index We considered different values for k (k=1,3,5,9,101)
Fuzzy kNN demonstrated to be robust to these changes, keeping its results invariant
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 94
Results: Performance Indexes
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 95
Results: Performance Indexes
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 96
Comparison to other classifiers In order to prove the effectiveness of Fuzzy kNN for the considered
problem, we evaluated other families of classifiers: Classic kNN Linear and Quadratic classifiers Feedforward Artificial Neural Network
sigmoidal activation function one hidden layer with 3 neurons
All obtained results were comparable or worst than those achieved by Fuzzy kNN
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 97
Comparison to other classifiers In order to prove the effectiveness of Fuzzy kNN for the considered
problem, we evaluated other families of classifiers: Classic kNN Linear and Quadratic classifiers Feedforward Artificial Neural Network
sigmoidal activation function one hidden layer with 3 neurons
All obtained results were comparable or worst than those achieved by Fuzzy kNN
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 98
Comparison to other classifiers
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 99
Comparison to other classifiers Performing a student ttest between all pairs of classifiers, no
relevant differences emerged All implemented classifiers result comparable for the considered
problem The robustness showed by Fuzzy kNN to k changes is not
verified in the Classic kNN, that leads to different results according to the value of k
Moreover the output given by Fuzzy kNN can be used to extract informations about the degree of membership to the assigned class In further works, this information could be used to understand
the relation between the given pattern and the stage of the disease
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 100
Comparison to other classifiers Performing a student ttest between all pairs of classifiers, no
relevant differences emerged All implemented classifiers result comparable for the considered
problem The robustness showed by Fuzzy kNN to k changes is not
verified in the Classic kNN, that leads to different results according to the value of k
Moreover the output given by Fuzzy kNN can be used to extract informations about the degree of membership to the assigned class In further works, this information could be used to understand
the relation between the given pattern and the stage of the disease
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 101
Comparison to other classifiers Performing a student ttest between all pairs of classifiers, no
relevant differences emerged All implemented classifiers result comparable for the considered
problem The robustness showed by Fuzzy kNN to k changes is not
verified in the Classic kNN, that leads to different results according to the value of k
Moreover the output given by Fuzzy kNN can be used to extract informations about the degree of membership to the assigned class In further works, this information could be used to understand
the relation between the given pattern and the stage of the disease
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 102
Comparison to other classifiers Performing a student ttest between all pairs of classifiers, no
relevant differences emerged All implemented classifiers result comparable for the considered
problem The robustness showed by Fuzzy kNN to k changes is not
verified in the Classic kNN, that leads to different results according to the value of k
Moreover the output given by Fuzzy kNN can be used to extract informations about the degree of membership to the assigned class In further works, this information could be used to understand
the relation between the given pattern and the stage of the disease
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 103
Comparison to the state of the art methods In literature there are three main research works on lung cancer diagnosis by an enose
Sensors 8 QMB sensors 32 Polymer sensors Virtual SAW MOS
Classifier PLSDA PCADA+SVM BPANN
Population 62 volunteers 76 volunteers 52 volunteers 101 volunteers
Test Set 1 sample * 62 times 10 volunteers 2 sample*101 times
Validation Leave1Out ABSENT ABSENT Modified Leave1OutAccuracy 90.3% 88.1% 80.0% 92.6%Sensitivity 100.0% 71.4% 80.0% 95.3%
Specificity 91.9% 80.0% 90.5%
100.0% 66.7% 100.0% 88.2%
93.4% 100.0% 96.3%
Tor Vergata University(1) Lerner Research Institue(2) Zhejang University(3)
Politecnico di Milano
FSNPLDAFuzzy kNN
76 volunteers (59 of them used in the training set+17 new)
44.4% (w.r.t. diseased after surgery)94.4% (w.r.t. healthy people)
PrecisionPOS
PrecisionNEG
80.0% (w.r.t. diseased after surgery)
77.27% (w.r.t. healthy people)
American and Chinese researches did not consider any crossvalidation technique Tor Vergata group considered 3 classes (Healthy, Diseased
before surgery and Diseased
after surgery):
No medical foundation of it How would classifier perform merging the two diseasedsubclasses?
