Con en s lis s a ailable a ScienceDi ec
Compu a ional and S uc u al Bio echnology Jou nal
jou nal homepage: www.else ie .com/loca e/csbj
Resea ch A icle
Towa ds gene alizable Fede a ed Lea ning in medical imaging: A
eal-wo ld case s udy on mammog aphy da a
Ioannis N. Tzo zis a, ,∗,1, Albe o Gu ie ez-To e b, ,1, S a os Sykio is a,, Fe an Agulló b,,
Nikolaos Bakalosa, Anas asios Doulamisa,, Nikolaos Doulamis a, Josep Ll. Be al c,b,
aSchool o Ru al, Su eying and Geoin o ma ics Enginee ing, Na ional Technical Uni e si y o A hens, He oon Poly echneiou 9, A hens, 15773, A ica, G eece
bBa celona Supe compu ing Cen e , Plaça d’Eusebi Güell, 1-3, Les Co s, Ba celona, 08034, Ca alunya, Spain
cUni e si a Poli ècnica de Ca alunya, Ba celona Tech, Ca e Jo di Gi ona, 1-3, Les Co s, Ba celona, 08034, Ca alunya, Spain
A R T I C L E I N F O A B S T R A C T
Keywo ds:
Fede a ed Lea ning
Medical imaging
P i acy p ese a ion
Mammog aphy
Deep lea ning
BIRADS classifica ion
Fede a ed Lea ning has been apidly gaining in popula i y in medical applica ions, due o he inc eased p i acy
offe ed, since medical da a doesn’ need o lea e he hospi als’ p emises o AI model aining. Howe e , a di ec
ansla ion o a classic expe imen o a ede a ed one is no always s aigh o wa d. In his wo k, we del e in o he
in icacies o ede a ed lea ning o a b eas cance classifica ion ool. We compa e classic model aining wi h a
ede a ed a ian , and highligh he adap a ions ha need o be aken ca e o o ensu e he equi alence be ween
he wo. Specifically, we in oduce he B eas A ea De ec ion ool as an essen ial componen o he p e-p ocessing
pipeline o enhance he obus ness o Fede a ed Lea ning by offe ing da a ha moniza ion. On op o ha , we
p esen an end- o-end Fede a ed Lea ning amewo k ha is effec i e o eal-wo ld da a and scena ios. Among
he h ee eal-wo ld hospi als in ol ed in he expe imen al p ocedu e, he p oposed amewo k significan ly
imp o es pe o mance a he fi s hospi al, p o iding consis en esul s simila o hose achie ed in he classic
app oach. Expe imen al esul s demons a e ha he in e en ions in oduced imp o ed model pe o mance by
app oxima ely 35%, aligning ede a ed lea ning and cen alized model pe o mance.
1. In oduc ion
Heal hca e is one o he domains ha is hea ily domina ed by human
decision making. Al hough he c i ical na u e o heal hca e signifies he
need o human expe ise, he hea y eliance on human in e p e a ion
and assessmen c ea es bo lenecks in he clinical wo kflow, leading o
issues such as delays in he examina ion p ocess o significan wo kload
on medical p o essionals (MP) [1]. The e o e, he e is an inc easing need
o au oma ed ools o assis medical p o essionals in he clinical wo k-
flow. The ecen ise o A ificial In elligence (AI) c ea es unp eceden ed
oppo uni ies o he design and de elopmen o such ea u es [2–5]. I
can e en be said ha AI holds he key o unlocking he un apped po-
en ial o so wa e-based medical applica ions, which would au oma e
ce ain s eps in he clinical wo kflow o help clinicians by boos ing hei
p oduc i i y.
Howe e , he de elopmen o AI algo i hms equi es cen alized ac-
cess o la ge amoun s o da a o ensu e gene aliza ion, which, in he
case o medical applica ions, comes in conflic wi h p i acy conce ns,
*Co esponding au ho .
E-mail add ess: [email p o ec ed] (I.N. Tzo zis).
1Au ho s con ibu ed equally.
as well as egula ions (e.g. GDPR). To ha end, esea che s ha e been
in es iga ing echniques which can alle ia e hese p i acy issues. One
p omising echnique in he heal hca e domain is Fede a ed Lea ning,
which is a machine lea ning app oach whe e mul iple dis ibu ed de-
ices collabo a i ely ain a sha ed model [6,7]. In a Fede a ed Lea ning
F amewo k, he da a do no need o be ga he ed in a cen alized en i y.
Ins ead, in he con ex o medical applica ions, sensi i e pa ien da a can
s ay wi hin a hospi al’s p emises. A copy o he model will be ans e ed
o he hospi al, ained on he local da a, and e u ned o an ex e nal
se e , which will be esponsible o me ging he models om diffe en
hospi als in an op imal way.
E en hough Fede a ed Lea ning (FL) appea s p omising on pape ,
he adap a ion o AI wo kflows o a decen alized a chi ec u e is no
s aigh o wa d. Limi ed o no access o he aining da a equi es as-
sump ions abou hei in insic p ope ies, s a is ical dis ibu ions, e c.
In addi ion, he aining p ocess is ha d o debug i he ini ial esul s
a e no as desi ed. Da a- ela ed phenomena, such as da a he e ogenei y
o domain shi , can significan ly impac he model aining p ocedu e,
h ps://doi.o g/10.1016/j.csbj.2025.03.031
Recei ed 15 Janua y 2025; Recei ed in e ised o m 14 Ma ch 2025; Accep ed 16 Ma ch 2025
Compu a ional and S uc u al Bio echnology Jou nal 28 (2025) 106–117
A ailable online 20 Ma ch 2025
2001-0370/© 2025 The Au ho s. Published by Else ie B.V. on behal o Resea ch Ne wo k o Compu a ional and S uc u al Bio echnology. This is an open access
a icle unde he CC BY license (
h p://c ea i ecommons.o g/licenses/by/4.0/ ).
I.N. Tzo zis, A. Gu ie ez-To e, S. Sykio is e al.
while being ha d o iden i y. Finally, FL en ails inc eased implemen a-
ion complexi y, as well as model agg ega ion challenges [8–10].
