2013 physica

Upload: dorin-pleava

Post on 03-Jun-2018

215 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/12/2019 2013 Physica

    1/5

    This content has been downloaded from IOPscience. Please scroll down to see the full text.

    Download details:

    This content was downloaded by: dramican

    IP Address: 79.101.164.228

    This content was downloaded on 15/11/2013 at 17:43

    Please note that terms and conditions apply.

    Artificial neural networks for processing fluorescence spectroscopy data in skin cancer

    diagnostics

    View the table of contents for this issue, or go to thejournal homepagefor more

    2013 Phys. Scr. 2013 014057

    (http://iopscience.iop.org/1402-4896/2013/T157/014057)

    ome Search Collections Journals About Contact us My IOPscience

    http://localhost/var/www/apps/conversion/tmp/scratch_10/iopscience.iop.org/page/termshttp://iopscience.iop.org/1402-4896/2013/T157http://iopscience.iop.org/1402-4896http://iopscience.iop.org/http://iopscience.iop.org/searchhttp://iopscience.iop.org/collectionshttp://iopscience.iop.org/journalshttp://iopscience.iop.org/page/aboutioppublishinghttp://iopscience.iop.org/contacthttp://iopscience.iop.org/myiopsciencehttp://iopscience.iop.org/myiopsciencehttp://iopscience.iop.org/contacthttp://iopscience.iop.org/page/aboutioppublishinghttp://iopscience.iop.org/journalshttp://iopscience.iop.org/collectionshttp://iopscience.iop.org/searchhttp://iopscience.iop.org/http://iopscience.iop.org/1402-4896http://iopscience.iop.org/1402-4896/2013/T157http://localhost/var/www/apps/conversion/tmp/scratch_10/iopscience.iop.org/page/terms
  • 8/12/2019 2013 Physica

    2/5

    IOP PUBLISHING PHYSICASCRIPTA

    Phys. Scr.T157(2013) 014057 (4pp) doi:10.1088/0031-8949/2013/T157/014057

    Artificial neural networks for processing

    fluorescence spectroscopy data in skincancer diagnostics

    L Lenhardt, I Zekovic, T Dramicanin and M D Dramicanin

    Institute for Nuclear Sciences Vinca, University of Belgrade, PO Box 522, 11001 Belgrade, Serbia

    E-mail:[email protected]

    Received 25 August 2012

    Accepted for publication 14 January 2013

    Published 15 November 2013

    Online atstacks.iop.org/PhysScr/T157/014057

    Abstract

    Over the years various optical spectroscopic techniques have been widely used as diagnostic

    tools in the discrimination of many types of malignant diseases. Recently, synchronous

    fluorescent spectroscopy (SFS) coupled with chemometrics has been applied in cancer

    diagnostics. The SFS method involves simultaneous scanning of both emission and excitation

    wavelengths while keeping the interval of wavelengths (constant-wavelength mode) or

    frequencies (constant-energy mode) between them constant. This method is fast, relatively

    inexpensive, sensitive and non-invasive. Total synchronous fluorescence spectra of normal

    skin, nevus and melanoma samples were used as input for training of artificial neural

    networks. Two different types of artificial neural networks were trained, the self-organizing

    map and the feed-forward neural network. Histopathology results of investigated skin sampleswere used as the gold standard for network output. Based on the obtained classification

    success rate of neural networks, we concluded that both networks provided high sensitivity

    with classification errors between 2 and 4%.

    PACS numbers: 87.19.xj, 87.64.K, 87.64.kv

    1. Introduction

    Skin cancer is one of the most common malignancies

    worldwide. Late detection delivers high mortality rates, but

    if diagnosed in the early stages it is one of the most treatableforms of cancer. Taking this problem into account, developingnew methods for cancer diagnosis is of crucial significance.In the last few decades, in the field of cancer diagnostics,fluorescence spectroscopy has proven to be a very promising

    technique (Alfano et al 1987,Sterenborg et al 1994,Palmeret al2003).

