Pechko a, Sijche; Venge , Lyudmyla; Andono ski, D agana; Andono ic, Be i
A icle
B eas Cance De ec ion om The mal Images using
Machine Lea ning
ENTRENOVA - ENTe p ise REsea ch InNOVA ion
P o ided in Coope a ion wi h:
IRENET - Socie y o Ad ancing Inno a ion and Resea ch in Economy, Zag eb
Sugges ed Ci a ion: Pechko a, Sijche; Venge , Lyudmyla; Andono ski, D agana; Andono ic, Be i
(2025) : B eas Cance De ec ion om The mal Images using Machine Lea ning, ENTRENOVA -
ENTe p ise REsea ch InNOVA ion, ISSN 2706-4735, IRENET - Socie y o Ad ancing Inno a ion and
Resea ch in Economy, Zag eb, Vol. 10, Iss. 1, pp. 567-577,
h ps://doi.o g/10.54820/en eno a-2024-0042
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/317985
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h ps://c ea i ecommons.o g/licenses/by-nc/4.0/
567
ENTRENOVA - ENTe p ise REsea ch InNOVA ion
Vol. 10 No. 1
B eas Cance De ec ion om The mal Images using
Machine Lea ning
Sijche Pechko a
Facul y o Technology and Me allu gy, Skopje, No h Macedonia
Lyudmyla Venge
IPec us P ojec , Be lin, Ge many
D agana Andono ski
No h Kansas Ci y Hospi al, Missou i, Uni ed S a es
Be i Andono ic
Facul y o Technology and Me allu gy, Skopje, No h Macedonia
Abs ac
In his s udy, he au ho s p opose an ad anced s a egy o analyze he mal images
o b eas cance de ec ion employing machine lea ning echniques. By ocusing on
c i ical ea u es ha cap u e geome ic and s uc u al in o ma ion in he mal images,
he aim is o ele a e he p ecision and uni o mi y o b eas cance diagnos ics. The
da ase comp ises he mal images om pa ien s wi h b eas cance ; hese i al
ea u es a e ex ac ed and in eg a ed in o p oposed decision ee model, esul ing in
a classi ica ion accu acy o 92%. This highligh s he u ili y o combining specialized
ea u es wi h machine lea ning algo i hms in medical image analysis. Consequen ly,
he indings sugges ha his app oach can subs an ially enhance adi ional imaging
me hods, es ablishing a obus basis o ea ly and accu a e b eas cance de ec ion.
Keywo ds: b eas cance , he mal images, machine lea ning
JEL classi ica ion: Y80
Pape ype: Resea ch a icle
Recei ed: 12 Feb ua y 2024
Accep ed: 29 Jun 2024
DOI: 10.54820/en eno a-2024-0042
568
ENTRENOVA - ENTe p ise REsea ch InNOVA ion
Vol. 10 No. 1
In oduc ion
The e a e many s udies ha suppo he mog aphy as a p omising supplemen a y ool
a he han a s andalone solu ion o b eas cance sc eening. They sugges i may
bene i pa ien s a high isk o wi h speci ic b eas cha ac e is ics, bu ha
s anda diza ion and u he esea ch a e needed o op imize i s ole. The consensus is
ha he mog aphy’s non-in asi e na u e and po en ial o de ec hea i egula i ies
make i a aluable addi ion o mammog aphy, especially wi h ad ances in AI and
machine lea ning imp o ing i s diagnos ic capaci y.
S udies ha e alua e he e ec i eness o he mog aphy as a diagnos ic ool o
b eas cance , ocusing on i s po en ial o complemen mammog aphy. Kim e al.
(2015) and Mok e al. (2013) conduc ed me a-analyses and ound ha he mog aphy
can de ec abno mal hea pa e ns linked o cance ous issue, bu i s sensi i i y and
speci ici y a e inconsis en . Thei esul s indica e ha he mog aphy alone canno
eliably eplace mammog aphy bu may ha e alue when used as an adjunc ool.
Ahmad e al. (2014) u he explo ed he mog aphy’s abili y o assis mammog aphy,
pa icula ly o pa ien s wi h dense b eas issue, whe e adi ional sc eening may miss
ea ly signs o cance .
Zhang e al. (2017) e iewed in a ed he mog aphy’s ole in sc eening and ound
ha while i can de ec cance - ela ed he mal anomalies, i s diagnos ic p ecision alls
sho o mammog aphy. Cze winski e al. (2015) echoed hese indings, poin ing ou
ha while he mog aphy is non-in asi e and adia ion- ee, i s accu acy a ies, likely
due o di e ences in echnique and echnological limi a ions.