This pape highligh s he challenges and adap a ions equi ed o
ansla e a cen alized medical AI model aining wo kflow o a ed-
e a ed one. Mo e specifically, we con end ha , by placing emphasis on
da a p epa a ion and eliable communica ion mechanisms, e en simple
agg ega ion me hods can achie e app op ia e pe o mance in a FL se -
ing. We u ilize he aining p ocedu e o a BIRADS classifica ion model
on mammog aphy da a as a case s udy o demons a e he s eps ha
need o be ollowed o ansi ion om classic aining, o a ede a ed
a ian wi h simila con e gence bounds. Ou app oach is implemen ed
using h ee hospi als as ede a ed nodes ( wo in G eece and one in Se -
bia) and highligh s he lessons lea ned, as well as sugges s mi iga ion
measu es o issues encoun e ed. The main ou comes o his wo k a e
summa ized as ollows:
•P e-p ocessing: he significan a iance in he pe o mance o he
ini ial ede a ed lea ning app oach among diffe en da a p o ide s,
compa ed o he classic app oach, led us o assume ha his disc ep-
ancy was due o da a he e ogenei y. To add ess his, we concei ed
he idea o building a pipeline specifically a ge ed a his issue.
This pipeline goes beyond adi ional me hods o medical image
p ocessing by in oducing he B eas A ea De ec ion ool, a supe -
ised lea ning-based so wa e designed o emo e anno a ion labels
and o he a i ac s om mammog ams.
•Adap ing he classic expe imen o a ede a ed scheme appea ed
o be a s aigh o wa d solu ion o c i ical issues in he medical
domain, such as p i acy p ese a ion. Howe e , in his wo k, we
demons a e ha he adap a ion is no as simple as i seems, pa ic-
ula ly when dealing wi h eal-wo ld da a and hospi als. Ul ima ely,
we p opose an end- o-end amewo k ha pe o ms effec i ely in
his demanding con ex .
•Lessons lea ned om he expe imen a ion: A majo con ibu ion
o his wo k is he aluable expe ience gained on his opic, wi h
conclusions ha can guide simila effo s. Fo each obse a ion/les-
son, we p opose specific mi iga ion ac ions and link hem o he
FUTURE-AI [11] ini ia i e.
1.1. Li e a u e e iew
A ificial In elligence (AI) has d i en significan ad ancemen s in
heal hca e, pa icula ly in medical imaging o cance diagnosis and
p ognosis. Among hese, deep lea ning models like Mi ai and AsymMi ai
ha e demons a ed excep ional capabili ies. Mi ai, o ins ance, p edic s
b eas cance isk up o fi e yea s in ad ance based on mammog a-
phy da a, while AsymMi ai emphasizes in e p e abili y by inco po a ing
bila e al asymme y ea u es, making i mo e anspa en and clinician-
iendly [12,13].
The e a e se e al publica ions ela ed o he classifica ion o BIRADS
sco es using mammog ams o ain deep lea ning models. In [14], he
au ho s in oduce he RN-BCNN model, a cus omized e sion o he ypi-
cal ResNe , o classi ying mammog ams acco ding o he BIRADS scale.
A mo e complex app oach is p esen ed in [15], whe e T ans e Lea n-
ing p inciples a e adop ed o enhance classifica ion esul s by inco po-
a ing s a e-o - he-a ne wo ks such as NASNe Mobile, VGG16, and
VGG19. Bo h s udies apply augmen a ion echniques o expand he lim-
i ed size o he INb eas da ase . Ano he no ewo hy s udy, p esen ed in
[16], demons a es he use o a DNN-based model o lesion localiza ion
while inco po a ing he co esponding BIRADS sco e. Addi ionally, a
wo-s age pipeline is in oduced in [17], whe e a YOLO 5-CBAM model
de ec s egions o in e es con aining lesions, which a e hen segmen ed
and classified acco ding o he BIRADS sco e in he second s age.
Howe e , classic cen alized AI app oaches in heal hca e ace nu-
me ous challenges. P i acy conce ns and s ingen egula ions like
HIPAA and GDPR es ic he agg ega ion o pa ien da a in cen alized
eposi o ies. Addi ionally, da a he e ogenei y—caused by a ia ions in
imaging p o ocols, equipmen , and pa ien demog aphics— u he com-
plica es he de elopmen o obus , gene alizable models [18,19]. The
inabili y o effec i ely ha monize da ase s ac oss ins i u ions o en leads
o biased models ha ail o pe o m well in di e se clinical se ings.
Fede a ed Lea ning (FL) has eme ged as a p omising solu ion o hese
challenges. FL enables collabo a i e aining o AI models while ensu -
ing ha sensi i e da a emains localized. This pa adigm has been suc-
cess ully applied in a ious medical imaging asks. Fo example, Bakas e
al. demons a ed he easibili y o FL in b ain umo segmen a ion, en-
abling mul i-ins i u ional collabo a ion wi hou comp omising pa ien
confiden iali y [20,21], while [22] pe o med an ex ensi e e alua ion in
clinical se ings, and highligh ed he need and ongoing conside a ions o
add ess secu i y and p i acy issues. Simila ly, FL has been employed in
his opa hology image analysis o he classifica ion o umo -infil a ing
lymphocy es, u he showcasing i s e sa ili y in add essing di e se
medical imaging p oblems [23].
Despi e i s p omise, FL p esen s i s own se o challenges. The non-IID
(non-Independen and Iden ically Dis ibu ed) na u e o medical imag-
ing da a, whe e da ase s om diffe en ins i u ions exhibi significan
a iabili y, poses a subs an ial hu dle o model con e gence and gen-
e aliza ion. Ad anced echniques like FedP ox and FedMA ha e been
p oposed o add ess hese issues by imp o ing model agg ega ion and
accommoda ing he e ogenei y in aining da a [24,25]. Fu he mo e,
FL sys ems ace challenges ela ed o communica ion o e head and sys-
em he e ogenei y, whe e a ia ions in compu a ional esou ces and
in as uc u e ac oss pa icipa ing ins i u ions complica e implemen a-
ion [19,26].