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 104
Comparison to the state of the art methods In literature there are three main research works on lung cancer diagnosis by an enose
Sensors 8 QMB sensors 32 Polymer sensors Virtual SAW MOS
Classifier PLSDA PCADA+SVM BPANN
Population 62 volunteers 76 volunteers 52 volunteers 101 volunteers
Test Set 1 sample * 62 times 10 volunteers 2 sample*101 times
Validation Leave1Out ABSENT ABSENT Modified Leave1OutAccuracy 90.3% 88.1% 80.0% 92.6%Sensitivity 100.0% 71.4% 80.0% 95.3%
Specificity 91.9% 80.0% 90.5%
100.0% 66.7% 100.0% 88.2%
93.4% 100.0% 96.3%
Tor Vergata University(1) Lerner Research Institue(2) Zhejang University(3)
Politecnico di Milano
FSNPLDAFuzzy kNN
76 volunteers (59 of them used in the training set+17 new)
44.4% (w.r.t. diseased after surgery)
94.4% (w.r.t. healthy people)
PrecisionPOS
PrecisionNEG
80.0% (w.r.t. diseased after surgery)
77.27% (w.r.t. healthy people)
American and Chinese researches did not consider any crossvalidation technique Tor Vergata group considered 3 classes (Healthy, Diseased
before surgery and Diseased
after surgery):
No medical foundation of it How would classifier perform merging the two diseasedsubclasses?
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 105
Comparison to the state of the art methods In literature there are three main research works on lung cancer diagnosis by an enose
Sensors 8 QMB sensors 32 Polymer sensors Virtual SAW MOS
Classifier PLSDA PCADA+SVM BPANN
Population 62 volunteers 76 volunteers 52 volunteers 101 volunteers
Test Set 1 sample * 62 times 10 volunteers 2 sample*101 times
Validation Leave1Out ABSENT ABSENT Modified Leave1OutAccuracy 90.3% 88.1% 80.0% 92.6%Sensitivity 100.0% 71.4% 80.0% 95.3%
Specificity 91.9% 80.0% 90.5%
100.0% 66.7% 100.0% 88.2%
93.4% 100.0% 96.3%
Tor Vergata University(1) Lerner Research Institue(2) Zhejang University(3)
Politecnico di Milano
FSNPLDAFuzzy kNN
76 volunteers (59 of them used in the training set+17 new)
44.4% (w.r.t. diseased after surgery)
94.4% (w.r.t. healthy people)
PrecisionPOS
PrecisionNEG
80.0% (w.r.t. diseased after surgery)
77.27% (w.r.t. healthy people)
American and Chinese researches did not consider any crossvalidation technique Tor Vergata group considered 3 classes (Healthy, Diseased
before surgery and Diseased
after surgery):
No medical foundation of it How would classifier perform merging the two diseasedsubclasses?
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 106
Comparison to current diagnostic methods The use of an electronic nose as lung cancer diagnostic tool is
reasonable if it gives some advantages compared to current diagnostic techniques
Electronic nose results better in terms of performance Electronic noses are cheaper, more robust, smaller (and thus eventually
portable), very fast and non invasive instruments
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 107
Comparison to current diagnostic methods The use of an electronic nose as lung cancer diagnostic tool is
reasonable if it gives some advantages compared to current diagnostic techniques
Electronic nose results better in terms of performance Electronic noses are cheaper, more robust, smaller (and thus eventually
portable), very fast and non invasive instruments
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 108
Comparison to current diagnostic methods The use of an electronic nose as lung cancer diagnostic tool is
reasonable if it gives some advantages compared to current diagnostic techniques
Electronic nose results better in terms of performance Electronic noses are cheaper, more robust, smaller (and thus eventually
portable), very fast and non invasive instruments
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 109
Extension and further direction of research The work could be extended into three parallel directions:
1. Improvement of Sensors Technology:• Development of longerlyfe and more stable sensors• Development of hybrid systems, able to provide both selective
and sensitive abilities 2. Improvement of Olfactory Signal Analysis and Classification
Algorithms3. Medical Analysis
• According to the scientific literature, there are no studies on the variation of VOCs in the breath before and after the surgery: this suggest as interesting the analysis, by an electronic nose, of the patterns' changes due to surgery
• Individuation of risk factors connected to lung cancer• Involving a larger population, partitioning it according to
different stages and using the Fuzzy output information, it would be possible to study the possibility of early diagnose
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 110
Extension and further direction of research The work could be extended into three parallel directions:
1. Improvement of Sensors Technology:• Development of longerlyfe and more stable sensors• Development of hybrid systems, able to provide both selective
and sensitive abilities 2. Improvement of Olfactory Signal Analysis and Classification
Algorithms3. Medical Analysis
