{"title": "Spectroscopic Detection of Cervical Pre-Cancer through Radial Basis Function Networks", "book": "Advances in Neural Information Processing Systems", "page_first": 981, "page_last": 987, "abstract": null, "full_text": "Spectroscopic Detection of Cervical \n\nPre-Cancer through Radial  Basis \n\nFunction Networks \n\nKagan Tumer \n\nkagan@pine.ece.utexas.edu \n\nNirmala Ramanujam \nnimmi@ccwf.cc.utexas.edu \n\nDept.  of Electrical and Computer Engr. \n\nThe University of Texas at Austin, \n\nBiomedical  Engineering Program \nThe University of Texas at Austin \n\nRebecca Richards-Kortum \n\nkortum@mail.utexas.edu \n\nJoydeep  Ghosh \n\nghosh@ece.utexas.edu \n\nBiomedical Engineering Program \nThe University of Texas at Austin \n\nDept.  of Electrical and  Computer Engr. \n\nThe University of Texas at Austin \n\nAbstract \n\nThe  mortality  related  to cervical  cancer  can  be  substantially  re(cid:173)\nduced  through  early  detection  and  treatment.  However,  cur(cid:173)\nrent  detection  techniques,  such  as  Pap  smear  and  colposcopy, \nfail  to  achieve  a  concurrently  high  sensitivity  and  specificity.  In \nvivo  fluorescence  spectroscopy  is  a  technique  which  quickly,  non(cid:173)\ninvasively and quantitatively probes the biochemical and morpho(cid:173)\nlogical  changes that occur in pre-cancerous tissue.  RBF ensemble \nalgorithms based on such spectra provide automated, and near real(cid:173)\ntime implementation of pre-cancer  detection in the  hands of non(cid:173)\nexperts.  The  results  are  more  reliable,  direct  and  accurate  than \nthose achieved  by  either human experts or multivariate statistical \nalgorithms. \n\n1 \n\nIntroduction \n\nCervical  carcinoma is  the  second  most  common  cancer  in  women  worldwide,  ex(cid:173)\nceeded  only by  breast cancer  (Ramanujam et al.,  1996).  The mortality related to \ncervical cancer can be reduced if this disease is  detected at the pre-cancerous state, \nknown  as  squamous intraepitheliallesion (SIL).  Currently, a  Pap smear is  used  to \n\n\f982 \n\nK. Turner,  N.  Ramanujam, R.  Richards-Kortum and J.  Ghosh \n\nscreen for  cervical  cancer {Kurman et al.,  1994}.  In a  Pap test,  a  large number of \ncells  obtained  by scraping the cervical epithelium  are  smeared onto a  slide  which \nis  then  fixed  and stained  for  cytologic  examination.  The  Pap smear  is  unable  to \nachieve a  concurrently high sensitivityl  and high specificity2  due to both sampling \nand  reading  errors  (Fahey  et  al.,  1995).  Furthermore,  reading  Pap  smears  is  ex(cid:173)\ntremely  labor  intensive  and requires  highly  trained  professionals.  A  patient with \na  Pap smear  interpreted  as  indicating the presence of SIL  is  followed  up  by  a  di(cid:173)\nagnostic  procedure called  colposcopy.  Since  this  procedure involves  biopsy,  which \nrequires histologic evaluation, diagnosis is not immediate. \nIn vivo fluorescence spectroscopy is a technique which has the capability to quickly, \nnon-invasively and quantitatively probe the biochemical and morphological changes \nthat occur as  tissue becomes neoplastic.  The measured spectral information can be \ncorrelated to tissue histo-pathology, the current \"gold standard\" to develop clinically \neffective  screening and  diagnostic  algorithms.  These mathematical  algorithms can \nbe  implemented  in  software  thereby,  enabling  automated,  fast,  non-invasive  and \naccurate pre-cancer screening and diagnosis in hands of non-experts. \n\nA screening and diagnostic technique for  human cervical pre-cancer based on laser \ninduced fluorescence spectroscopy has been developed  recently  (Ramanujam et al., \n1996).  