A c i ical a ea o ocus in FL is da a ha moniza ion, which seeks
o educe a iabili y ac oss da ase s and enhance model gene alizabil-
i y. S udies ha e highligh ed he impo ance o me ada a-d i en p e-
p ocessing and domain adap a ion echniques o align da a dis ibu ions
[18,27]. The applica ion o Kube ne es-based a chi ec u es in FL ame-
wo ks has also been no able, enabling scalabili y and adap abili y in
eal-wo ld implemen a ions [26,27].
Ou wo k ad ances he s a e o he a in Fede a ed Lea ning (FL)
o medical imaging by add essing c i ical challenges iden ified in p e-
ious s udies. Building on he impo ance o ha monizing he e ogeneous
da ase s as emphasized by Kilim e al. [18] and Zhou e al. [27], we in-
oduce a dynamic ha moniza ion pipeline ha adap s p e-p ocessing o
a ia ions in imaging p o ocols and pa ien demog aphics. While me h-
ods like FedP ox [24] and FedMA [25] aim o ackle non-Independen
and Iden ically Dis ibu ed (non-IID) da a, hey ha e se e al limi a-
ions. Fedp ox adds a p oximal e m o he local aining objec i e o
each node, which discou ages d as ic de ia ions om he global model.
E en hough his app oach helps in add essing da a he e ogenei y be-
ween nodes, he model may s uggle o gene alize in he case whe e no
many nodes a e pa icipa ing in model aining. This is usually he case
in c oss-silo scena ios, whe e he numbe o pa icipan s in FL is lim-
i ed. FedMA ollows a diffe en app oach, whe e laye -wise ma ching
is in oduced o align neu ons be ween nodes based on hei simila i y.
This leads o an adap i e model s uc u e ha conside s da a he e o-
genei y. Howe e , his app oach en ails significan addi ional compu a-
ional/communica ion complexi y in aining p ocedu e, as nodes need
o implemen addi ional in o ma ion exchange ounds wi h he agg e-
ga ion se e . In addi ion, FedMA only suppo s model a chi ec u es
comp ising o con olu ional and ully connec ed laye s, and laye -wise
ma ching canno be implemen ed in building blocks such as Long-Sho
Te m Memo y (LSTM) and T ans o me laye s. These ac o s es ic i s
p ac icali y in ede a ed medical se ings. In ou app oach, we a gue
ha mo e emphasis should be pu on da a p epa a ion and obus com-
munica ion mechanisms a he han sophis ica ed agg ega ion me hods,
as some o he cases can be sol ed by doing so. We use BIRADS classi-
fica ion on mammog ams as a case s udy, and implemen a Fede a ed
Lea ning p ocedu e ac oss h ee hospi al ins i u ions om he INCISIVE
p ojec (A hens, No i Sad, Thessaloniki). Ou wo k highligh s he s eps
ha need o be ollowed o he success ul implemen a ion o an FL
Compu a ional and S uc u al Bio echnology Jou nal 28 (2025) 106–117
107
I.N. Tzo zis, A. Gu ie ez-To e, S. Sykio is e al.
Fig. 1. The INCISIVE P ojec Cloud and Fede a ion in as uc u e.
in as uc u e, as well as highligh s impo an lessons lea ned and p o-
poses mi iga ion measu es.
2. Con ex and modeling me hodology o e iew
2.1. Da ase and pla o m de ails
The wo k desc ibed in his pape has been conduc ed unde he con-
ex o he INCISIVE Eu opean P ojec .2This has allowed us o access
mul iple da ase s om diffe en medical ins i u ions, and in pa icula
o his wo k, mammog ams. Da a collec ion was join ly o ganized wi h
he medical ins i u ions associa ed o he INCISIVE p ojec . In o de o
acili a e he da a collec ion and o la e use he da a, a Join Con olle
Da a Sha ing ag eemen was pu in place. All da a we e ga he ed in a
common cloud and also s o ed a he hospi al le el. An ex ensi e de-
sc ip ion abou he da a collec ion p ocess can be ound in he wo k by
Lazic e al. [28], co e ing bo h echnical and legal aspec s o he p ocess.
While he FL in as uc u e was no eady, he eams de eloping A i-
ficial In elligence (AI) models such as he one men ioned in his wo k,
used he Cloud o ain hei models in he classical se up.
On he o he hand, mos o he medical ins i u ions had hei own
on-p emise se e s whe e he da a was also s o ed in o de o p o ide
an in as uc u e o he FL amewo k. When he medical ins i u ion
could no se up hei own se e due o he lack o IT pe sonnel o
o he issues, a i ual machine was se up in he Cloud o hem. In he
case o ou wo k, all he ins i u ions had hei own on-p emise se e s.
All he machines me he minimum equi emen s se which we e 12 h
gene a ion In el i9, AMD Ryzen 5900X o be e , 64 GB o RAM and, i
possible, a G aphics P ocessing Uni (GPU) wi h a leas 10 GB o RAM.
A de ailed desc ip ion o he implemen ed FL amewo k is p esen ed
in Sec ion 3.1. The INCISIVE FL amewo k will se e as he backbone
o he EUCAIM
3p ojec , which aims o se e as he main pla o m o
eu opean ede a ed esea ch in heal hca e. The final in as uc u e is
depic ed in Fig. 1, whe e he e a e wo main pa s, he cloud which con-
ained he copy o all da a, and he ede a ion which con ained he nodes
om each hospi al as men ioned be o e. Whene e we make e e ence
o FL, we e e o use he ede a ion in an FL se ing, no a simula ion
in he cloud.
2.2. Da ase s desc ip ion
In o de o ain he BIRADS classifica ion model, we u ilized da a
om h ee diffe en hospi al ins i u ions; Hellenic Cance Socie y, Uni-
e si y o No i Sad and A is o le Uni e si y o Thessaloniki. I should
be no ed ha he o me p o ided da a om mul iple hospi al ins i u-
ions, whe eas he la e wo had di ec access o da a om he hospi al’s
2h ps://incisi e-p ojec .eu/.