Ras ghalam & Pou ghassem (2013) and Selama e al. (2018) ocus on e ining he
p ocess o ea u e ex ac ion in he mal imaging. Ras ghalam & Pou ghassem
de eloped a spec al ea u e ex ac ion app oach ha success ully iden i ies
asymme ies in he mog ams, which o en indica e abno mali ies. Selama e al.
p o ide a comp ehensi e e iew o in a ed imaging echniques, discussing how
inc eased he mal ac i i y in cance ous issue can be e ec i ely dis inguished om
heal hy issue h ough digi al analysis, hus enhancing he diagnos ic p ecision o
he mog aphic imaging.
P amanik e al. (2019) conduc ed de ailed s udies on segmen a ion me hods o
he mog aphic b eas images. One s udy by P amanik e al. uses le el-se
segmen a ion o isola e suspicious egions by applying s a is ical and ex u e-based
analysis, while ano he examines ad anced s a is ical segmen a ion echniques o
di e en ia e malignan a eas, o e ing g ea e p ecision in pinpoin ing abno mal
egions.
Mohamed e al. (2022) and Goncal es e al. (2022) explo e he applica ion o
con olu ional neu al ne wo ks (CNNs) o classi y and p edic b eas cance in
he mog aphic da a. Mohamed e al. achie ed high diagnos ic accu acy by using a
CNN-based classi ica ion sys em, highligh ing i s po en ial o au oma ed de ec ion in
he mog aphy. Goncal es e al. u he e ined CNN pe o mance by applying bio-
inspi ed algo i hms o op imize he a chi ec u e, which signi ican ly imp o ed
diagnos ic esul s.
Cha e jee e al. (2022) de eloped deep ea u e selec ion me hods o ex u e
analysis, which allow o he enhancemen o classi ica ion ou comes in he mal
imaging. In one o hei s udies, hey use G unwald-Le niko -based ea u e selec ion
wi h a deep lea ning model, demons a ing imp o ed classi ica ion accu acy o
de ec ing ea ly cance indica o s.
S udies on b eas cance de ec ion using he mal imaging inc easingly in eg a e
machine lea ning (ML) and a i icial in elligence (AI) o imp o e diagnos ic accu acy.
The mal imaging cap u es su ace empe a u e dis ibu ions, highligh ing hea
569
ENTRENOVA - ENTe p ise REsea ch InNOVA ion
Vol. 10 No. 1
anomalies linked o cance ous issue. Models like con olu ional neu al ne wo ks
(CNNs) ha e achie ed high accu acy in de ec ing b eas cance on he mog ams,
especially wi h image p ep ocessing echniques like segmen a ion. App oaches in ol-
ing ex u e analysis, ea u e ex ac ion and colo mo phology u he e ine de ec ion.
In his s udy, we p esen a me hod o u ilizing ad anced machine lea ning
echniques o analyze he mal images o b eas cance de ec ion. Ou ocus lies in
iden i ying c i ical ea u es ha cap u e c ucial geome ic and s uc u al in o ma ion
om hese images, hus imp o ing he accu acy and consis ency o b eas cance
diagnosis. Employing a da ase comp ised o he mal images om pa ien s wi h
con i med b eas cance , we ex ac and inco po a e hese i al ea u es in o ou
decision ee model, achie ing a classi ica ion accu acy o 92%. This demons a es
he e ec i eness o combining specialized ea u es wi h machine lea ning algo i hms
in medical image analysis. Ou indings sugges ha his app oach can signi ican ly
enhance adi ional imaging me hods, laying a s ong ounda ion o ea ly and
p ecise b eas cance de ec ion.
The ask in ol es p ocessing and analysing he mal images o app oxima ely 400
pa ien s, each consis ing o a se ies o he mal images. The pa ien coho includes
indi iduals bo h wi h and wi hou con i med cance diagnoses. The goal is o
de e mine i he e a e disce nible pa e ns o ea u es in he he mal images ha can
help iden i y po en ial cance cases among he gi en da ase .
The In ui ion behind ou App oach
Asymme y is a common cha ac e is ic o cance g ow h as i sp eads une enly and
i egula ly in he body, making i di icul o iden i y consis en pa e ns o symme ies
ha can be eliably used o analysis.