    Human tissues are a complex mixture of differentmolecules and some of these molecules, called endogenousfluorophores, have the ability to absorb and emit light ofdifferent wavelengths. The concentration and distribution ofvarious fluorophores in the skin such as nicotinamide adenine

    dinucleotide (NADH), flavine adenine dinucleotide (FAD),

    keratin, collagen, elastin, amino acids, lipids and porphyrinsare the fundamentals for discrimination between cancer andnormal tissue by fluorescence spectroscopy. In recent years ithas been shown that synchronous fluorescence spectroscopy

    (SFS) can be successfully used in cancer detection (Vo-Dinh2000, Vengadesan et al 2002, Dramicanin et al 2005,2006, 2011). This method involves simultaneous scanningof both emission and excitation wavelengths while keeping

    the interval of wavelengths (constant-wavelength mode) orfrequencies (constant-energy mode) between them constant.The advantage of a synchronous spectrum over ordinaryemission spectra is that it often has more features andprovides more information. The obtained data are usuallyhighly correlated with subtle differences between abnormaland normal tissues, and for this purpose artificial neuralnetworks (ANN) can be very useful for building a diagnosticmodel. ANN is an adaptive system that modifies structurebased on input findings to generate robust output. The benefitof this method is its consistency and objectivity due tolack of human fatigue and bias. In this paper, two differenttypes of ANN were trained, the self-organizing map (SOM)

    and the feed-forward neural network. SOM is one of themost popular neural networks that convert high-dimensionalnonlinear statistical relationships into simple geometricrelationships in an unsupervised way (Kohonen 2001).

    0031-8949/13/014057+04$33.00 1 2013 The Royal Swedish Academy of Sciences Printed in the UK

    http://dx.doi.org/10.1088/0031-8949/2013/T157/014057mailto:[email protected]://stacks.iop.org/PhysScr/T157/014057http://stacks.iop.org/PhysScr/T157/014057mailto:[email protected]://dx.doi.org/10.1088/0031-8949/2013/T157/014057
  • 8/12/2019 2013 Physica

    3/5

    Phys. Scr.T157(2013) 014057 L Lenhardtet al

    Figure 1. Total synchronous fluorescence spectra of different skin sample types: (a) melanoma, (b) nevus and (c) normal skin.

    Unlike SOM, the feed-forward neural network is trained in

    a supervised way with prior knowledge of a samples group

    membership by data entering at the inputs passing through

    layers of neurons until it arrives at the outputs without any

    feedback information between layers (Jiang et al 2003).

    ANN are often trained to be further used as classification

    models (Rhee et al 2005, Dramicanin et al 2009). The aim

    of this work was to investigate the possibility of building a

    skin cancer diagnostic method by training ANN using total

    fluorescence spectra of different types of skin lesions.

    2. Materials and methods

    Synchronous fluorescence spectra of 50 skin samples (12

    melanoma, 18 nevus and 20 normal skin samples) were

    measured ex vivo using a fluorescence spectrophotometer

    (Perkin Elmer LS45) in constant-wavelength mode. The

    spectra were collected at a scan rate of 200 nm min1 in the

    excitation range from 330 to 550 nm and the wavelength

    offset range from 30 to 120 nm and automatically

    normalized to excitation by the instrument. Samples were

    obtained from human patients soon after surgical resection,

    histopathologically identified and then measured by the

    instrument at room temperature. In order to reduce the

    dimensionality of obtained spectra, principal component

    analysis (PCA) was applied. For training, calculated PCA

    components were introduced in SOM and the feed-forward

    neural network. To assess classification error, they were tested

    with new samples unknown to both ANN.

    3. Results and discussion

    Figure 1 presents total synchronous fluorescence spectra in

    the form of contour diagrams of different skin samples:

    normal skin, nevus and melanoma. Emission patterns

    reflect the specificity of endogenous fluorophores and their

    microenvironments in skin. It can be clearly seen that the main

    differences in the fluorescence of skin lesions are observable

    in several spectral regions. The first region spans the excitation

    wavelength from 330 to 400 nm and the synchronous interval

    from 30 to 55 nm. Structural proteins of the extracellular

    matrix, collagen and elastin have maximum excitation in this

    region (Ramanujam 2000). The second and third spectralregions cover the excitation wavelength interval from 430 to