3h ps://cance image.eu/.
Fig. 2. Da ase spli pe ins i u ion. This spli is used on all he expe imen s.
uni e si y clinic. Th oughou his wo k, hese en i ies will be called Hos-
pi al 1, Hospi al 2 and Hospi al 3 espec i ely.
An ini ial da ase was c ea ed and u ilized as desc ibed in Fig. 2,
which highligh s h ee main a eas: a) da a collec ion and p ocessing,
b) da a combina ion and spli ing, and c) he aining p ocedu e. The
p ocessing mechanism, which will be desc ibed in de ail in Sec ions 3.2
and 3.3 is applied independen ly o each hospi al’s da a, esul ing in he
c ea ion o Da ase 1, Da ase 2, and Da ase 3.
The da ase c ea ion p ocess is shown in Fig. 2. Each o he fi s
wo da ase s is spli in o aining, alida ion, and es se s ollowing
he pe cen ages 70%, 20% and 10%, espec i ely. Al hough hese se s
con ain s andalone mammog ams wi hou any in o ma ion abou he
co esponding pa ien s o ela ed mammog ams, he spli is pe o med
on a pa ien -wise basis. This ensu es ha no in o ma ion leaks om he
aining se in o he alida ion and es se s. The aining and alida-
ion se s a e hen combined in o a single final aining se and a single
final alida ion se , espec i ely. The es se s, howe e , a e kep sepa-
a e, as i is c ucial o e alua e he model’s pe o mance o each da a
p o ide indi idually. Due o i s limi ed numbe o images, Da ase 3
is no included in he aining p ocedu e. Ins ead, i is con e ed in o a
single es se o e alua e he model’s pe o mance specifically on his
hospi al’s da a.
A e he obse a ion ha he pe o mance o he model in he ini ial
FL expe imen was subop imal (mo e de ails in Sec ion 4.3), he da ase s
we e e-e alua ed in consul a ion wi h he medical p o essionals om
each hospi al o iden i y whe he he issue is da a- ela ed. In pa icula ,
we ound ha 6 pa ien s had b eas implan s and 11 we e no co ec ly
scanned. The e o e hese we e emo ed. Mo eo e we ound ha he e
we e se e al issues such as some images including labels on hem, some
o hem we e scanned and some had su gical clips in he depic ion. I
should be no ed ha such a i ac s a e no uncommon in mammog a-
phy images [29]. These diffe ences did no affec he classic aining as
when all images a e pu oge he he e a e mo e o hese samples, bu
Compu a ional and S uc u al Bio echnology Jou nal 28 (2025) 106–117
108
I.N. Tzo zis, A. Gu ie ez-To e, S. Sykio is e al.
Fig. 3. Examples o images wi h a i ac s which affec model pe o mance.
B eas implan s (le ) in oduce a o eign body which con uses he model. Su -
gical clips ( igh ) a e depic ed as whi e small spo s wi h may be con used by he
model as abno mali ies.
Table 1
Final coun o images o each hospi al and BIRADS.
BIRADS H1 (G ) H2 (Se) H3 (G ) To al
0 0 0 0 0
1 30 10 0 40
2 212 297 21 530
3 0 33 0 33
4 136 43 6 185
5 112 72 22 206
6 0 0 0 0
To al Images 490 455 49 994
To al Pa ien s 132 145 17 294
i affec ed g ea ly in FL. F om he wo hospi als, he one ha had mo e
a iabili y in hese images was Hospi al 1, he one whose F1 sco e was
lowe . In ac , Hospi al 1 included images om smalle hospi als, he e-
o e including mo e a iabili y on he machine y pa ame e s (Physical
Image Pa ame e s (PIP)) used o ake he images, which can lead o ou-
ble when c ea ing a classifie as seen in Killim e al. [30]. In addi ion, i
was iden ified ha images be ween hospi als 1 and 2 had significan ly
diffe en pixel in ensi y anges, which was aken in o conside a ion o
imp o e FL model pe o mance. Fig. 3p esen s some examples wi h
imaging a i ac s which would affec model pe o mance and we e e-
mo ed.
Finally, medical p o essionals indica ed he need o emo e any im-
ages labeled BIRADS 0 o 6, as hese labels a e conside ed noisy and
would add con usion o he model. The final numbe o images pe BI-
RADS ype is he one ha can be seen in Table 1.
3. Ma e ials and me hods
In his sec ion we p o ide de ails on he me hods used. In pa icula
we fi s desc ibe wha FL is and hen we b iefly in oduce ou amewo k
o pe o m FL. A e ha , all he de ails ega ding he p e-p ocessing a e
gi en, including he B eas A ea De ec ion (BAD) p e-p ocessing s ep.
3.1. Fede a ed Lea ning
Fede a ed Lea ning (FL) is a way o aining Machine Lea ning (ML)
models wi hou ha ing o mo e he da a om whe e i is o iginally
s o ed. Ins ead, he model is ained whe e he da a is loca ed, i.e., he
hospi als, and hen combined cen ally. Fig. 4 ep esen s his aining
Fig. 4. FL p ocess. The p ocess s a s wi h he se e p o iding a base model
o e e y clien . Then he clien s ain hei local models. Finally, all he models
a e me ged. Then, i a new ound s a s, he me ged model is again sen o he
clien s o es a he p ocess.
schema. The FL se e is he coo dina o o he aining p ocedu e and
he one ha me ges he models ha a e p oduced in he FL clien s.
The FL clien s a e he pieces o so wa e ha un in he p emises o he
hospi als, whe e he da a is s o ed. In his way, he da a does no ha e o
lea e he p emises o he hospi al, a oiding unnecessa y da a ans e s
o machines ha migh no be us ed.
The p ocess wo ks as can be seen in Fig. 4. Fi s , he FL se e e-
ques s a aining p ocess o be pe o med, hen he FL clien s ain on
hei da a and e u n he model o he FL se e . Then, hese models a e
agg ega ed (e.g., a e aged). By doing so, i he p ocess is co ec ly con-
figu ed, he esul ing model summa ises he aining pe o med by he
clien s. All his p ocess is conside ed an FL ound and can be epea ed
se e al imes o ensu e he quali y o he model. A e all he ounds,
he final model is se ed by he FL se e .