• According to the scientific literature, there are no studies on the variation of VOCs in the breath before and after the surgery: this suggest as interesting the analysis, by an electronic nose, of the patterns' changes due to surgery
• Individuation of risk factors connected to lung cancer• Involving a larger population, partitioning it according to
different stages and using the Fuzzy output information, it would be possible to study the possibility of early diagnose
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 111
Extension and further direction of research The work could be extended into three parallel directions:
1. Improvement of Sensors Technology:• Development of longerlyfe and more stable sensors• Development of hybrid systems, able to provide both selective
and sensitive abilities 2. Improvement of Olfactory Signal Analysis and Classification
Algorithms3. Medical Analysis
• According to the scientific literature, there are no studies on the variation of VOCs in the breath before and after the surgery: this suggest as interesting the analysis, by an electronic nose, of the patterns' changes due to surgery
• Individuation of risk factors connected to lung cancer• Involving a larger population, partitioning it according to
different stages and using the Fuzzy output information, it would be possible to study the possibility of early diagnose
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 112
Extension and further direction of research The work could be extended into three parallel directions:
1. Improvement of Sensors Technology:• Development of longerlyfe and more stable sensors• Development of hybrid systems, able to provide both selective
and sensitive abilities 2. Improvement of Olfactory Signal Analysis and Classification
Algorithms3. Medical Analysis
• According to the scientific literature, there are no studies on the variation of VOCs in the breath before and after the surgery: this suggest as interesting the analysis, by an electronic nose, of the patterns' changes due to surgery
• Individuation of risk factors connected to lung cancer• Involving a larger population, partitioning it according to
different stages and using the Fuzzy output information, it would be possible to study the possibility of early diagnose
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 113
Extension and further direction of research The work could be extended into three parallel directions:
1. Improvement of Sensors Technology:• Development of longerlyfe and more stable sensors• Development of hybrid systems, able to provide both selective
and sensitive abilities 2. Improvement of Olfactory Signal Analysis and Classification
Algorithms3. Medical Analysis
• According to the scientific literature, there are no studies on the variation of VOCs in the breath before and after the surgery: this suggest as interesting the analysis, by an electronic nose, of the patterns' changes due to surgery
• Individuation of risk factors connected to lung cancer• Involving a larger population, partitioning it according to
different stages and using the Fuzzy output information, it would be possible to study the possibility of early diagnose
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 114
Extension and further direction of research The work could be extended into three parallel directions:
1. Improvement of Sensors Technology:• Development of longerlyfe and more stable sensors• Development of hybrid systems, able to provide both selective
and sensitive abilities 2. Improvement of Olfactory Signal Analysis and Classification
Algorithms3. Medical Analysis
• According to the scientific literature, there are no studies on the variation of VOCs in the breath before and after the surgery: this suggest as interesting the analysis, by an electronic nose, of the patterns' changes due to surgery
• Individuation of risk factors connected to lung cancer• Involving a larger population, partitioning it according to
different stages and using the Fuzzy output information, it would be possible to study the possibility of early diagnose
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 115
Extension and further direction of research The work could be extended into three parallel directions:
1. Improvement of Sensors Technology:• Development of longerlyfe and more stable sensors• Development of hybrid systems, able to provide both selective
and sensitive abilities 2. Improvement of Olfactory Signal Analysis and Classification
Algorithms3. Medical Analysis
• According to the scientific literature, there are no studies on the variation of VOCs in the breath before and after the surgery: this suggest as interesting the analysis, by an electronic nose, of the patterns' changes due to surgery
• Individuation of risk factors connected to lung cancer• Involving a larger population, partitioning it according to
different stages and using the Fuzzy output information, it would be possible to study the possibility of early diagnose
Fuzzy kNN Lung Cancer Identification by an Electronic Nose 116
The end
Thanks!