Int ern at i onal  Journ al of  El e ctrical  an d  Co mput er  En gin eeri ng   (IJ E C E)   Vo l.   10 ,  No.   5 ,  Octo be r   2020 ,  pp.  4738 ~ 4744   IS S N:  20 88 - 8708 ,  DOI: 10 .11 591/ ijece . v 10 i 5 . pp 4738 - 47 44           4738       Journ al h om e page :  http: // ij ece.i aesc or e.c om/i nd ex .ph p/IJ ECE   Gliobla s tomas   b rain  t umour   s eg m entation  b ase d  on   c on vo lut ional  n eu ra l  n etworks       Moh ’ d   R asoul   Al - Hadi di 1 ,   B ayan   AlS aaid ah 2 ,   M ohamm ed   Y.   Al - G aw ag z eh 3   1,3 Depa rtment   of   Elec tr ical   Pow e r   Engi n ee ring ,   Depa rtment   of   Co m pute r   Engi n ee r ing,   Fa cul t y   of   E ngine er ing ,     Al - Bal qa   Applied   Univer sit y ,   Jor dan   2 Depa rtment   of   Com pute r   Scie n ce ,   Princ e   Abdul la h   b in   Gha zi   Fa cul t y   of   Inform a ti on   Te chno log y   and   Com m unicati ons ,   Al - Bal qa   Applied   Univer sit y ,   Jor dan       Art ic le   In f o     ABSTR A CT     Art ic le   hist or y:   Re cei ved   Oct   15 ,   2019   Re vised   Ma r   14 ,   2020   Accepte d   Ma r   2 5 ,   2020       Brai n   tumour   segm ent at ion   c an   improve   di agnosti cs   eff ic i ency ,   rise     the   pre d ic t ion   rat e   and   tr ea tm ent   pl anni ng.   T his   will   he lp   t he   doct ors     and   expe r ts   in   the ir   work.   W h ere   m an y   t y pes   of   bra in   tumour   may   be   cl assifi ed   ea sil y ,   the   gli om as   tumour   is   cha ll eng ing   to   be   segm ent ed   because   of   the   diffusion   bet wee n   th e   tu m our   and   the   surrounding   edem a.   Another   important   ch al l e nge   with   thi s   t ype   of   bra in   tum our   is   tha t   the   t um our   m ay   grow   an y wher e   in   the   bra in   w it h   diffe r ent   shape   and   siz e.   B rai n   ca n ce r   pre sents   one   of   the   m ost   famous   disea ses   over   t he   world,   which   enc our age   the   rese ar che rs   to   find   a   high - throughput   s y stem   for   tumour   det ec t ion   and   cl assifi ca t ion.   S eve ra l   appr o ac h es   have   b ee n   p r oposed   to   desig n   aut om at i c   det e ct ion   and   c la ss ifi c at ion   s y s te m s.   Thi s   pap er   pre sents   an   int egr at e d   fra m ework   to   se gm ent   the   gl io m as   bra in   tumour   aut om at i call y   using   pixe l   cl uster ing   for   th e   MRI   images   f ore ground   and   bac kground   and   cl assif y   its   t y p e   base d   on   d ee p   l ea rn ing   m ec hani sm ,   which   is   the   convo lut i onal   neur a l   net work .   In   thi s   work,   a   novel   segm ent at ion   a nd   cl assificat ion   sy st em   is   proposed   to   detec t   th e   tumour   ce l ls   and   cl assif y   the   br ai n   ima ge   if   it   is   hea l th y   or   not .   After   col l ec t in g   dat a   for   heal th y   and   non - he al th y   bra in   images,   sat isfact or y   r esult s   are   f ound   and   r egi st ere d   using   computer   v ision   appr oac h es.   Thi s   appr oa ch   can   be   used   as   a   p art   of   a   b igge r   dia gn osis   s y stem   for   bre ast   tumou r   detec t ion   and   m ani pula ti o n.   Ke yw or d s :   Brai n   tum our     Conv olu ti onal   neural   net works   su pe r pix el   Im age   segm entat ion     Pixel   cl us te rin g   Copyright   ©   202 0   Instit ut e   of   Ad vanc ed   Engi n ee r ing   and   S cienc e .     All   rights   reserv ed .   Corres pond in g   Aut h or :   Moh ’ d   Ra soul   Al - H adi di ,   Dep a rtm ent   of   Ele ct rical   Pow er   E ng i neer i ng,   De pa rtm ent   of   Com pu te r   En gin ee rin g,   Faculty   of   E ngineerin g,   Al - B al qa   A ppli ed   U niv e rsity ,   King   Tal al   St r eet ,   Salt   19 117,   Jor dan .   Em a il :   m oh a m m ad_ ha did i @ bau.ed u.j o       1.   INTROD U CTION   Brai n   cance r   present   one   of   the   highest   dea th   causes   besi des   seve ral   cancer   ty pes   wit h   the   highest   death   c om par ed   with   t he   nu m ber   of   patie nt s   [1 ] .   T he   bra in   tum ou r   is   a   gro up   of   a bn or m al   cel ls   tha t   grow     in   the   brai n   [ 2].   Detect   t his   m ass   an d   ide ntif y   the   locat io n   of   it   hel ps   the   do ct or s   to   treat   the   patie nts;   in   m os t   cases,   they   nee d   to   rem ov e   t he   tum ou r   sur gical ly .   