In his wo k ou se up ollows he FedA g me ging unc ion [31].
This unc ion akes all he model weigh s (which is known as model
gene ally) and a e ages hem wi h all he weigh s p o ided by he all
he clien s/hospi als. Al hough he e a e exis ing amewo ks such as
Flowe [32], usually hey lack o in eg a ion wi h mode n esou ce man-
age s such as Kube ne es, which is based in con aine ized applica ions
ha can be un whe e e is needed. The e o e, o co e ha gap we de-
signed a language agnos ic Kube ne es based amewo k. The main idea
o ou amewo k4is o enable he de elope o code in wha e e lan-
guage hey wan and hen communica e h ough a s anda d Applica ion
P og amming In e ace (API). This p ocedu e has he ad an age ha
he sys em adminis a o s o each hospi al ha e jus o se up a machine
wi h an ope a i e sys em compa ible wi h Kube ne es and p o ide ba-
sic in o ma ion he ede a ion IT manage . This se up b ings cen alized
con ol o he so wa e deployed while eleasing he IT om he hospi al
om main aining and co ec ing mos o he e o s in he se ice.
As shown in Fig. 1, his in as uc u e is ac ually dis ibu ed. This
means ha he e is a coo dina ion cen al clus e and hen he diffe en
local clus e s/compu e s loca ed a he hospi als’ p emises. Fig. 5shows
he main componen s o he amewo k. Fi s o all, in he cen alized
clus e we can find he O ches a o , ha is in cha ge o managing all
he o he componen s. Then, he Model as a Se ice (MaaS) hos s all he
models and all in e media e da a ha is equi ed o be s o ed. Those
wo componen s a e s a ic in he in as uc u e and always p esen .
On he o he side, he FL Manage is deployed on demand. This means
ha when a aining p ocedu e is launched, he O ches a o deploys his
componen o manage he p ocess. Then, he O ches a o deploys all he
4h ps://gi hub.com/INCISIVE-T aining-In e ence-F amewo k.
Compu a ional and S uc u al Bio echnology Jou nal 28 (2025) 106–117
109
I.N. Tzo zis, A. Gu ie ez-To e, S. Sykio is e al.
Table 2
The DICOM codes u ilized o he p ocessing o he digi al mammog ams.
Code Desc ip ion Values Example Values
(0028, 1052) Rescale In e cep A ibu e Decimal S ing -1024, 0, 00
(0028, 1053) Rescale Slope A ibu e Decimal S ing 1, 01, 0.001
(2050, 0020) P esen a ion LUT Shape A ibu e Code S ing IDENTITY,
INVERSE
(0028, 1041) Pixel In ensi y Rela ionship Sign A ibu e Signed Sho -1, + 1
(0028, 0004) Pho ome ic In e p e a ion A ibu e Code S ing MONOCHROME1,
MONOCHROME2
(0028, 0030) Pixel Spacing A ibu e Decimal S ing [0.703125,
0.703125]
(0020, 0062) Image La e ali y A ibu e Code S ing L, R
Fig. 5. Fede a ed Lea ning amewo k diag am.
componen s o he Local da a p ocesso , howe e his is no deployed in
he cen al in as uc u e bu in all he clus e /compu e s loca ed a he
hospi als’ p emises. This allows us o ha e componen s ha a e able o
ain models wi h he da a s o ed in he local eposi o ies.
As seen in Fig. 5, he e a e h ee componen s in he Local da a p o-
cesso . Fi s , he FL Clien , in cha ge o managing he local aining and
coo dina ion wi h he cen al manage . Second, he P ocess Resou ce
Manage (PRM) which is in cha ge o managing all he esou ces and a -
chi ec u e specific p ocesses o make he in e ac ion wi h he amewo k
mo e anspa en o he use . And hi d, he AI Engine which con ains
he code ha he de elope builds. The AI Engine can be p og ammed in
any kind o language as long as i offe s an Hype Tex T ans e P o ocol
(HTTP) API ha is complian wi h he specifica ions.
3.1.1. Adap ing code o he amewo k
As ou amewo k wo ks wi h an API as main en y poin o he code
and i is agnos ic o he language and ools used o build he p e iously
men ioned AI Engine, he effo on adap ing he cen alized code o he
en i onmen is minimal. The modifica ions a e he ollowing:
•C ea e an API endpoin ha can be pinged o check i he ini ializa-
ion has finished and i is eady o ecei e eques s
•C ea e an API endpoin o un he diffe en use cases
•Include he unc ionali y o me ge models
•Include he unc ionali y o ain om an al eady ained model
(ini ialized weigh s)
No ice ha he diffe en use cases needed o be implemen ed a e he
ollowing: ain a new model, ain a model ha is p e ained, e alu-
a e he model, me ge models and pe o m in e ence. E en hough i is
echnically easy o adap he code o he amewo k, he e a e new de-
ails ha ha e o be aken in o accoun , like new pa ame e s such as
he numbe o FL ounds o pe o m. By de aul , he me ging unc ion
is p o ided by he amewo k i he use makes use o Py o ch, bu can
be cus omized by he use . Implemen ing bo h aining and me ging al-
low use s o pe o m whiche e ope a ion o use any model ype (e.g.,
andom o es s) hey wan ins ead o ha ing one se o de aul ype o
models.
3.2. Image p e-p ocessing
The DICOM files o he mammog ams o he da ase con ain essen-
ial in o ma ion no only o he echnical aspec s o he image bu also
o he con en o i . Thus, specific DICOM a ibu es, desc ibed in Ta-
ble 2, a e u ilized o he p e-p ocessing s eps.
