@techreport{TR-IC-PFG-16-02,
   number = {IC-PFG-16-02},
   author = {William Tustumi and Helio Pedrini},
   title = {{Skin Lesion Classification in Dermoscopy Images}},
   month = {December},
   year = {2016},
   institution = {Institute of Computing, University of Campinas},
   note = {In Eglish, 10 pages.
    \par\selectlanguage{brazil}\textbf{Resumo}
       Melanoma  \'e um dos tipos mais agressivos de cancer de pele. O
       sucesso   do   tratamento   \'e   dependente  do  diagn\'ostico 
       prematuro.  Esse  projeto  procura  estudar  e  desenvolver uma
       ferramenta   autom\'atica   para  assitir  o  diagn\'ostico  de 
       les\~oes  de  pele.  Inicialmentem,  as  imagens  de  fer\'idas 
       cut\^anias   s\~ao   segmentadas   atrav\'es   de  algor\'itmos 
       utilizados  em  an\'alise  de  imagens,  como Otsu's threshold,
       Chan-Vese    and    Statistical    Region    Merging.   Depois,  
       caracte\'irsticas s\~ao extra\'idas atrav\'es das regras ABCD e
       rela\c{c}\~oes  de  segunda  ordem  da  imagem. A partir dessas
       caracter\'isticas,   uma   decis\~ao  \'e  feita  utilizando  a 
       combina\c{c}\~ao,   atrav\'es   do   m\'etodo   de   voto,  dos 
       classificadores  Extra  Tree,  Decision  Tree, AdaBoost, Linear
       Discrimination    e    Random   Forest.   Os   resultados   das  
       classifica\c{c}\~oes  atingiram  uma  taxa  de  acerto  de 77\%
       usando a base de dados ISIC, que \'e compar\'avel ao sucesso de
       dermatologistas   especialistas  que  apenas  fizeram  o  exame 
       dermatol\'ogico e inspecionaram o hist\'orico do paciente.
    \par\selectlanguage{english}\textbf{Abstract}
       Melanoma  is  one  of the most aggressive types of skin-cancer.
       The   success  of  the  treatment  is  very  reliant  on  early 
       diagnosis.  This project aims to study and develop an automatic
       tool  to help physicians diagnose skin lesions. Initially, skin
       lesion  images  are segmented through well known image analysis
       algorithms,  such  as  Otsu,  Chan-Vese  and Statistical Region
       Merging.  Then,  features  are  extracted by following the ABCD
       rule   and  second  order  image  characteristics.  From  these 
       features,  a  decision  is  made  by means of a voting strategy
       combining   Extra   Tree,   Decision   Tree,  AdaBoost,  Linear 
       Discrimination  and  Random  Forest classifiers. Classification
       results  achieved  a  success rate of 77\% on the ISIC dataset,
       which  is  comparable  to  expert  dermatologists that only had
       access to dermatology exams and patient history.
  }
}