Wh ere   t he   brai n   t um ou r   has   m any   typ es,   gliom as   pr esent   the   m os t   diff ic ult   one   for   pr e dicti on .     In   the   gliom a s   ty pe,   the   tu m ou r   area   po or ly   co ntrasts   and   diff ic ult   to   segm ent   reg ar ding   its     diffusi ng.   Furtherm or e,   the   t um ou r   sprea d   in   m any   siz e   and   s ha pes   in   the   br ai n   [3 ] .   In   sp it e   of   the   la st   i m pr ovem ent   in   the   brai n   ca ncer   treat m ent   that   happen e d   recently ,   but   the   m or bid it y   sti ll   cor relat ed   with     the   po or   diag no sis .   Accord i ng   to   t he   Am erican   B rain   t um ou r   Associ at ion   sta te s,   t her e   are   120   ty pes   of     the   br ai n   tum ou r   a nd   it   becom es   the   m os t   death   ca us e   of   the   young   pe op le   w hose   ag e   under   40   ye ars   [ 4].   Desp it e   al l   the   i m pr ovem ents   in   the   brai n   ca ncer   treat m ent   but   the   survi va l   rate   sti ll   low,   wh ic h   as   re por te d   in   the   cu re  br ai n ca ncer f oundat ion   a nd   s how n   in   Fig ure   1   [ 5].   Evaluation Warning : The document was created with Spire.PDF for Python.
In t J  Elec  &  C om p  En g     IS S N:  20 88 - 8708       Gliob l as to m as   br ai n  t umo ur  s egm e nta ti on  base d on co nv olu ti onal... ( Mo h’d  Raso ul Al - H ad i di )   4739   Early   detect io n   of   br ai n   ca nc er   can   hel p   the   patie nt   to   be   su r vi ved   a nd   ov e rc om e   cancer   treat m ent   pro blem s.   The   low   sur viv al   pe rcen t,   the   high   cost   of   the   treatm ent,   the   sever it y   beh i nd   t he   surge ry   treatm ent,   and   a   la r ge   nu m ber   of   brai n   ty pes   prese nt   dem and   for   ea rly   detect ion   wit h   an   e ff ect ive   diag nosis.   T he   m os t   popula r   im aging   m et ho d   for   m edical   pu r po ses   is   the   m agnet ic   resonan ce  i m aging   ( MR I )   m et ho d   [6]   in   wh ic h   a   strong   m agn et ic   fiel d   is   use d   be sides   t he   rad i o   wa ves   a nd   t he   fiel d   gr adients.   De pendin g   on   t he   cl inica l   app li cat io n,   dif fer e nt   ty pes   of   co ntrast   t hat   a re   us ed   in   MR   i m aging   li ke   T1 – an d   T2 – w ei gh te d   im aging   [7 ] .   An   exam ple   of   the   MRI   im ag es   is   sho wn   in   Figure   2.             Figure   1.   S urvival   rate   for   t he   per i od   bet ween   1984 - 20 13   [ 5]       Figure   2.   Brai n   i m age   us i ng     MRI   te ch nique       Cl assify ing   the   br ai n   cel ls   if   it   is   healt hy   or   not,   t he   t um ou r   cel ls   s hould   be   seg m ented   fir st.     The   m os t   popula r   se gm entation   m et ho d   is   a   reg i on   gr ow i ng   m et ho d   whic h   de pe nds   on   a   see d   point   that   is   grow i ng   acc ordin g   to   the   E uclidia n   distan ce   bet ween   pi xels   [ 8].   H owever,   the   se gm entat ion   pro cess   is   consi der e d   as   a   chall eng e   for   researc her s   be cause   of   the   im age   un if or m it y   and   the   var i at ion   of   the   cel ls   size   and   s hap e   [ 9].     Super pix el   m eth od   is   a   si m ple   ty pe   of   cl us te r ing   that   is   us e d   for   im age   par ti ti on ing   proces s   [10]   and   base d   on   the   m os t   i m po rtan t   par t   of   any   im age   wh ic h   is   the   pix el   valu e   [11].   Using   s ever al   pa ram eter s   an d   dep e ndin g   on   the   distance   be tween   pix el s,   these   pa rtit ion s   are   segm ented   an d   la belle d   with   va riant   siz es .   These   sub - im a ges   a re   us e d   as   input   for   cl ass ific at ion   m od el s   f or   cl assifi cat ion   pur poses.     Conv olu ti onal   neural   netw ork   (CN N )   is   c on sidere d   as   a   robu st   cl assifi cat ion   m od el   t hat   is   trai ne d   and   le a rn e d   on   a   huge   num ber   of   data   set s   a nd   desi gn e d   usi ng   a   c om bin at ion   of   netw ork s   as   la ye rs.   Us ing     the   CNN   m eans   the   abili ty   to   extract   featu re s   from   the   raw   input   data   us i ng   its   com plica t ed   hier arc hy   w it ho ut   need   for   the   m anu al   featu r e   extracti on   [ 12 ] .   