To add ess his challenge, a ious echniques a e explo ed in o de o ex ac
ea u es om medical images in a symme ical way. Symme ical ea u e ex ac ion
e e s o he p ocess o iden i ying and analyzing pa e ns o ea u es ha ha e
e lec i e o mi o -like p ope ies ac oss he le and igh sides o an image. This
app oach is based on he assump ion ha i cance exhibi s asymme y, hen i s
e lec ion o mi o image on he o he side o he body may p o ide aluable
in o ma ion o diagnosis and p ognosis.
Howe e , ex ac ing symme ical ea u es om medical images can be a complex
ask due o se e al challenges. Fi s , as men ioned in he ex , i is o en unclea which
ea u es a e ele an o analysis since asymme y may no be easily desc ibed o
quan i ied. Second, medical images can con ain signi ican a iabili y and noise,
making i challenging o eliably iden i y symme ical pa e ns. Thi d, he p esence o
a i ac s o o he s uc u al di e ences be ween he le and igh sides o an image
can u he complica e he p ocess o iden i ying symme ical ea u es.
Despi e hese challenges, esea che s a e explo ing a ious app oaches o
symme ic ea u e ex ac ion in medical imaging, including he use o deep lea ning
models, egis a ion echniques, and s a is ical analysis me hods. These app oaches
aim o au oma ically iden i y and ex ac symme ical ea u es om medical images
while accoun ing o a iabili y and noise, as well as he p esence o s uc u al
di e ences o a i ac s. Ul ima ely, he goal is o de elop obus and accu a e
me hods o analyzing asymme ical medical da a ha can help imp o e diagnosis,
ea men planning, and pa ien ou comes.
Me hodology
The wo k in ol es collec ing he mal images om abou 400 pa ien s, labeled by hei
heal h s a us, o c ea e a obus da ase o aining a machine lea ning model. A e
570
ENTRENOVA - ENTe p ise REsea ch InNOVA ion
Vol. 10 No. 1
p ep ocessing he images— esizing, no malizing, and augmen ing hem o ensu e
consis ency and imp o e obus ness—meaning ul ea u es will be ex ac ed using
di e en me hods. Then selec ed machine lea ning algo i hms, such as Random
o es s o Decision T ees will be ained on he da ase , and hei pe o mance
e alua ed h ough me ics o p edic he likelihood o cance in new pa ien s based
on hei he mal images. This a e he main poin s:
1. Da a Collec ion: Ga he he mal images om app oxima ely 400 pa ien s, ensu ing
o documen hei heal h s a us (ei he heal hy o diagnosed wi h cance ) o
c ea e a comp ehensi e labeled da ase . This da ase will se e as he ounda ion
o aining and es ing ou machine lea ning model.
2. Fea u e Ex ac ion: To de i e signi ican ea u es om he he mal images, we will
employ echniques like His og am o O ien ed G adien s (HOG), Local Bina y
Pa e ns (LBP), o Con olu ional Neu al Ne wo ks (CNN). These ex ac ed ea u es
will be u ilized as inpu s o ou machine lea ning model.
3. Model T aining: Selec an app op ia e machine lea ning algo i hm, such as Suppo
Vec o Machines (SVM), Random Fo es Classi ie , o deep lea ning a chi ec u es.
The da ase will be di ided in o aining and es ing subse s, employing me hods
like k- old c oss- alida ion o assess model pe o mance.
4. Model E alua ion: E alua e he pe o mance o he ained machine lea ning
model by analyzing a ious me ics. The model will hen be used o p edic whe he
a new pa ien ep esen ed by an unlabeled he mal image is likely o ha e cance .
I will p oduce a p obabili y sco e indica ing he likelihood o cance p esence.
Da a Collec ion
We collec ed he mal images om app oxima ely 400 pa ien s, ca e ully eco ding
each indi idual’s heal h s a us (whe he heal hy o diagnosed wi h cance ) o
cons uc a well-labeled and de ailed da ase . This da ase is c ucial o aining and
es ing ou machine lea ning model, o e ing a wide a ay o he mal pa e ns
associa ed wi h bo h cance ous and non-cance ous condi ions.
Figu e 1
Example o pa ien da a.
Sou ce: Au ho s’ wo k
571
ENTRENOVA - ENTe p ise REsea ch InNOVA ion
Vol. 10 No. 1
Each pa ien ’s da a includes se e al on al images, along wi h side- iew images
cap u ed a 90 and 45-deg ee angles, hus p o iding a ied pe spec i es o enhance
model accu acy and obus ness in de ec ing he mal asymme ies ha migh signi y
unde lying abno mali ies.