Screening and diagnosis was  achieved  using a  multivariate statistical  algo(cid:173)\nrithm  (MSA)  based on principal component analysis and logistic discrimination of \ntissue  spectra acquired  in  vivo.  Furthermore,  we  designed  Radial  Basis  Function \n(RBF)  network  ensembles  to  improve  the  accuracy  of the  multivariate  statistical \nalgorithm, and to simplify the decision making process.  Section 2 presents the data \ncollection/processing techniques.  In Section  3,  we  discuss  the MSA,  and  describe \nthe neural network based methods.  Section 4 contains the experimental results and \ncompares the neural network results to both the results of the MSA  and to current \nclinical detection methods.  A discussion of the results is given  in Section 5. \n\n2  Data Collection and Processing \n\nA  portable fluorimeter  consisting of two nitrogen pumped-dye lasers,  a  fiber-optic \nprobe and a  polychromator coupled  to an intensified  diode  array controlled  by  an \noptical multi-channel analyzer was utilized to measure fluorescence spectra from the \ncervix  in  vivo  at three excitation wavelengths:  337,  380  and 460  nm  (Ramanujam \net al.,  1996).  Tissue biopsies were obtained only from  abnormal sites identified  by \ncolposcopy and subsequently analyzed by the probe to comply with routine patient \ncare  procedure.  Hemotoxylin  and  eosin  stained  sections  of each  biopsy  specimen \nwere  evaluated  by  a  panel  of four  board  certified  pathologists  and  a  consensus \ndiagnosis was  established  using  the  Bethesda classification  system.  Samples  were \nclassified  as  normal  squamous  (NS),  normal  columnar  (NC),  low  grade  (LG)  SIL \nand high grade (HG)  SIL.  Table  1 provides the number of samples in  the training \n(calibration)  and  test  sets.  Based  on  this  data set,  a  clinically  useful  algorithm \nneeds to discriminate SILs from  the normal tissue types. \n\nFigure 1 illustrates average fluorescence spectra per site acquired from cervical sites \nat 337 nm excitation from a typical patient.  Evaluation of the spectra at 337 nm ex-\n\nlSensitivity is the correct classification percentage on the pre-cancerous tissue samples. \n2Specificity is the correct classification  percentage on  normal tissue samples. \n\n\fSpectroscopic Detection of Cervical Pre-cancer through RBF Networks \n\n983 \n\nTable  1:  Histo-pathologic classification of samples. \n\nHisto-pathology \n\n1faining Set \n\nTest Set \n\nNonnal \n\nSIL \n\n107  (SN:  94;  SC:  13) \n58  (LG:  23;  HG:  35) \n\n108  (SN:  94;  SC:  14) \n59  (LG:  24;  HG:  35) \n\ncitation highlights one of the classification difficulties,  namely that the fluorescence \nintensity of SILs  (LG  and HG)  is  less than that of the corresponding normal squa(cid:173)\nmous tissue and greater than that of the corresponding normal columnar tissue over \nthe  entire emission  spectrum3 .  Fluorescence  spectra at all  three excitation  wave(cid:173)\nlengths comprise of a  total of 161  excitation-emission wavelengths pairs.  However, \nthere is a significant cost penalty for  using all  161  values.  To alleviate this concern, \na  more cost-effective fluorescence  imaging system was  developed,  using component \nloadings  calculated  from  principal  component  analysis.  Thus,  the  number  of re(cid:173)\nquired  fluorescence excitation-emission wavelength pairs were reduced  from  161  to \n13  with a  minimal  drop in classification accuracy (Ramanujam et al.,  1996). \n\n0.50  ,----,-----,-----,----,-----.-----,-----.----, \n\n0.30 \n\n.~  0.40 \nI/) c \n<I> \nC \n<I> g \n<I> o \nI/) \n.... \n<I> \no \n:::l \nII \n\n0.20 \n\n0.10 \n\nNC \nNS \nHG \nLG \n\n0.00  <--__ --'-=.:=------'-____ ---l... __  --L-__  -'--__  ~ __  L...