T his   stu dy   aim s   to   s egm ent   the   gliom as   br ai n   tum ou r   autom at ic ally   us in g   pix el   cl ust ering   f or   the   MRI   im ages   fo re gro und   a nd   bac kgr ound   a nd   us e   the   res ults   to   cl assify   the   cel l   sta tus   ba sed   on   deep   le ar ning   m echan ism   wh ic h   is   t he   C NN .       2.   R EL ATED   W ORK   Segm entat ion   the   brai n   tum our   proces s   is   sti l l   a   chall eng e   f or   the   resea rchers   an d   the   m os t   com m on   m et ho d   f or   bra in   tum ou r   se gm entat ion   is   t he   re gion   gro wing   m et ho d   [13 ] .   The   segm entat ion   process   us in g   reg i on   grow i ng   nee d   f or   a   m anu al   sel ect i on   for   a   seed   in   w hich   the   se le ct ed   po i nt   m ay   cause   an   in te ns it y   distance   e rror   in   the   hom og en ei ty   of   the   of   pi xels.   A nothe r   m et ho d   m ay   be   the   th reshold ing   [14]   de pe ndin g   on   t wo   grey   le vels   (0   an d   255)   this   m ay   cause   losi ng   s om e   of   the   ac tual   tum ou r   ce ll s.   Ba sed   on   i m age   processi ng   te c hn i qu e s   an d   usi ng   ANNs ,   the   cancer   cel ls   wer e   detect ed   and   cl assifi e d   [15].   This   w ork   is   insp ire d   to   m erg e   a   c om patip le   te ch niques   to   get   t he   m os t   us ef ul   i nfo rm ation   from   the   im ages   ba sed   on     the   RO I   us in g   i m age   proces sing   te ch niques.     Dep e ndin g   on   the   sy m m et ric al   po ints   of   t he   le ft   and   the   ri gh t   side s   of   the   br ai n,   s om e   m et ho ds   wer e   pro po se d.   E xtr act   the   feat ur e s   al ong   the   line   betwee n   t he   two   sides   w he re   low   sym m et ry   m ean s   there   is   diff e re nt   ti ssu e   w hich   m eans   tum ou r   e xisti ng   [16 ,  17] .   B ut   this   way   can not   be   ef fici ent   with   gliom as   t um ou r   ty pe   beca us e   t his   ty pe   a ppear s   in   s om e   cases   in   var i ou s   loca ti on s   with   dif fe ren t   s ha pe   a nd   siz e.     Using   t he   c onvo l ution al   net works   in   cl as sific at ion   a ble   to   e xtract   s ophisti cat ed   fea tures   w hich   m akes   them   w el l - m eaning .   T his   is   done   by   prov i ding   the   ou t pu t   feat ur e   m aps   of   a   C onvoluti onal   la ye r   as   input   cha nn el s   to   the   subse quent   C onvolut ion al   la ye r   [ 18] .   The   buil din g   blo c ks   in   CNN   al lo w   f orm ing   diff e re nt   ty pes   of   CN N s .   T his   ty pe   of   de ep   le ar ning   netw orks   is   ve ry   effe ct ive   for   high - perf orm ance   com pu te r   visio n   m od el ,   a nd   t hey   ef fici ently   le arn   a nd   e xtra ct   m any   visu al   featur e s   for   we ll   gen e rali zi ng   ta sk s   without   the   ne ed   for   hand - cra fted   featu re   e xtracti on   [ 19 ] .   Most   of   t he   e xisted   m et ho ds   are   base d   on   cl us t eri ng   al gorithm s,   m achine   le ar ning,   or   us in g   the   w hole   i m a ge   base d   on   deep   le ar ni ng   al go rithm s   [20 - 23 ] .     Evaluation Warning : The document was created with Spire.PDF for Python.
                          IS S N :   2088 - 8708   In t J  Elec  &  C om p  En g,   V ol.  10 , No .  5 ,  Oct ob e r   2020   :    47 38   -   47 44   4740   The   perform ance   of   these   m eth ods   de pends   on   t he   qu al it y   and   the   ty pe   of   the   ext racted   f eat ur es   w hich   can   be   var ie d   [15 ,  24] .   The   m ai n   aim   of   t his   pa per   is   to   de velo p   an   integrate d   cl ust ring   a nd   dee p   le ar ning   bas ed   ap proa c h     to   de te ct   an d   extract   the   br a in   tum ou r   a nd   cl assify   its   ty pe.   Ba sed   on   su pe r pix el   cl ust ring   al gorith m   for   tum ou r   se gm e ntati on   is   e xpe ct ed   to   wor k   pro per ly   withou t   need i ng   f or   t he   m anu al   det ect ion   of   t he   tum ou r   cel ls.   More ov e r,   us in g   the   de ep   le ar ning   f or   cl assi ficat ion   pu rposes   wi ll   be   ind e pe ndent   f ro m   the   f eat ure   extracti on   pro cess   w hich   is   tradit ion al ly   us e d   in   m achine   le ar ning.   