DICOM code pai s (0028, 1041), (0028, 0004) and (2050, 0020)
should be e iewed o de e mine i he pixel alues need be in e ed. In
pa icula , i Pixel In ensi y Rela ionship Sign equals o 1 o Pho ome -
ic In e p e a ion A ibu e is se o “MONOCHROME1” o P esen a ion
LUT Shape A ibu e is assigned he alue “INVERSE”, hen each pixel
ge s in e ed by sub ac ing i s alue om he maximum pixel alue o
he image.
A his second s ep, he pixel a ay should be ho izon ally flipped
o align wi h he p ede e mined la e ali y, he eby elimina ing an addi-
ional ac o ha could in oduce bias in o he model. Technically, he
Image La e ali y A ibu e (0020, 0062) is inspec ed and i i is assigned
he alue “R”, he image will be flipped. In his way, he a ea o in e -
es ( he b eas a ea) will always be loca ed a he le side o he pixel
a ay.
As mammog ams a e used o de ec e y small lesions in sc eening,
hey can be o high esolu ion such as 3328 × 4096 pixels [33]. How-
e e , hese sizes make he aining p ocedu e eally challenging om
he pe spec i e o compu ing capaci y. Thus, ha ing compu a ional con-
s ain s due o limi ed esou ces, i is commonly accep ed o use educed
esolu ion images o eeding Machine Lea ning model du ing he ain-
ing p ocess. Aiming o achie e a p ope esizing o he images, i is
c ucial o engage he Pixel Spacing A ibu e (0028, 0030) o p ese -
ing spa ial ela ionships, main aining accu a e measu emen s, sus ain-
ing consis ency ac oss diffe en de ices and esolu ions and imp o ing
image analysis accu acy.
3.3. B eas a ea de ec ion
I is no a e o find p in ed labels on mammog ams, especially when
alking abou digi ized e sions. Digi al images a e usually clea bu ,
e en in such cases, labels o anno a ions may exis . These labels a e
mos ly ela ed o he la e ali y o he b eas o he iew o he mammo-
g am. Such elemen s can po en ially influence he lea ning capabili y o
he model du ing he aining p ocedu e; no only hey end o add an-
o he bias o he da ase bu , also, he ex ac ion o he ac ual egion o
in e es is ge ing much mo e difficul .
To add ess his p oblem, we in oduce he B eas A ea De ec ion
(BAD)5 ool which akes as inpu a mammog am and gene a es a clea ed
5h ps://gi hub.com/i zo zis/b eas _a ea_de ec ion.
Compu a ional and S uc u al Bio echnology Jou nal 28 (2025) 106–117
110
I.N. Tzo zis, A. Gu ie ez-To e, S. Sykio is e al.
Fig. 6. Pa hway om he ini ial INb eas da ase o he finalized AULD da ase .
e sion as ou pu , emo ing he useless in o ma ion by he ini ial image.
I consis s o h ee main pa s; a) he unsupe ised b eas a ea masks
gene a ion, b) he da a augmen a ion and c) he supe ised b eas a ea
mask gene a ion.
3.3.1. P epa ing he BAD da ase
Two auxilia y da ase s ha e been gene a ed o aining and e alua -
ing he BAD ool; a) he Unsupe ised Lea ning Da ase (ULD) and b) he
Augmen ed Unsupe ised Lea ning Da ase (AULD). Fo he pu poses o
his ask, he INB eas [34] mammog am da ase is u ilized. I includes
410 digi al mammog ams ha co espond o 115 cases, 90 o hem a e
women wi h bo h b eas s affec ed, while he emaining 25 women ha e
unde gone mas ec omy.
These images a e he inpu s o he K-means clus e ing algo i hm ha
gene a es he ini ial b eas anno a ion masks. Due o he unce ain y in-
oduced by his me hod, he p oduced anno a ion masks a e isually
inspec ed in o de o selec only he p ope ones. The ou pu o his se-
lec ion ask e e s o he ULD. In an aim o gene alize he capaci y o
he ool, he idea o a supe ised segmen a ion model is in oduced. To-
wa ds his di ec ion, he idea o augmen ing his ini ial da ase (ULD)
is eme ging due o a) he limi ed amoun o da a (less han 410 images
since we ha e elimina ed some a e he clus e ing p ocedu e) and b)
he demand o gene a e andomly sized and posi ioned boxes simula -
ing, in his way, he undesi able labels. Fo each selec ed mammog am,
fi e a ia ions a e used o he final augmen ed da ase ; he fi s co -
esponds o he o iginal one while he es co espond o copies o he
o iginal including a andomly-sized and andomly-placed, ake label.
This ou e om he o iginal INb eas da ase o his final AULD is de-
pic ed in Fig. 6. E en ually, his app oach pushes he UNe model o
lea n de ec ing he a ea o in e es (b eas a ea) in a mo e a ge ed way.
A he same ime, he gene a ed da ase con ains mo e han 1500 im-
ages which a e conside ed o be sufficien o he aining o he BAD
model.
Table 3
Compa a i e pe o mance e alua ion o
BAD on he Inb eas da ase .
Model Name Dice Sco e IoU
Th esholding [37] 94.70 90.06
MedSegDiff [38] 98.37 96.82
SAM [39] 92.04 85.65
SAM-adap e [35] 99.05 98.13
SAM-b eas [36] 99.27 98.55
BAD 98.48 97.01
3.3.2. T aining p ocedu e and usage
The aining p ocedu e is s aigh o wa d since he da ase is clea ,
a ge ed o he a eas o in e es and, hus, he hype -pa ame e s fine
uning is conside ed an easy ask. 1000 images a e used o he ain-
ing o he model, 300 o he alida ion a he end o each epoch and
he es o he es se o p ac ically e alua e he pe o mance o he fi-
nal model. Table 3con ains compa a i e e alua ion o BAD wi h o he
simila ools in he li e a u e. E en hough i s pe o mance is sligh ly
wo se han SAM-based ools [35,36], he compu a ional complexi y o
SAM is significan ly highe . The added compu a ional complexi y does
no wa an he small pe o mance inc ease.