Wi h o e 6,000 images in o al, his da ase cap u es an ex ensi e ange o body
sposi ions and he mal a ia ions. Such di e si y is in ended o imp o e he model’s
gene alizabili y by allowing i o lea n om di e en ana omical iews and he mal
signa u es unique o each pa ien . The wide ange o samples also con ibu es o
iden i ying sub le pa e ns and empe a u e disc epancies ha migh o he wise go
unno iced. By inco po a ing hese di e en angles and mul iple pa ien pe spec i es,
he da ase es ablishes a solid ounda ion o a machine lea ning model aimed a
achie ing high diagnos ic accu acy and eliabili y in b eas cance de ec ion h ough
he mal imaging.
Fea u e Ex ac ion
In o de o iden i y and ex ac a eas o in e es om an image, we i s need o ain
a sui able model. This could be achie ed using a ious machine lea ning o compu e
ision echniques. Once he a ea o in e es is iden i ied using he ained model, we
will p oceed o ex ac ea u es om i o u he analysis. The applica ion ha we
made has he abili y o ma k he a ea o in e es using ci cles. A e we se he a ea o
in e es o some images, we apply machine lea ning echniques o lea n he
coo dina es o he ci cles. Fo his we a e using il e ed da a as desc ibed below.
Figu e 2
Example o pa ien da a wi h a ea o in e es . The ci cles on he image ep esen he
a ea o in e es .
Sou ce: Au ho s’ wo k
S anda d s a is ical ea u es a e nume ical alues calcula ed om he image da a,
such as mean, median, a iance, s anda d de ia ion, his og am, and momen
in a ian ea u es. These ea u es help in desc ibing he in ensi y dis ibu ion and colo
in o ma ion o he a ea o in e es .
572
ENTRENOVA - ENTe p ise REsea ch InNOVA ion
Vol. 10 No. 1
To p epa e he image o ea u e ex ac ion, we apply se e al image p ep ocessing
il e s o enhance edges, supp ess noise, and educe unwan ed da a. Some o hese
il e s include:
● The Laplace il e is a second-o de de i a i e il e used o edge de ec ion in
images. I highligh s egions o apid in ensi y change and is commonly used o
ind edges and ex u es. The il e calcula es he Laplacian o he image,
emphasizing a eas whe e he in ensi y g adien changes signi ican ly. This il e
is pa icula ly e ec i e o de ec ing edges and is o en applied as a
p ep ocessing s ep o u he image analysis asks. The Gaussian g adien il e
combines Gaussian smoo hing wi h g adien compu a ion, p o iding a way o
de ec edges while educing noise.
● The Gaussian G adien il e is a powe ul image p ocessing echnique ha
in eg a es Gaussian smoo hing wi h g adien compu a ion o de ec edges
while minimizing noise. Ini ially, i applies a Gaussian blu o he image,
e ec i ely educing high- equency noise, which enhances he eliabili y o he
subsequen g adien calcula ion. By measu ing changes in pixel in ensi y,
ypically using ope a o s like Sobel, he il e highligh s edges and ansi ions in
he image. This dual- unc ionali y makes he Gaussian G adien il e
pa icula ly e ec i e o asks such as ea u e ex ac ion, objec de ec ion, and
segmen a ion, especially in applica ions whe e edge in eg i y and noise
educ ion a e c i ical.
● The Fou ie Gaussian il e ope a es in he equency domain, u ilizing he Fou ie
ans o m o apply a Gaussian il e o an image. This il e e ec i ely educes
high- equency noise while p ese ing low- equency componen s, making i
sui able o asks such as image smoo hing and noise educ ion. The Gaussian
shape in he equency domain co esponds o a smoo h blu ing e ec in he
spa ial domain, which can be bene icial in applica ions equi ing enhanced
image quali y.
● The Minimum il e eplaces each pixel's alue wi h he minimum alue wi hin a
speci ied neighbo hood a ound ha pixel. This il e is pa icula ly use ul o
emo ing small-scale noise o b igh spo s om images while p ese ing he
o e all shape and ea u es o la ge objec s. I is o en used in conjunc ion wi h
o he il e s o enhance he quali y o he image by supp essing isola ed high-
in ensi y pixels.
These il e s a e ke nel-based ans o ma ions ha change he in ensi y dis ibu ion
o an image based on speci ic ma hema ical ope a ions, he eby b inging ou he
desi ed ea u es.