F ur t her m or e,   the   propose d   a ppr oac h   sh owe d   prom i sing   res ults   w hich   pro ve   the   abili ty   of   the   dee p   le ar ning   al gorithm   to   pro du ce   a   r obus t   an d   accurate   detect ion   a nd   cl assifi cat ion   syst em   for   the   gliom as   br ai n   tum ou r .       3.   E X PERI MEN T   AND   RES U LT S     The   pro posed   stud y   ai m s   to   segm ent   the   brai n   t um ou r   usi ng   a   s up e rp i xe l   cl us te rin g   m et ho d   t he n   cl assify   the   la belle d   patches   us in g   C NN.   T his   w ork   was   carried   out   over   fi ve   m on ths   an d   will   be   i m pr ov e d   su bse que ntly   f or   bette r   res ults.       3.1.   Material   and   da t a   set   The   pro posed   al gorithm   was   carried   out   a nd   te ste d   us in g   a   data   set   from   t he  ca ncer  im aging  arc hiv e   (TCI A)   [25 ,  26] .   This   data   set   is   pu blicl y   avail able   an d   can   be   us e d   f or   researc h   a nd   academ ic   purposes.   The   ne uro rad i ologist s   in   Th om as   Jeff erson   U niv e rsity   (TJU )   H os pital   prov i de   the   im age   by   its   feature   char act e risat ion s.   T he   total   num ber   of   im ages   in   this   data   set   is   40 69;   the   healt hy   br ai n   is   pr esente d   by   988   i m ages   w her e   t he   non - healt hy   brai n   is   presen te d   by   3081   im ages.       3.2.   Experim en t   The   pro posed   s yst e m   con sist s   of   m ulti ple   stag es   as   sho wn   in   F i gure   3 .           Figure   3 .   Ge ne ral   m et ho do l ogy       3.3.   Pre - pr ocessin g   This   ste p   ai m s   to   pr e par e   the   im ages   an d   a dju st   thei r   co ntrast   us in g   filt erin g   a nd   norm al is e     the   i m ages   us i ng   sta ti sti cal   op erati ons   bb a s ed   on   the   fo ll ow i ng   e quat io n   [ 27 ] .   T his   st ep   was   a ppli ed   to   al l   i m ages   befo re   the   s up e rp i xel   segm entat ion   proces s.     =   − +   (1)     wh e re   C   is   the   con t rast,      an d      are   the   m axi m um   and   m ini m um   lu m inance   values .   Evaluation Warning : The document was created with Spire.PDF for Python.
In t J  Elec  &  C om p  En g     IS S N:  20 88 - 8708       Gliob l as to m as   br ai n  t umo ur  s egm e nta ti on  base d on co nv olu ti onal... ( Mo h’d  Raso ul Al - H ad i di )   4741   3.4.   Superpi xel   se gmen tatio n   Af te r   pr e par i ng   the   MRI   i m ages   an d   rem ov e   a ny   no ise   m ay   app ear   and   ca us e   seg m entat ion   or   cl assifi cat ion   e rror,   a   s uper pi xel   se gm entat i on   pro cess   wa s   ap plied   to   se gm ent   the   brai n   tum our   area .   The re   are   di ff e ren t   a lgorit hm s   can   be   use d   f or   s uper pix el   se gme ntati on   [28].   The   pro po se d   m et ho d   us e d   s i m ple  li near   it erati ve   cl us te rin g  ( S LIC)   al go rith m   [ 10 ],   wh ic h   adap ti ve ly   refi nes   the   c om pactness   pa ram et er   after   the   fir st   it erati on.   T he   first   st ep   of   this   al go rithm   is   init ia l i sing   center s   for   cl us te rs   on   a   gr i d   s paced   S   pix el .   Nex t,   t he   cl us t er   centers   a re   a lt ered   into   3   ×   3   nei ghbor hood   base d   on   the   lowest   gra dient   po sit io n.   Eac h   pix el   is   assigne d   to   the   nea rest   pix e l   based   on   the   m easur ed   distance   as   sho wn   in   ( 2 )   w hic h   is   m easur ed   us in g   ( 3 )   and   ( 4 )   w hich   f ind   t he   c olor   ne arn es s   a nd   the   sp at ia l   nea r ne ss   res pecti vely .       D = √ ( d c m ) 2 + ( d s S ) 2   (2)     d c = √ ∑ ( I ( x i , y i , s p ) − I ( x j , y j , s p ) ) 2 s p ∈ B   (3)     d s =   √ ( x j − x i ) 2 + ( y j − y i ) 2   (4)       is   the   sp ect ra l   band   that   ha s   the   pix el s   ( , , )   and   ( , , ) ,   m   par a m et er   is   us e d   to   c on tr ol     the   su pe r pix el s   com pactness,   B   pr esents   the   sp ect ral   band   set .   Finall y,   S   pr ese nts   the   sa m pling   interval   of   each   cl ust er   ce ntr oid   [ 29 ].   Sp li t   the   im age   into   la bels   a fter   se ve ral   at tem pts   to   fin d   t he   m os t   su it ab le   value   of   t he   num ber   of   su pe r pix el s   we   wan t   to   create ,   wh ic h   is   15   a reas.   