    480 nm, 350 to 400nm and the synchronous interval from

    30 to 50 nm, 65 to 90 nm, respectively. Fluorescence of

    Figure 2. Synchronous fluorescence spectra taken at = 70nmof melanoma (full line), normal skin (dash line) and nevus (dot line)samples.

    these regions originates from several fluorophores like the

    co-enzymes NADH and FAD.A skin cancer diagnostic method based on measurement

    of SFS and a classification model built using two different

    ANN was developed in several steps. First reduction of

    dimensionality was necessary due to the large size of gathered

    SFS. For that purpose PCA was applied on SFS data

    and a five-component PCA model was acquired. PCA is

    an unsupervised statistical method used to distinguish and

    identify patterns in data and express them in such a manner

    as to point out their similarities and differences. This method

    transforms a number of possibly correlated variables into

    a smaller number of uncorrelated variables called principal

    components, which account for most of the variance in theobserved variables. This allowed us to explore our data and

    to determine for which wavelength offset value () best

    separation between groups was achieved. After examining the

    results of PCA, it was concluded that the best differentiation

    between classes is for = 70 nm, figure2.In figure3 one

    can see an obvious distinction between classes. Taking this

    into consideration, 25 samples PCA scores for = 70nm

    were used to train SOM and the feed-forward neural network.

    In order to enlarge the number of samples for ANN training,

    we created a normally distributed set of data based on the

    mean value and standard deviation of 25 samples used for

    training. With a new data set of 2000 samples we trained both

    ANN.SOM learns to group data based on similarity and

    topology, and as a result it assigns the same indices to each

    class. It is used for reduction of dimensionality and clustering

    2

  • 8/12/2019 2013 Physica

    4/5

    Phys. Scr.T157(2013) 014057 L Lenhardtet al

    Figure 3. PCA scores obtained from SFS data taken at = 70nmof melanoma, normal skin and nevus samples.

    Figure 4. Schematic representation of the SOM architecture.

    data. We introduced a 2000 5 matrix (2000 samples and 5

    parameters describing every sample) to SOM with 3 neurons

    (size 1 3), figure4,and started training with 150 iterations.

    Through all this, SOM did not have any information about the

    samples group membership (unsupervised learning). When

    trained, SOM was tested with 25 new samples unknown to

    the network and the classification error was acquired. This

    process of SOM training and testing was repeated several

    times, every time using a different combination of data for

    training and testing. Doing this we obtained a more accurate

    efficiency overview of this method.

    Feed-forward networks, figure5,contain series of layers.

    The first layer gets network input data, and each subsequent

    layer is connected to the previous layer while the final layer

    gives network output. They are usually used for input to output

    mapping. We used the feed-forward network with one hidden

    layer to build a classification model by giving it as input the

    2000 5 matrix, the same matrix used for SOM training, and

    the matrix with information of class membership of all 2000

    samples (supervised learning). The process of testing was the

    same as for SOM; several models were built and tested using

    different combinations of data, and the mean classification

    error for all models was calculated.

    Based on the obtained classification errors of neuralnetworks, we found that both networks provide high

    sensitivity with classification errors between 2 and 4%. For

    different input data the SOM classification error was between

    Figure 5. Schematic representation of the feed-forward networkarchitecture.

    2 and 3%, while for the feed-forward network it was between

    3 and 4%.

    4. Conclusion

    In this work, we used statistical analysis and ANN to classify

    data of synchronous fluorescence spectra with the aim of

    developing a diagnostic method for skin cancer diagnosis. The

    fluorescence spectra of three different types of skin lesions,

    melanoma, nevus and normal skin, revealed differences

    between them due to differences in concentration of various

    fluorophores and their microenvironments. The application of

    PCA provided data with enlarged variances between sample

    groups. It is shown that both SOM and feed-forward neural

    networks gave promising results with a 9698% success tissueclassification rate. Moreover, the presented method is fast,

    sensitive, inexpensive and non-destructive. Based on these

    findings, we can conclude that SFS combined with neural

    network-based classification has good potential in melanoma

    diagnostics.

    Acknowledgment

    This work was supported by the Serbian Ministry of Science

    and Technological Development (project numbers 173049 and

    45020).