A e aining he UNe model, he BAD ool’s in e ence can be u i-
lized as a s andalone componen wi hin he p ocessing pipeline. As
shown in Fig. 7, he basic p e-p ocessing me hodology is applied o he
o iginal mammog am, and he ou pu is hen passed o he BAD ool’s
in e ence mechanism. The model’s p edic ion, in he o m o segmen a-
ion, is combined wi h he o iginal mammog am o gene a e a cleane
e sion. Fu he p ocessing can be applied o his final ou pu by pe -
o ming simple pe iphe al c opping, which is easible since he e a e
no labels o o he a i ac s. This esul s in a educ ion o unnecessa y
in o ma ion.
Fig. 8p esen s a isual o e iew on he impac o all p e-p ocessing
s eps on an example image. A e po en ial in e sion and clipping, he
Compu a ional and S uc u al Bio echnology Jou nal 28 (2025) 106–117
111
I.N. Tzo zis, A. Gu ie ez-To e, S. Sykio is e al.
Fig. 7. The BAD ool wi hin he image p ocessing pipeline, as desc ibed in Sec ion 3.2.
Fig. 8. Visual ep esen a ion o he impac o all p e-p ocessing o an example image om he Inb eas [34] da ase . The image was con as -in e ed o demons a ion
pu poses.
image is esized o 512x512 o be gi en as inpu o he BAD classifie .
The BAD ou pu is hen used o c op he image so ha he backg ound
a ea is minimized, as i does no p o ide any meaning ul in o ma ion
o he classifica ion model. Finally, he esul ing image is esized o
256x256 o se e as inpu o he BIRADS classifica ion model.
4. Resul s
4.1. Expe imen al se up
To e alua e ou amewo k on he BIRADS p edic ion model ain-
ing p ocedu e, we pe o med a wo-s ep expe imen . Fi s , he model is
ained in a cen alized/classic manne , whe e all da a a e ga he ed in
one place. We hen pe o med model aining using Fede a ed Lea ning,
whe e he da a is dis ibu ed in o sepa a e nodes/hospi als, o compa e
bo h app oaches. In o ma ion abou he da a and he spli s is de ailed
in Sec ion 2.2. The model a chi ec u e is based on Con olu ional Neu al
Ne wo ks (CNN), as shown in Fig. 9. We ha e used con olu ional lay-
e s since hey do no equi e a la ge numbe o pa ame e s and, hus,
hey do no c ea e a communica ion bo leneck when ans e ing he
models on each FL ound.
The objec i e is o quan i y how well do he ained models pe o m
o e he same da ase wi h a simila configu a ion. The only pa ame e
ha a ied om model o model was he numbe o epochs, i.e., num-
be o imes ha he model “sees” he comple e da ase . In he cen al
e sion, we ha e used 100 epochs, whe eas in he ede a ed e sion we
ha e used 3 diffe en combina ions: 100 ounds -1 epoch, 50 ounds -
2 epochs and 20 ounds -5 epochs.
Compu a ional and S uc u al Bio echnology Jou nal 28 (2025) 106–117
112
I.N. Tzo zis, A. Gu ie ez-To e, S. Sykio is e al.
Fig. 9. The s uc u e o he model used o BIRADS classifica ion.
Fig. 10. Model pe o mance pe hospi al du ing he ini ial FL adap a ion.
To pe o m he expe imen s we used he wo in as uc u e ypes
p esen ed in Sec ion 2: a) a cen alized eposi o y con aining all he
equi ed da a and b) a ede a ed in as uc u e. The ede a ed in as-
uc u e main se e was hos ed in he same cen alized clus e , whe eas
he clien s we e hos ed in a machine loca ed in each hospi al p emises
(G eece and Se bia) o es how a eal li e en i onmen can wo k, includ-
ing ne wo k delays, diffe ence o compu ing powe and o he issues ha
migh happen on a complex se up like his.
4.2. Classic expe imen
A his s age, a no maliza ion p ocedu e is applied o all se s based
on he minimum and maximum pixel alues o he aining se . The
aining se is di ided in o ba ches and ed in o he model o a single
epoch o aining. Following his, an e alua ion p ocess is conduc ed
o e he ba ches o he alida ion se . Subsequen ly, a decision is made
acco ding o he ollowing algo i hm: i he cu en aining epoch has
no exceeded he p edefined o al numbe o epochs (100 in his case),
he p ocedu e (single epoch aining and alida ion) is epea ed. O he -
wise, he aining is s opped, and he model is e alua ed on he es se s.
Table 4p esen s he se o hype pa ame e s used du ing model aining.
4.3. Adap a ion o he p oblem o Fede a ed Lea ning
4.3.1. Ini ial adap a ion
Wi h p e iously men ioned model hype pa ame e s, se e al expe i-
men s we e pe o med ega ding he FL, mainly he ade-off be ween
local epochs and FL ounds. The op imal choice in his case was 50 FL
ounds wi h 2 local epochs. To adap he classic e sion, he code was
Table 4
Bes hype pa ame e se ound in he
ini ial expe imen .
Hype pa ame e Value
Loss unc ion C oss-en opy loss
Lea ning a e 10−8
Op imize Adam
Ba ch size 10
Epochs 100
modified sligh ly o adap i o he FL amewo k, howe e only com-
munica ion pa s we e modified, he aining pa was un ouched.
As can be seen in Fig. 10b, he F1 sco e o he Hospi al 2 (in o -
ange) is usually highe and ha happens bo h in he cen al and he
ede a ed expe imen s. No ice ha he sco e is sligh ly lowe in he ed-
e a ed expe imen , which is expec ed in FL se ings. On he o he hand,
Hospi al 1 (in eal) suffe s om a se ious dec ease in pe o mance com-
pa ed o he classic expe imen (∼ 40% dec ease in F1). As we can see
in Fig. 10a, om ound 4 he model d i s away om he es se o
Hospi al 1, which means ha he e a e e y diffe en images ha a e
no accoun ed in aining o Hospi al 1 o o Hospi al 2, as in each
ound he model is synch onized. The e o e we need o go deepe o
unde s and wha is happening.