A fter   c om pu ti ng ,   the   num ber   of   s up e rp i xels,   w hich   is   16,   the   col our   of   e ach   pix el   was   s et   us in g   t he   m e an   value   of   the   super pix el   re gi on .   T his   gr ouping   process   is   done   dep e ndin g   on   the   s patia l   distance   a nd   al s o   the   inte ns it y   di sta nce   bet wee n   t he   pix el s.   F igure   4   s hows   thes e   su pe r pix el s   w he re   F i gure   5   s hows   the   la belle d   reg i on s   after   set ti ng   the   pi xe l   values .   Applyi ng   thes e   ste ps   a nd   bin a rize   the   r esulta nt   im age,   the   require d   se gm ented   i m age   for     the   non - he al th y   cel ls   is   pr oduced .  It  is   sho wn   in   F ig ur e   6.   The   se gm ented   im ages   will   be   us e d   in   C NN   for   trai ning   purpos es   to   pr e dict   the   sta tus   of   the   br ai n   cel ls.             Figure  4. Pixel  v al ue s  set ti ng       Figure  5. Im age  su pe r pix el s       Figure  6. Se gme nted   im age       3.5.   Convoluti onal   neur al ne two rks   In   this   sta ge,   the   resu lt ed   patc hes   or   the   s ub - areas   f ro m   the   su pe r pix el   seg m entat ion   ste p   are   la belle d   then   trai ned   usi ng   the   CN N   to   cl assify   the   brai n   cel ls   norm al it y.   The   tradi ti on al   way   for   cl assifi cat ion   a lway s   carried   out   by   extracti ng   the   f eat ur es   m anu al ly   then   us e   one   of   the   m achin e   le arn in g   cl as sifie rs   su c h   as   neural   netw orks   a nd   SV M.   By   us in g   the   dee p   le ar ning   net work,   wh ic h   is   CNN ,   sign ific a nt   fea tures   will   be   extracte d   us in g   the   ra w   i m ages   wh ic h   are   he re   the   resu lt e d   patc hes   f r om   the   su pe r pix el   ste p.   T he   C NN   s tructu re   com pr ise s   of   m any   la ye rs:   the   in put   la ye r,   the   co nvol ution al   la ye rs ,   poolin g   la ye rs ,   dro pout   la ye rs ,   f ully   connecte d   la ye rs,   a nd   finall y   the   ou t pu t   la ye r .   T hese   la ye rs   are   e xp la ine d   be low   a s s how n i n  Fig ure  7 .   a.   Conv olu ti onal   la ye r :  This  is  the  fi rst  la ye r  that  deals  with  the  ra w  i m age.  This  l ay er  co ns ist s  of    m any  filt ers  th at   are  c onvolv ed  to   ha ve   wei gh ts   f or  each   r egio n  of  the  i m age  that  is  presented   as  a   fe at ure   m ap  [ 30 ] .    b.   Pooli ng   la ye r:   Af te r   ha ving   a  huge  num ber   of  feat ur es ,   these  feat ur es   are  re duced  us in g  the  pool in g    la ye r  that wil l r edu ce  the  com pu ta ti onal  c omplexit y o f  the  net work [ 32 ] .     Evaluation Warning : The document was created with Spire.PDF for Python.
                          IS S N :   2088 - 8708   In t J  Elec  &  C om p  En g,   V ol.  10 , No .  5 ,  Oct ob e r   2020   :    47 38   -   47 44   4742   c.   Fu ll y  connecte d  la ye r :  This  t he  la st  la ye r  wh ere  eac h  neur on   in  this  la ye r  is  connecte d  with  al l  neu r ons     in the p rev i ou s  lay er.           Figure   7 .   Co nvolu ti onal   ne tw ork   st ru ct ur e       The   arc hitec tu re   of   the   pro pose d   CN N   is   sh ow n   in   T a bl e   1.   In   e ve ry   sing le   la ye r   of   the   CN N   pro du ces   a   res pons e   for   the   input   i m age.   In   the   CN N,   t her e   a re   a   fe w   su it able   la ye rs   f or   im age   featur e   extracti on   pro cess.   The   first   la ye rs   of   its   structu re   ca pture   only   the   gl ob al   f eat ur e s   of   the   im age,   su c h   as     the   ed ges   an d   t he   blobs,   see   F igure   8,   w hich   sh ows   a   set   of   weig hts   f r om   the   first   la ye r.   In   e ver y   sin gle   la ye r   of   the   CNN   pr oduces   a   resp onse   for   the   input   i m age.   In   the   CNN ,   there   a r e     a   few   s uitable   la ye rs   for   im age   featu re   ex tract ion   pr oces s.   The   fi rst   la ye rs   of   its   str uctu re   captu re   on ly     the   global   featur es   of   the   im a ge,   su c h   as   the   edg es   an d   the   blobs,   see   F ig ure   8,   w hich   s hows   a   set   of   w ei gh ts   from   the   first   la ye r.   