-_---' \n\n300.0 \n\n400.0 \n\n500.0 \n\n600.0 \n\n700.0 \n\nWavelength (nm) \n\nFigure  1:  Fluorecsence spectra from  a  typical patient at 337 nm excitation. \n\n3  Algorithm Development \n\n3.1  Multivariate Statistical Algorithms \n\nThe multivariate statistical algorithm development described in (Ramanujam et al., \n1996)  consists of the following  five  steps:  (1)  pre-processing to reduce inter-patient \nand  intra-patient variation of spectra from  a  tissue  type,  (2)  dimension  reduction \nof  the  pre-processed  tissue  spectra  using  Principal  Component  Analysis  (PCA), \n(3)  selection of diagnostically relevant  principal components,  (4)  development of a \nclassification algorithm based on logistic discrimination, and finally (5) retrospective \nand  prospective evaluation of the  algorithm's accuracy on  a  training  (calibration) \nand  test  (prediction)  set,  respectively.  Discrimination  between  SILs  and  the  two \nnormal  tissue  types  could  not  be  achieved  effectively  using  MSA.  Therefore  two \n\n3Spectral  features observed in  Figure 1 are representative of those measured at 380 nm \n\nand 460  nm excitation  (not shown  here). \n\n\f984 \n\nK.  Tumer,  N.  Ramanujam, R.  Richards-Kortum and 1. Ghosh \n\nconstituent algorithms were developed:  algorithm (1), to discriminate between SILs \nand normal squamous tissues, and algorithm  (2), to discriminate between SILs and \nnormal columnar tissues  (Ramanujam et al.,  1996). \n\n3.2  Algorithms based on Neural Networks \n\nThe second stage of algorithm development consists of evaluating the applicability of \nneural networks to this problem.  Initially,  both Multi-Layered Perceptrons (MLPs) \nand Radial  Basis function  (RBF)  networks were considered.  However, MLPs failed \nto improve upon  the MSA  results for  both algorithms  (1)  and  (2),  and frequently \nconverged to spurious solutions.  Therefore, our study focuses on RBF networks and \nRBF network ensembles. \n\nRadial Basis Function Networks:  The first  step in applying  RBF networks to \nthis problem consisted of retracing the two-step process outlined for the multivariate \nstatistical algorithm.  For constituent algorithm (1) the kernels were initialized using \na  k-means  clustering  algorithm  on  the  training  set  containing  NS  tissue  samples \nand  SILs.  The  RBF  networks  had  10  kernels,  whose  locations  and  spreads  were \nadjusted during training.  For  constituent algorithm (2), we selected 10 kernels, half \nof which were fixed  to patterns from  the columnar normal class, while the other half \nwere  initialized  using  a  k-means algorithm.  Neither  the kernel  locations nor  their \nspreads were adjusted during training.  This process was adopted to rectify the large \ndiscrepancy  between  the samples from  each category  (13  for  columnar normal  vs. \n58  for  SILs).  For each algorithm,  the  training time  was  estimated  by  maximizing \nthe performance on one validation set.  Once the stopping time was established,  20 \ncases were run for  each algorithm4. \n\nLinear  and  Order  statistics  Combiners:  There  were  significant  variations \namong  different  runs  of  the  RBF  networks  for  all  three  algorithms.  Therefore, \nselecting  the  \"best\"  classifier  was  not  the  ideal  choice.  First,  the  definition  of \n\"best\"  depends on the selection of the validation set, making it difficult to ascertain \nwhether one network will outperform all others given a different test set, as the val(cid:173)\nidation sets are small.  Second, selecting only one classifier discards a large amount \nof potentially relevant  information.  