    References

    Alfano R R, Tang G C, Pradhan A, Lam W, Choy D S J and Opher E1987 Fluorescence spectra from cancerous and normal humanbreast and lung tissuesIEEE J. Quantum Electron. 23180611

    Dramicanin T, Dimitrijevic B and Dramicanin M D 2011Application of supervised self-organising maps in breastcancer diagnosis by total synchronous luminescencespectroscopyAppl. Spectrosc.652937

    Dramicanin T, Dramicanin M D, Dimitrijevic B, Jokanovic V andLukic S 2006 Discrimination between normal and malignantbreast tissues by synchronous luminescence spectroscopyActaChim. Slov.534449

    Dramicanin T, Dramicanin M D, Jokanovic V,Nikolic-Vukosavljevic D and Dimitrijevic B 2005Three-dimensional total synchronous luminescencespectroscopy criteria for discrimination between normal andmalignant breast tissuesPhotochem. Photobiol.8115548

    3

    http://dx.doi.org/10.1109/JQE.1987.1073234http://dx.doi.org/10.1109/JQE.1987.1073234http://dx.doi.org/10.1366/10-05928http://dx.doi.org/10.1366/10-05928http://dx.doi.org/10.1562/2005-02-15-RA-442http://dx.doi.org/10.1562/2005-02-15-RA-442http://dx.doi.org/10.1562/2005-02-15-RA-442http://dx.doi.org/10.1366/10-05928http://dx.doi.org/10.1109/JQE.1987.1073234
  • 8/12/2019 2013 Physica

    5/5

    Phys. Scr.T157(2013) 014057 L Lenhardtet al

    Dramicanin T, Zekovic I, Dimitrijevic B, Ribar S andDramicanin M D 2009 Optical biopsy method for breastcancer diagnosis based on artificial neural networkclassification of fluorescence landscape dataActa Phys. Pol. A1166902

    Jiang X and Wah A H K S 2003 Constructing and trainingfeed-forward neural networks for pattern classificationPatternRecognit.3685367

    Kohonen T 2001Self-Organizing Maps3rd edn (Berlin: Springer)Palmer G M, Keely P J, Breslin T M and Ramanujam N 2003

    Autofluorescence spectroscopy of normal and malignanthuman breast cell linesPhotochem. Photobiol.784629

    Ramanujam N 2000Encyclopedia of Analytical Chemistry(Chichester, NY: Wiley) pp 2056

    Rhee J I, Lee K I, Kim C K, Yim Y S, Chung S W, Wei J andBellgardt K H 2005 Classification of two-dimensionalfluorescence spectra using self-organizing mapsBiochem. Eng. J.2213544

    Sterenborg H J C M, Motamedi M, Wagner R F, Duvic M,Thomsen S and Jacques S L 1994 In vivofluorescencespectroscopy and imaging of human skin tumours Laser Med.Sci.9191201

    Vengadesan N, Anbupalam T, Hemamalini S, Ebenezar J,Muthvelu K, Koteeswaran D, Aruna P R and Ganesan S C2002 Characterization of cervical normal and abnormal tissuesby synchronous luminescence spectroscopyProc. SPIE4613137

    Vo-Dinh T 2000 Principle of synchronous luminescence (SL)technique for biomedical diagnosticsProc. SPIE3911429

    4

    http://dx.doi.org/10.1016/S0031-3203(02)00087-0http://dx.doi.org/10.1016/S0031-3203(02)00087-0http://dx.doi.org/10.1562/0031-8655(2003)078%3C0462:ASONAM%3E2.0.CO;2http://dx.doi.org/10.1562/0031-8655(2003)078%3C0462:ASONAM%3E2.0.CO;2http://dx.doi.org/10.1016/j.bej.2004.09.008http://dx.doi.org/10.1016/j.bej.2004.09.008http://dx.doi.org/10.1007/BF02590223http://dx.doi.org/10.1007/BF02590223http://dx.doi.org/10.1007/BF02590223http://dx.doi.org/10.1016/j.bej.2004.09.008http://dx.doi.org/10.1562/0031-8655(2003)078%3C0462:ASONAM%3E2.0.CO;2http://dx.doi.org/10.1016/S0031-3203(02)00087-0