The se  fe at ur es  will   be  processe d  us in g  dee pe r  net w orks  f or  m or e  detai le d  featu r es.  A fter   hav i ng  a  trai ne d  m od el ,  the  e valuati on  proc ess  is  do ne  usi ng  the  te st  la be ls  with  the   pr edict ed  la bels   to  f i nd   the  cl assifi er  perform ance  and   acc ur acy .  Af te r  trai n  the   segm ented  pa tc hes  from   the  su pe rp i xel  process,     the  eval uation  process  s houl d  be  a pp li ed   by   rep eat in g  t he  sam e  ste ps   on   unknow n  im age  to  cl assify   it   and  fin d  the  accu ra cy  o f  the  pro posed  syst em .       Table  1.  C onvoluti onal   n et w ork  p aram et ers   Lay e r   Para m eter   I m ag e  Inp u t  ‘ d ata ’   2 5 6 x 2 5 6  ( n o r m aliz ed )   Co n v o l u tio n   ‘ co n v 1 ’   9 6  11 x 1 1 x 3  con v o lu tio n s   Max Po o lin g   ‘ p o o l 1 ’   3 x 3   m ax  po o lin g   Co n v o l u tio n   ‘ co n v 2 ’   2 5 6  5x5x 4 8  con v o lu tio n s   Max Po o lin g   ‘ p o o l 2 ’   2 x 2   m ax  po o lin g   Fu lly  Co n n ected  ‘ f c6 ’   4 0 9 6   f u lly  con n ected   Drop o u t  ‘ d rop 7 ’   50%   Clas sif icatio n   ‘ o u t p u t ’   2           Figure  8. First  conv olu ti onal  l ay er w ei gh t   Evaluation Warning : The document was created with Spire.PDF for Python.
In t J  Elec  &  C om p  En g     IS S N:  20 88 - 8708       Gliob l as to m as   br ai n  t umo ur  s egm e nta ti on  base d on co nv olu ti onal... ( Mo h’d  Raso ul Al - H ad i di )   4743   4.   RESU LT S   A ND   DI SCUS S ION   The   pro po se d   m od el   has   di fferent   trai ning   accuracy   us in g   a   dif fer e nt   num ber   of   e poch s   as   s ho w n     in   F ig ure   9 .   By   us in g  CN N,   t he  need  f or   a  la r ge  num ber   of   e poc hs   is  re du ce d  w her e   the  tr ai ni ng     accuracy  becom es  sta ble,  sta rting   from   12 5  epo c h.   T he  sy stem   per f or m a nce  was  e val ua te d,   an d  the  r esulte d  accuracy  was  repor te d  for  f urt her   e nh a nce m ent  in   the  fu ture.   T he  over al l  te st ing   accuracy  was  75 %,  an d    the accu racy  for  eac h  cl ass is  sh ow n  in  the   F igure  10 .   The  pr opos e d  m et ho d  trie d  t o  m erg e  deep  le arn in g  with   the  cl us te rin g  fo r  rob us tnes s  purpose s .   These  res ults  cou l d  be  im pr ov e d  by  par a m et er  tun ing,  op ti m isa t ion ,  a nd   a pp ly   on  a no t her   ty pe  of   m od el s  su c h  as  t he  de ci sion   tr ee  cl assifi er  by  cl a ssifyi ng  each   pa tc h  i ndivid ua ll y  then  ta ke   the  m os t  re dunda nt   cat egory  by  voti ng   f r om   al l  t he  im age  par ti ti on s.  This  w ork  c ould  be  e xt end e d  f or  m ulti cl ass  cl assifi cat ion   us in g  S VM  cl assifi er.  T he  c la ssific at ion   pr ocess  co ul d  co ver   m or e  br ai n  tum ou r  ty pe s  by  extracti ng  m or e  featur e s  based  on m achine learn i ng.           Figure   9 .   Trai ni ng   a cc uracy           Figure   10 .   C onfu si on   m at rix       5.   CONCL US I O N     The   brai n   can cer   rate   rises   r ecentl y,   wh ic h   le ad   the   research   to   fi nd   a   hig h - th rou ghput   detect io n   syst e m .   In   this   stud y,   an   a utom at ic   segm ent at ion ,   detect io n,   an d   cl assifi c at ion   syst e m   wer e   pro po s ed   to   detect   the   ab norm al   c el ls   and   ide ntif y   its   t ype.   The   propose d   ap pr oach   ai m s   to   find   a   rob us t   segm entat ion   process   besides   us i ng   t he   dee p   le ar ni ng   al gorithm ,   wh ic h   is   the   C NN .   T he   segm entat ion   us in g   su pe r pix el   s hows   an   eff ect ive   way   to   segm ent   the   br ai n   tum ou r   c el ls   and   by   us i ng   t he   patche s   wh ic h   sp eci fy   the   i m age   featur es .   Using   the   CN N   after   the   seg m entat ion   ste p   abr id ges   the   f eat ur e   extracti on   ste p,   w hich   is   a   big   chall eng e   for   t he   researc hers   in   m achine   lear ni ng   al gorith m s.   This   syst e m   can   be   extend e d   to   co ver   oth e r   ty pes   of   br ai n   cancer .   T his   sy stem   can   be   ap plied   us in g   a   di ff e ren t   num ber   of   the   supe rp i xel   patc hes.       ACKN OWLE DGE MENTS     This   researc h   has   bee n   car ri ed   out   duri ng   sab batic al   le ave   gr a nted   to   the   aut hor   M oh ’ d   Ra s oul     Al - H adi di   f rom   Al - Ba lqa   A pp li ed   U niv e rs it y   (BAU),   Salt ,   Jor dan   du rin g   the   aca dem ic   ye ar   20 17 / 2018.   Evaluation Warning : The document was created with Spire.PDF for Python.