In  order  to use  all  the  available  data,  and  to \nincrease  both the  performance and  the  reliability  of the  methods,  the  outputs  of \nRBF  networks were pooled before a classification decision was  made. \n\nThe concept of combining  classifier outputs5  has  been  explored  in  a  multitude of \narticles  (Hansen  and  Salamon,  1990;  Wolpert,  1992).  In  this  article  we  use  the \nmedian combiner, which  belongs  to the class order statistics combiners introduced \nin (Turner and Ghosh,  1995), and the averaging combiner, which performs an arith(cid:173)\nmetic average of the corresponding outputs. \n\n4  Results \n\nTwo-step algorithm:  The ensemble results reported are based on pooling 20 dif(cid:173)\nferent  runs  of RBF  networks,  initialized  and trained  as  described  in  the previous \nsection.  This  procedure  was  repeated  10  times  to ascertain  the  reliability  of the \n\n4Each  run has a different initialization set of kernels/spreads/weights. \n5 An extensive bibliography is available in  (Turner and Ghosh,  1996). \n\n\fSpectroscopic Detection of Cervical Pre-cancer through RBF Networks \n\n985 \n\nresult and to obtain the standard deviations.  For an application such as pre-cancer \ndetection,  the  cost of a  misclassification  varies  greatly from  one  class  to another. \nErroneously labeling  a  healthy  tissue as  pre-cancerous can be  corrected  when  fur(cid:173)\nther tests  are performed.  Labeling a  pre-cancerous tissue as  healthy however,  can \nlead to disastrous consequences.  Therefore, for algorithm (1), we have increased the \ncost  of a  misclassified  SIL  until  the sensitivity6  reached  a  satisfactory  level.  The \nsensitivity and specificity values for  constituent algorithm  (1)  based on both MSA \nand RBF ensembles are provided in Table 2.  Table 3 presents sensitivity and speci(cid:173)\nficity values for  constituent algorithm (2) obtained from  MSA and RBF ensembles 7 . \nFor  both  algorithms  (1)  and  (2),  the  RBF  based  combiners  provide  higher speci(cid:173)\nficity  than the MSA.  The median combiner provides results similar to those of the \naverage combiner,  except for  algorithm  (2)  where it provides better specificity.  In \norder to obtain the final  discrimination between normal tissue and SILs,  constituent \nalgorithms (1)  and (2) are used sequentially, and the results are reported in Table 4. \n\nTable 2:  Accuracy of constituent algorithm  (1)  for  differentiating SILs  and normal \nsquamous tissues,  using MSA  and RBF ensembles. \n\nAlgorithm \n\nSpecificity  Sensitivity \n\nMSA \n\nRBF-ave \nRBF-med \n\n63% \n\n90% \n\n66%  \u00b11% \n66%  \u00b11% \n\n90%  \u00b1O% \n90%  \u00b11% \n\nTable 3:  Accuracy of constituent algorithm  (2)  for  differentiating SILs  and normal \ncolumnar tissues,  using  MSA  and  RBF ensembles. \n\nAlgorithm  Specificity  Sensitivity \n\nMSA \n\nRBF-ave \nRBF-med \n\n36% \n\n97% \n\n37%  \u00b15% \n44%  \u00b17% \n\n97%  \u00b1O% \n97%  \u00b1O% \n\nOne-step  algorithm:  The  results  presented  above  are  based on  the  multi-step \nalgorithm  specifically  developed  for  the  MSA,  which  could  not  consolidate  algo(cid:173)\nrithms (1)  and (2)  into one step.  Since the ultimate goal of these two algorithms is \nto separate SILs  from  normal  tissue samples,  a  given  pattern has  to be  processed \nthrough both algorithms.  In order to simplify  this decision process, we  designed a \none step  RBF  network to perform  this separation.  Because the pre-processing for \nalgorithms (1)  and (2) is different8 ,  the input space is now  26-dimensional.  We  ini(cid:173)\ntialized  10 kernels using a k-means algorithm on a trimmed9  version of the training \nset.  The kernel  locations  and spreads were not adjusted during training.  