Addi ionally, we manually label a subse o images (app oxima ely 6000) o c ea e
a aining da ase o ou machine lea ning model. This labeled da a will be used o
each he model how o iden i y and classi y di e en a eas o in e es . Once he
model is ained, i can be applied o a la ge da ase o unlabeled images o
au oma ically ex ac and label hei espec i e a eas o in e es based on he lea ned
ea u es.
In ou image p ocessing sys em, we ca e ully ex ac dis inc ea u es om he
speci ied egions o an inpu image o comp ehensi e analysis. The p ocess is ca ied
ou sepa a ely o bo h he le and igh sides o he image. To cla i y, he ollowing
a e he speci ic ea u es we ex ac :
1. Fos (Fi s O de S a is ics): This ea u e ep esen s he i s o de s a is ical
measu es, including mean, s anda d de ia ion, a iance, skewness, ku osis,
and ene gy,
2. ex ac ed om he in ensi y alues o he pixels in he egion.
573
ENTRENOVA - ENTe p ise REsea ch InNOVA ion
Vol. 10 No. 1
3. Ng dm (Neighbo hood G ay Tone Di e ence Ma ix): The Neighbo hood G ay
Tone Di e ence Ma ix is a ex u e ea u e ha desc ibes he spa ial dis ibu ion
o g ayscalein ensi y di e ences wi hin a de ined neighbo hood. This ea u e
helps us unde s and he coa seness and con as o he ex u e in he egion.
4. S m (S a is ical Fea u e Ma ix): S a is ical Fea u e Ma ix ep esen s highe o de
s a is ical ea u es, such as au oco ela ion, en opy, co ela ion ma ix, and
ene gy spec al densi y, ha p o ide mo e de ailed in o ma ion abou he
dis ibu ion o pixel in ensi ies wi hin a egion.
5. Fd a (F ac al Dimension Tex u e Analysis): F ac al Dimension Tex u e Analysis is
a ea u e ha quan i ies he i egula i y o complexi y o he ex u e by
de e mining i s ac al dimension. This ea u e helps in iden i ying ea u es such
as c acks, po es, and ough su aces.
6. Fps (Fou ie Powe Spec um): The Fou ie Powe Spec um analyzes he image
in e ms o i s equency componen s by calcula ing he powe spec al densi y
o each pixel in he egion. This ea u e is pa icula ly use ul o de ec ing
pe iodic pa e ns o shapes.
7. Glszm (G ay Le el Size Zone Ma ix): G ay Le el Size Zone Ma ix is a ex u e
ea u e ha cha ac e izes spa ial ela ionships be ween pixels wi h simila
g ayscale in ensi y alues wi hin a ious-sized neighbo hoods. I desc ibes he
dis ibu ion o such zones in he image, p o iding insigh s in o he uni o mi y and
coa seness o he ex u e.
8. Lbp (Local Bina y Pa e n): Local Bina y Pa e n is a ea u e ha ep esen s he
spa ial a angemen o pixel in ensi y di e ences wi hin a de ined window o
neighbo hood. This ea u e helps in iden i ying local s uc u es and pa e ns in
he image.
9. Dw (Disc e e Wa ele T ans o m): Disc e e Wa ele T ans o m is a mul i-
esolu ion ans o m used o analyze images a di e en scales, p ese ing bo h
spa ial and equency in o ma ion. I can e ec i ely cap u e ea u es such as
edges and co ne s.
Fo each speci ied egion in he image, we ex ac all hese ea u es. In o al, his
esul s in a ea u e ec o wi h 732 elemen s.
Resul s
Model T aining
Fo ou p edic i e model we a e using Random Fo es o Classi ica ion T ees. I is an
ensemble lea ning me hod ha combines mul iple decision ees o c ea e a powe ul
and obus classi ie . The algo i hm educes o e i ing by in oducing andomness in
bo h he selec ion o aining da a and he ea u es used o spli ing, ul ima ely
imp o ing he accu acy and gene aliza ion capabili ies o he esul ing model.
The algo i hm ope a es by gene a ing nume ous decision ees, each ained on a
andom subse o he da a wi h a unique selec ion o ea u es a each spli . This
p ocess, known as boo s apping and ea u e andomiza ion, educes o e i ing by
dec easing co ela ions among ees. Each ee is ained independen ly o make
accu a e p edic ions on i s subse . When classi ying a new da a poin , all ees in he
o es make independen p edic ions. Fo classi ica ion asks, he inal classi ica ion is
de e mined by majo i y o ing ac oss he ees, whe e he mos equen class is
selec ed. Al e na i ely, a e aging p obabili ies ac oss ees can yield a e ined
p obabili y es ima e, use ul o p o iding nuanced esul s in bo h classi ica ion and
eg ession asks.