                          IS S N :   2088 - 8708   In t J  Elec  &  C om p  En g,   V ol.  10 , No .  5 ,  Oct ob e r   2020   :    47 38   -   47 44   4744   REFERE NCE S     [1]   R.   L .   Si ege l ,   et a l.   “ Can ce r   statist ic s,   2016 , ”   CA:   a   cancer   journal   for   c li ni ci ans ,   v ol.   66 ,   no .   1 ,   pp .   7 - 30 ,   2016 .   [2]   [Online ] ,   Avai lable:  htt p :/ /www . khcc . jo/ sec ti on/b rai n - tumors - 0,   [ Ret ri eve d  10 - D e ce m ber - 2017 ] .     [3]   M.   Hava e i ,   et  a l. ,   “ Brai n   tumour   segm ent ation   with   Dee p   Neur al   Ne tworks ,”   Me dic a l   Image   Anal ysis ,   v ol.   3 5.     pp.   18 - 31 ,   2017 .   [4]   “ Brai n   Tumor   Sy m p toms ,   Treat m ent ,   Support,   Resea rch ,”   Ame rican   Brain   Tum o r   Associat ion ,   [ Online ] .   Availab le :   ww w.a bta . org .   [ Ret ri eve d   15 - D e ce m ber - 2017 ] .   [5]   “ Fact s   and   Stat s ,”   Cure  Brain   Cance r   Foundat io n ,   [Onlin e] .   Avai la bl e :   htt ps:// ww w.c ur ebr ai n ca n ce r. o rg . au/ p age /8 /facts - stat s.   [ R et ri eve d   20 - Novem ber - 2017 ] .   [6]   M.  R.   Al - Hadid i ,   A.  Ala rab e yy a t ,   and   M.  Alhanahnah,   “ Brea st   cance r   d et e ct ion   u sing   k - nea r est   n ei ghbor   m ac h ine   le arn ing   al gori t hm , ”  in  2016   9th   Inte rnational   Confe renc e   on   Dev el opments   in   eSy stems   En gine ering   ( DeSE ) ,    pp.   35 - 39 ,   2016 .   [7]   P.  Hagm ann ,   et   al. ,   “ Understa nd ing   dif fusion   MR   imaging   techn ique s:   from   sca l ar   diffusion - wei ghte d   imaging   to   diffusion   t ensor   imaging   and   be yond ,”   Radi ograp hic s ,   vo l.  26,   sup.  l1 ,   pp .   S205   −   S223,  2006.   [8]   H.  Hooda,   O.  P.  Verm a,   and   T.  Singhal ,   “ Brai n   tumour   segm en ta ti on :   A   per for m anc e   anal y s is   using   K - Mea ns,   Fuzz y   C - Me an s   and   Regi on   growing   al gorit hm , ”  in  Ad v ance d   Comm unic ati on   Con trol   and   Computing   Technol ogi es   ( ICACCCT) ,   2014   Inte rnat ional   C onfe renc e ,  pp.   1 621 - 1626 ,   2014 .     [9]   N.  Sauwen ,   et   al . ,  “ Hier a rch i ca l   nonnega t ive   m atrix   factori za t ion   to   cha r acte ri ze   b rai n   tumour   he terogene ity   using   m ult ipa ramet r ic   MRI,”   NMR   in   Bi omedi ci ne ,   vo l .   28 ,   no .   12 ,   pp .   1599 - 1624 ,  201 5 .     [10]   R.   Acha nta,   et   al .   “ SLIC   super pixe ls   compare d   to   stat e - of - the - art   superpixel   m et hods,”   IEE E   transacti ons   on   patt ern   ana ly sis   and   machine   in t el li g ence ,   vol .   3 4,   no .   11 ,   pp .   22 74 - 2282,   2012 .   [11]   B.   AlSaai d ah ,   et  al . ,   “ Z ebr af ish   La rva e   Cla ss ific at ion   b ase d   on   Dec ision   Tr ee   Model:   A   Com par ative   Anal y sis , ”   Adv anc es   in   Scie nce ,   Techno logy   and   Engi n ee ring   Syste ms   Journal ,   v ol .   3,   n o.   4,   pp .   347 - 353 ,   2018 .   [12]   O.  Ronnebe rge r ,   P.  Fis che r,   and   T.  Brox,   “ U - net :   Convolut ion al   net works   for   biomedi cal   imag e   segm ent at ion , ”    in  Int ernati onal   Confe renc e   on   Me dic a l   Image   Computing   and   Computer - Assisted   Int erv en ti on ,  pp.   234 - 241 ,   20 15 .   [13]   T.   Kal ai se lvi   an d   P.  Naga ra ja,   “ A   rap id   au tomat ic   bra in   tumour   det e ct ion   m e tho d   for   MRI   images   using   m odifi e d   m ini m um   err or   t hre sholding   te ch nique , ”   In t.   J.   Im aging   Syst ems   and   Technol og y ,   vol.   25 ,   no .   1 ,   pp .   77 - 85 ,   2015 .   [14]   D.  Cobza s ,   et   a l . ,   “ 3D   v ari a ti ona l   bra in   tumour   s egmenta t ion   usi ng   a   h igh   d imensional   f ea tur e   se t , ”   in   2007  IE E E   11th  Int ernati on al  Conf ere nce o n  Computer  V isi on ,  pp .   1 - 8 ,   200 7 .     [15]   M.  R.   Al - Hadidi ,   M.  Y.  Al - Gawagz eh,   and   B.   A.  