The cost \nof a  misclassified  SIL  was set at 2.5  times  the cost of a  misclassified  normal tissue \n\nSIn  this case,  the cost of misclassifying a SIL was three times the cost of misclassifying \n\na  normal  tissue sample. \n\n7In  this case,  there was  no  need to increase  the cost of a  misclassified  SIL,  because of \n\nthe high  prominence of SILs in  the training set. \n\n8Normalization  vs.  normalization  followed  by mean scaling. \n9The  trimmed set  has  the same number of patterns from  each  class.  Thus,  it  forces \neach  class  to  have  a  similar  number  of kernels.  This  set is  used  only for  initializing  the \nkernels. \n\n\f986 \n\nK.  Tumer;  N.  Ranulnujam. R.  Richards-Kortum and 1.  Ghosh \n\nsample,  in  order  to provide  the  best sensitivity/specificity pair.  The average and \nmedian combiner results  are obtained by pooling 20  RBF networks10 . \n\nTable 4:  One step RBF algorithm compared to multi-step MSA and clinical methods \nfor  differentiating SILs and normal tissue samples. \n\nAlgorithm \n2-step MSA \n\n2-step RB F -ave \n2-step RBF-med \n\nRBF-ave \nRBF-med \n\nPap smear (human expert) \nColposcopy  (human expert) \n\nSpecificity \n\nSensitivity \n\n63% \n\n65%  \u00b12% \n67%  \u00b12% \n67%  \u00b1.75% \n65.5%  \u00b1.5% \n68%  \u00b121% \n48%\u00b123 % \n\n83% \n\n87%  \u00b11% \n87%  \u00b11% \n91%  \u00b11.5% \n91%  \u00b11% \n62%  \u00b123% \n94%  \u00b16% \n\nThe results of both the two-step  and one-step  RBF  algorithms and  the results  of \nthe  two-step  MSA  are compared to the  accuracy of Pap smear  screening and  col(cid:173)\nposcopy  in  expert  hands  in  Table  4.  A  comparison of one-step  RBF  algorithms \nto  the two-step  RBF  algorithms indicates  that the one-step algorithms have  simi(cid:173)\nlar specificities, but a  moderate improvement in sensitivity relative to the two-step \nalgOrithms.  Compared  to the  MSA,  the one-step  RBF  algorithms  have  a  slightly \ndecreased  specificity,  but  a  substantially improved  sensitivity.  In  addition  to the \nimproved sensitivity, the one step RBF algorithms simplify the decision making pro(cid:173)\ncess.  A comparison between the one step RBF algorithms and Pap smear screening \nindicates  that  the  RBF  algorithms  have  a  nearly 30%  improvement  in  sensitivity \nwith no compromise in  specificity;  when compared to colposcopy in expert  hands, \nthe RBF ensemble algorithms maintain the sensitivity of expert colposcopists, while \nimproving the specificity by almost 20%.  Figure 2 shows the trade-off between speci(cid:173)\nficity  and  sensitivity for  clinical  methods,  MSA  and  RBF  ensembles,  obtained  by \nchanging  the misclassification cost.  The RBF ensembles provide better sensitivity \nand higher reliability than  any other method for  a  given specificity value. \n\n1.0 \n\n.z. \n\u00b7S  0.8 \n~ \n'(;; \nc \nQ) \n(fJ  0.6 \n\nA \n\n+ \n\n+ \n\n+ \n\n+ \n\n+ \n\nA \n\nA  A \n\n+ \n\nA \n\n~RBF-ave \nG- ~ RBF-mad \n~MSA \n\n+ Pap smear \nA  Colposcopy \n\n0.4  ~~ __ ~ ____ ~ ____ ~ ____ ~ ______ ~ ____ ~ ____ ~ ____ - J  \n0.8 \n\n0.2 \n\n0.0 \n\n0.4 \n\n0.6 \n\n1 - Specificity \n\nFigure 2:  Trade-off between sensitivity and specifity for  MSA  and RBF ensembles. \nFor reference, Pap smear and colposcopy results from the literature are included (Fa(cid:173)\nhey et al.,  1995). \n\nlOThis  procedure is repeated  10  times, in  order to determine the standard deviation. \n\n\fSpectroscopic Detection of Cervical Pre-cancer through RBF Networks \n\n987 \n\n5  Discussion \n\nThe  classification  results  of  both  the  multivariate  statistical  algorithms  and  the \nradial  basis function  network ensembles demonstrate that significant improvement \nin classification accuracy can be achieved over current clinical  detection modalities \nusing cervical tissue spectral data obtained from  in  vivo fluorescence  spectroscopy. \nThe one-step RBF algorithm has  the potential  to significantly reduce  the number \nof pre-cancerous cases  missed  by  Pap  smear  screening  and  the number  of normal \ntissues misdiagnosed by expert colposcopists. \n\nThe qualitative nature of current clinical detection modalities leads to a significant \nvariability in  classification accuracy.  