They o e se e al ad an ages: hey a e highly adap able and can model nonlinea
ela ionships be ween ea u es and a ge classes, making hem applicable o a wide
574
ENTRENOVA - ENTe p ise REsea ch InNOVA ion
Vol. 10 No. 1
ange o p oblems. By u ilizing mul iple decision ees, hey educe o e i ing isk
h ough an a e aging e ec , which mi iga es noise om indi idual weak lea ne s.
Sui able o bo h classi ica ion and eg ession asks, Random Fo es s pe o m
e ec i ely e en in high-dimensional spaces wi h la ge da ase s. They also p o ide
insigh s in o ea u e impo ance, helping use s iden i y he mos in luen ial a iables in
he p edic ions, and hey a e ela i ely easy o ain.
Model E alua ion
We a e using 10 old c oss- alida ion o e alua ing he p edic i e pe o mance o
he model. C oss- alida ion is a s a is ical me hod used o e alua ing models and
assessing how well he model gene alizes o unseen da a. The main idea behind c oss-
alida ion is o pa i ion ou da a in o di e en subse s: one o aining he model ( he
" aining se ") and ano he o es ing i ( he " es se "). Fo 10-Fold C oss-Valida ion he
da ase is andomly di ided in o 10 subse s o olds. In each i e a ion, 9 olds a e used
o aining, while he emaining old is used as a es ing se o e alua e he model's
pe o mance. The p ocess is epea ed 10 imes. The p ima y ad an age o c oss-
alida ion is i s abili y o p o ide a mo e obus es ima e o ou model's gene aliza ion
e o compa ed o using a single aining/ es spli . This is because i helps o a e age
o e he a iabili y in oduced by di e en andom ain/ es spli s, leading o mo e
eliable es ima es o pe o mance.
To ensu e obus alida ion and p e en o e i ing, we ca e ully assign di e en
pa ien s o sepa a e olds when pa i ioning he da a. This app oach mi iga es he isk
o " inge p in ing," a phenomenon whe e a model lea ns o ecognize speci ic
indi iduals' unique cha ac e is ics a he han gene alizable pa e ns indica i e o
b oade condi ions, like cance . By keeping each pa ien 's da a exclusi e o a single
old, he model is es ed on en i ely new indi iduals in each alida ion cycle, simula ing
eal-wo ld condi ions whe e he model mus gene alize o new pa ien s i has ne e
encoun e ed be o e. This echnique a oids da a leakage, whe e he model migh
o he wise exploi simila i ies wi hin he same pa ien 's da a ac oss mul iple olds,
leading o a i icially in la ed accu acy sco es du ing c oss- alida ion.
This pa ien -speci ic pa i ioning s a egy is pa icula ly impo an in medical
imaging, as e en sligh pe sonal he mal a ia ions be ween pa ien s can bias he
model i no ca e ully managed. By ensu ing ha no pa ien appea s in mo e han one
old, we educe he chance ha he model will ely on unique indi idual ai s, such as
ana omy o na u al he mal pa e ns, a he han lea ning meaning ul indica o s o
disease. This p ocess enhances he eliabili y o pe o mance me ics and helps c ea e
a model ha is bo h accu a e and gene alizable o new, unseen pa ien s in p ac ical
diagnos ic se ings.
In he inal model p edic ions, we use a simple o ing echnique o agg ega e he
esul s om mul iple images o he same pa ien , imp o ing diagnos ic consis ency
and accu acy. This means ha , ins ead o elying on a single he mal image o each
diagnosis, we combine p edic ions om se e al images. By agg ega ing hese
p edic ions, he model educes he isk o e o s due o image-speci ic a ia ions, such
as sligh empe a u e inconsis encies o mino posi ioning di e ences, which migh
o he wise mislead a single-image p edic ion.
This o ing app oach le e ages he di e si y o images o s abilize and enhance he
model's o e all diagnos ic eliabili y. I ensu es ha he inal classi ica ion is based on
a mo e comp ehensi e iew o he pa ien 's he mal p o ile, e ec i ely minimizing he
impac o any anomalies o noise in indi idual images. This p ocess helps educe alse
posi i es and alse nega i es, as he collec i e o e emphasizes pa e ns consis en
ac oss mul iple images a he han ou lie s. As a esul , agg ega ing p edic ions om