Alsaa ida h ,   “ Solving   m am m o gra ph y   probl ems   of   bre ast   ca nc er   det e ct ion   using   art i ficial   neur a l   net works   and   image   pro ce ss i ng   technique s, ”   Indian   journal   of   sci ence   an d   te chno logy ,   vol .   5 ,   no .   4 ,   pp .   252 0 - 2528,   2012 .   [16]   K.  Popuri ,   et   al. ,   “ 3D   var ia ti on al   bra in   tumour   segm ent at ion   using   Diric hle t   p riors   on   a   cl ustere d   feature   set , ”   Inte rnational   jou rnal   of   compute r   ass iste d   radiolo gy   and   sur gery ,   vol.   7 ,   no .   4 ,   pp .   493 - 506,   2012 .   [17]   A.  Krizh evsk y ,   I .   Suts keve r ,   and   G.  E.  Hinton,   “ Im age net   cl assifi ca t ion   with   d ee p   convol ut iona l   n eur al   net works , ”  Adv anc es   in   neu ral   inf orm ati on   proce ss ing   syste ms ,   pp.   1097 - 11 05,   2012 .   [18]   J.  Donahue ,   et  a l .  “ Dec a f:   A   d eep   convol ut iona l   ac t iva t ion   fe at ur e   for   g ene ri c   vis ual   r ec ogni ti on , ”   in  Int ernati onal   conf ere n ce   on   m achi ne   le arning ,  pp.   647 - 655 ,   20 14 .     [19]   M.  Jafa ri   and   S.  Kasae i ,   “ Autom at ic   bra in   tis sue   det ec t ion   in   MRI   images   using   see ded   reg ion   growing   segm ent at ion   an d   neur al   ne twork   cl assificat ion , ”   Australi an   Journal   of   Basic   and   Appl ie d   Sc ie n c es ,   vol.   5,   no .   8 ,  pp.   1066 - 1079 ,   2011 .   [20]   E.  Tor ti ,   et   al . ,   “ The   HELI CoiD   proje ct:   par al l el   SVM   for   bra in   ca nc er   class ifi c a ti on , ”  in  Digit a l   Syste m   De sign   ( DS D ) ,   2017   Eu rom ic ro   Confe re nce ,  pp.   445 - 450 ,   2017 .     [21]   V.  Panca   and   Z .   Rustam,   “ Applicati on   of   m achine   learni ng   on   bra in   c anc e r   m ult ic la ss   class ifi cation , ”  in  A I P   Confe renc e   Pro c ee dings ,   AI P Pu bli shing ,   v ol.   18 62,   n o.   1 ,  p p .   03 0133 ,   2017 .     [22]   S.  Jain,   “ Brai n   ca nc er   class ifica t ion   using   GLCM   base d   fe at ure   ext ra ct ion   in   art if ic i al   n eur al   n et work, ”     Int   J   Comput   S ci   Eng   Te chnol ,   v ol.   4 ,   no .   7 ,   pp .   9 66 - 970,   2013 .   [23]   J.  J.  Corso ,   et   a l . ,   “ Eff ic i ent   m ul ti le v el   b rai n   tumour   segm ent atio n   with   in te gr at e d   Ba y esi an   m odel   cl assifi ca t ion,”   IEE E   transacti o ns   on   medical   imaging ,   vo l.  27 ,   n o.   5 ,   pp .   629 - 64 0,   2008 .   [24]   M.  R.   Al - Hadid i ,   D.  Al - Hadid i,   and   R.   S.  R az ou q,   “ Pneum onia   Ide nti fi cation   usi ng   Organi zi ng   Map   Algorit hm , ”   APRN   Journal   of   Eng ine ering   an d   Applied   S cienc es ,   vol .   11 ,   no .   5 ,   pp .   1819 - 6608 ,   2016 .   [25]   L.  Sc arp a ce ,   e t   al. ,   “ Data   Fro m   REMBRANDT   –   Th e   C an ce r   Im agi ng   Ar chi ve ,”   2015.   [ Online ] ,   Avai lable :   htt p://doi. o rg/10.7937/K9/T CIA. 2015. 588OZUZ B     [26]   K..   Cl ark ,   et  a l . ,   “ The   C anc e r   Im agi ng   Archi ve   (TCIA):   Ma int ai n ing   and   Opera ti ng   a   Pu bli c   In for m at io n   Reposit or y , ”   Jou rnal   of   Dig it al   I maging ,   v ol .   26 ,   no.   6,   pp   1045 - 1 057 ,  2013 .     [27]   E.  Pe li,   “ Contra s t   in   complex   ima ges ,”   JOSA   A ,   v ol.   7 ,   no .   10 ,   pp .   2032 - 2040,   199 0.   [28]   P.  Neube rt   and   P.  Protzel,   “ Superpixel   ben chma rk   and   compari s on , ”  i n   Proc.   F orum   Bi ldv erarbeit ung ,   v ol.   6 ,    pp.   1 - 12 ,   2012 .   [29]   C.   A.  Ortiz   Tor o ,   et   al. ,   “ Superpixe l - b ase d   roughne ss   m ea sure   for   m ult ispec tr al   satelli t e   imag e   segm ent at ion , ”   Re mote   sensing ,   vol.   7 ,   no .   11 ,   pp .   14620 - 14645 ,   2015.   [30]   W .   B.   Park ,   et  al. ,   “ Cla ss ifi c a ti on   of   cr y s ta l   struct ure   using   a   convol u ti ona l   neur al   net work , ”   IUCr J ,   vol.  4,     no.   4 ,   2017 .     [31]   N.  Ta jb akhsh   a nd   K.  Suzuki. ,   “ Com par ing   tw o   cl asses   of   en d - to - end   m ac hin e - learni ng   m ode ls   in   lung   nodul e   det e ct ion   and   cla ss ifi ca ti on :   MT AN Ns   vs   CNN ,”   Pattern   R ec ogn i ti on ,   vol.  63 ,   pp .   476 - 486 ,   2017.     Evaluation Warning : The document was created with Spire.PDF for Python.