4.3.2. Imp o ed p e-p ocessing adap a ion
A e he p ep ocess, he FL aining was pe o med in he same se -
ings. No ice ha only Hospi al 1 and 2 a e used o he aining as in he
p e ious expe imen . We include Hospi al 3 only o be used in model
es ing as seen in Fig. 2. As seen in Fig. 11a, pe o mance o bo h, he
Compu a ional and S uc u al Bio echnology Jou nal 28 (2025) 106–117
113
I.N. Tzo zis, A. Gu ie ez-To e, S. Sykio is e al.
Fig. 11. Resul s o he final expe imen including he whole p ep ocessing pipeline.
Table 5
Resul s o he hype pa ame e sea ch on he ade-off be ween Rounds and
Epochs, showing he F1-sco e o each hospi al. The 3 solu ions a e almos
iden ical as he model con e ges o simila weigh s.
Rounds Epochs Hospi al 1 Hospi al 2 Hospi al 3 T ain ime (min)
100 1 0.549 0.743 0.559 260
50 2 0.550 0.743 0.559 192
20 5 0.549 0.743 0.559 247
classic expe imen and he combina ion o 50 ounds and 2 epochs is
inc eased wi h he ull p ep ocess. The pe o mance o he new Hos-
pi al 3 (da k blue) is simila o Hospi al 1, showing gene aliza ion o
he app oach. Howe e , he same happens when we analyze he es o
ound/epoch combina ions. Again, we es ed 3 diffe en combina ions:
100 ounds -1 epoch, 50 ounds -2 epochs and 20 ound -5 epochs.
This means ha in gene al we will do 100 epochs in each combina ion.
As can be obse ed in Table 5, in his pa icula p oblem, all he com-
bina ions con e ge o simila solu ions. As seen in he able and figu e,
now Hospi al 1 accu acy is no a om he classic e sion. In pa icu-
la , Fig. 11b shows he es F1 sco e while aining o all combina ions.
Now Hospi al 1 does no all in o e fi . The as es solu ion was he 50-2
combina ion as seen in Table 5. Howe e , his is because he in as uc-
u e is being sha ed and is dependen o he ne wo k and machine y
om o he ins i u ions. Usually, he as es app oach should be he 20-5
as i equi es less communica ion be ween he nodes.
5. Lessons lea ned and ecommenda ions
Fede a ed Lea ning (FL) is a e y in e es ing esea ch opic ha
p omises benefi s o si ua ions in which is impossible o sha e he da a.
Howe e , i canno be seen as a plug and play solu ion, since many
de ails needs o be ca e ully conside ed o ensu e success ul implemen-
a ion. In his Sec ion, we summa ize impo an lessons lea ned, and
p opose mi iga ion ac ions o specific encoun e ed p oblems.
•Inhomogeneous da a leads o inaccu a e models: gi en he na-
u e o da a, da a om diffe en p o ide s can be diffe en , as can
be seen om ou case s udy as well as o he wo ks [30]. Va ia-
ions migh no affec a human adiologis , bu hey migh ha e a
big impac on he accu acy o he models p oduced. This ac is ac-
cen ua ed in FL, as we a e building models independen ly in each
Da a P o ide , we migh find a si ua ion in which he models do
no con e ge o ha a e biased owa ds some conc e e popula ion.
Da a homogenei y may also di ec ly impac FL hype pa ame e un-
ing, as he local model may o e fi he espec i e p o ide ’s da a,
hus impac ing model agg ega ion and hampe ing i s gene aliza ion
capabili ies [40].
Mi iga ion ac ions: S anda dized da a collec ion p o ocols ac oss
da a-sha ing ins i u ions should be es ablished o ensu e consis-
ency in da a o ma s, imaging modali ies and anno a ions. These
p o ocols should be enhanced h ough da a anno a ion p ocedu es
and guidelines, which should be commonly-defined by medical p o-
essionals. Finally, each hospi al/medical ins i u ion should epo
da ase me ada a (e.g. scanne ype, acquisi ion se ings) alongside
images o enable model adjus men s o po en ial domain shi s.
•Da a should be homogenized: ollowing he p e ious poin , ha -
moniza ion is he mos impo an pa o AI medical se ice de el-
opmen . One o he main challenges ha he AI de elope s ace
du ing model de elopmen is he high deg ee o ha moniza ion e-
qui ed be ween diffe en Da a P o ide s. One canno expec ha
Da a P o ide s om diffe en ins i u ions will use he same equip-
men o da a gene a ion. This ac is explo ed by Kilim e al. [30]
whe e hey explo e he pa ame e s om diffe en imaging machines
and diffe en hospi als. They saw ha he e is a clea diffe ence be-
ween images om diffe en cen e s and used he pa ame e s om
he machines o imp o e he accu acy. The e o e, ha moniza ion o
medical images be ween Da a P o ide s is one o he mos impo -
an s eps in he de elopmen pipeline, and can “make o b eak”
he effec i eness and gene aliza ion capabili ies o a model. Fol-
lowing Kilim e al. [30], o he impo an s ep o do is o keep he
image me ada a ega ding he machine y used. This me ada a can
be used o he ha moniza ion o a e wa ds o he p e-p ocessing
o he da a in he FL en i onmen .
Mi iga ion ac ions: The esea che shall ensu e he s anda diza-
ion o p e-p ocessing pipelines ac oss all pa icipa ing ins i u ion.
On a communica ion le el, in o ma ion exchange be ween nodes
(e.g. no maliza ion alues, da ase size e c.) should be enabled. In
addi ion, he esea che should in es iga e he u iliza ion o da a
augmen a ion and ha moniza ion echniques o educe a iabili y.
Finally, he egula o y bodies a e encou aged o define minimum
da ase s anda d o FL pa icipa ion (e.g. image esolu ion, label-
ing consis ency).
•The da a should be cu a ed p io FL aining: e en hough Fed-
e a ed Lea ning is abou secu ing he da a by no showing i o he
de elope s, in he end de elope s need o ha e a glimpse on he
da a o unde s and how o wo k wi h i and o debug he models.
This could be done wi h agg ega ions ha ulfill he p i acy e-
Compu a ional and S uc u al Bio echnology Jou nal 28 (2025) 106–117
114