For example, estimates of the sensitivity and \nspecificity of Pap smear screening have  been shown to range from  11-99%  and  14-\n97%,  respectively  (Fahey  et  al.,  1995).  This  limitation  can  be  addressed  by  the \nRBF  network  ensembles  which  demonstrate  a  significantly  smaller  variability  in \nclassification  accuracy  therefore  enabling  more  reliable  classification.  In  addition \nto  demonstrating  a  superior  sensitivity,  the  RBF  ensembles  simplify  the  decision \nmaking  process  of the  two-step  algorithms  based  on  MSA  into  a  single  step  that \ndiscriminates  between  SILs  and  normal  tissues.  We  note  that for  the given  data \nset,  both MSA  and MLP were unable to provide satisfactory solutions in one step. \n\nThe one-step  algorithm  development  process can  be readily  implemented  in  soft(cid:173)\nware,  enabling  automated  detection  of cervical  pre-cancer.  It  provides  near  real \ntime  implementation of pre-cancer  detection in  the  hands of non-experts,  and can \nlead  to  wide-scale  implementation  of screening  and  diagnosis  and  more  effective \npatient management in the prevention of cervical cancer.  The success of this appli(cid:173)\ncation will  represent an important step forward in both medical laser spectroscopy \nand gynecologic oncology. \nAcknowledgements:  This research  was  supported in  part by  NSF grant ECS  9307632, \nAFOSR contract  F49620-93-1-0307,  and Lifespex,  Inc. \n\nReferences \n\nFahey,  M.  T.,  Irwig,  L.,  and  Macaskill,  P.  (1995).  Meta-analysis  of pap  test  accuracy. \n\nAmerican Journal  of Epidemiology,  141(7):680-689. \n\nHansen,  L.  K. and Salamon, P.  (1990).  Neural network ensembles.  IEEE  1Tansactions  on \n\nPattern  Analysis and Machine  Intelligence,  12(10):993-1000. \n\nKurman, R.  J ., Henson,  D.  E., Herbst,  A.  L.,  Noller,  K.  L.,  and Schiffman,  M.  H.  (1994). \nInterim guidelines of management of abnormal cervical cytology.  Journal of American \nMedical  Association,  271:1866-1869. \n\nRamanujam,  N.,  Mitchell,  M.  F.,  Mahadevan,  A.,  Thomsen,  S.,  Malpica,  A.,  Wright, \nT.,  Atkinson,  N.,  and Richards-Kortum,  R.  R.  (1996).  Cervical  pre-cancer detecion \nusing  a  multivariate statistical  algorithm  based  on  fluorescence  spectra at  multiple \nexcitation wavelengths.  Photochemistry  and Photobiology,  64(4):720-735. \n\nTurner,  K.  and  Ghosh,  .J.  (1995).  Order  statistics  combiners  for  neural  classifiers. \n\nIn \nProceedings  of the  World  Congress  on  Neural  Networks,  pages 1:31-34,  Washington \nD.C.  INNS  Press. \n\nTurner,  K.  and Ghosh,  J.  (1996).  Error correlation  and error reduction  in  ensemble clas(cid:173)\n\nsifiers.  Connection  Science.  (to appear). \n\nWolpert, D.  H.  (1992).  Stacked generalization.  Neural  Networks,  5:241-259. \n\n\f", "award": [], "sourceid": 1285, "authors": [{"given_name": "Kagan", "family_name": "Tumer", "institution": null}, {"given_name": "Nirmala", "family_name": "Ramanujam", "institution": null}, {"given_name": "Rebecca", "family_name": "Richards-Kortum", "institution": null}, {"given_name": "Joydeep", "family_name": "Ghosh", "institution": null}]}