AI4Li e – Policy B ie – 27.10.2023
POLICY BRIEF
RESEARCH INFRASTRUCTURES
INTRODUCTION: PROJECT TITLE AND PROJECT OBJECTIVES
CALL: HORIZON-INFRA-2021-SERV-01
TOPIC: HORIZON-INFRA-2021-SERV-01-06
PROJECT: A i icial In elligence o Image Da a Analysis in he Li e Sciences (AI4Li e).
h ps://ai4li e.eu obioimaging.eu/
Objec i es
(1) Democ a ized FAIR access o AI-based image analysis me hods h ough a common
applica ion in e ace s anda d – which we call he BioImage Model Zoo (BMZ) –
compu a ionally powe ed by he Eu opean Open Science Cloud (EOSC) in as uc u e. The
AI4Li e se ice landscape will make e en he la es and mos ad anced me hods ins an ly
a ailable o li e scien is s, pa icula ly o hose wi hou in-dep h compu a ional expe ise.
Addi ionally, ou se ices will be a ailable o local ins alla ion (“on-p emise”).
(2) Es ablish s anda ds o he submission, s o age and FAIR access o e e ence da a,
e e ence anno a ions (g ound- u h), ained AI models, and ainable AI me hods.
This will happen in close pa ne ship wi h exis ing Eu opean in as uc u es, e.g., in he
BioImage A chi e (BIA) a EMBL-EBI and he Eu opean Open Science Cloud (EOSC). The
BioImage A chi e is an in eg al pa o he image da a se ices o e ed by Eu o-BioImaging.
The access o ou in as uc u es will be i ual, open, and ee o cha ge o all in e es ed
use s in indus y and academia. All es ablished s anda ds will be open o os e
compa ibili y and syne gy also o AI-d i en indus y e o s.
(3) Simple model deploymen , sha ing, and dissemina ion o AI-based me hods as a
new de elope - acing se ice o he BioImage Model Zoo. Ou s anda ds and se ices
will enable compu a ional scien is s om he li e sciences, AI, ML and Compu e Vision
ields o each a la ge audience o li e science use s o hei la es AI-based analysis
me hods. This will allow compu a ional scien is s o con inuously push he bounda y o wha
AI-based me hods can achie e in he con ex o scien i ic imaging and image da a
analyses. While oday he deploymen and dissemina ion o a new me hod o ool is
labo ious and ime consuming, ou in as uc u es and se ices will educe his o e head
ime and g ea ly accele a e he a e o compu a ional inno a ion in academia and in SMEs.
(4) O ganise Open Calls and Challenges o ou s anding image analysis p oblems in he EU
Mission a eas o he Ho izon Eu ope amewo k p og amme. Building on he common
p ac ice o he AI and Compu e Vision ields o cas impo an analysis p oblems in o
well-de ined challenge asks on a p epa ed body o da a, wi h (a) gi en e e ence baseline
solu ion(s), hese compu a ional disciplines will engage in a iendly compe i ion o ind
be e and be e pe o ming me hods o he mos impo an p oblems. Bes pe o ming
app oaches o each Challenge we o ganise will immedia ely become a ailable h ough ou
BioImage Model Zoo in as uc u e, ollowing he pionee ing examples o he Human
P o ein A las (Ouyang, Winsnes, e al. 2019) image classi ica ion compe i ions which had
housands o pa icipa ing eams wo ld-wide.
Implemen a ion o esea ch in as uc u es.
AI4Li e is commi ed o enhancing he BioImage Model Zoo (BMZ), an AI model eposi o y ha
p o ides li e scien is s wi h democ a ised access o AI-based image analysis me hods and FAIR
image analysis in as uc u e se ices. In his con ex , AI4Li e has de eloped he BioEngine, a
cloud-based amewo k designed o acili a e he execu ion o p e- ained AI models di ec ly om
he BMZ websi e, which allows all use s - independen o hei le el o compu a ional expe ise o
local compu e esou ces - o es and un models o hei image da a analysis. Howe e , limi ed
compu e esou ces a ailable o he p ojec conso ium s ill ep esen a bo leneck on he way o
allow use s o ully scale up image analysis h ough he BMZ websi e. AI4Li e is he e o e in ending
o secu e compu a ional esou ces h ough he Eu opean Open Science Cloud (EOSC). Also, see
ecommenda ion 7.
Fu he mo e, o engage de elope s wi h la ge and/o well-known p e- ained model collec ions as
communi y pa ne s o AI4Li e in as uc u e, dedica ed e o s o b inging hei model collec ions
in o a FAIR and compa ible s a e would be needed.
RECOMMENDATION 1: Implemen Open Calls wi h cascading g an s o suppo model
de elope s wi h la ge and/o well-known p e- ained model collec ions, enabling hem o b ing he
model collec ions in o a FAIR-complian s a e h ough he BMZ.
RECOMMENDATION 2: Implemen Open Calls wi h cascading g an s suppo o ea ly
adop e s and so wa e ool de elope s o con ibu e o he ma u a ion o he BMZ by achie ing
ools in e ope abili y. This will help os e de elopmen o he BMZ in o a communi y in as uc u e
isible in he esea ch so wa e de elopmen a ena as well.
Access o esea ch in as uc u es.
AI4Li e is suppo ing use s ( h ough di ec suppo ) and de elope s ( h ough con ibu o se ices)
a ound he BMZ, imp o ing accessibili y o he in as uc u e. Toge he wi h a ge ed dissemina ion
and inc eased usage o he esou ces h ough Open Calls ha a e open o scien is s om all li e
science domains wo ldwide, he p ojec will b ing no el use s h ough he in as uc u e o he
EOSC.
AI4Li e – Policy B ie – 27.10.2023
(5) Empowe common image analysis pla o ms by AI ools. We will in eg a e all analysis
models in he BioImage Model Zoo in o popula end-use so wa e ools. We will s a wi h
some o he a guably mos common so wa e ools buil and main ained by membe s o ou
conso ium (i.e., Fiji, ilas ik, ImJoy, deepImageJ, Ze oCos DL4Mic). This uly makes he
sha ed communi y models in he Zoo eusable and will enable all model- empowe ed
so wa e ools o o e be e and e e -g owing analyses. Addi ionally, we will suppo
de elope s o o he so wa e ools ha a e no ye pa o ou conso ium o connec o ou
Model Zoo and become a consume (and/o p oduce ) o ained models and analysis
me hods hemsel es.
(6) O ganise ou each and aining e en s o li e scien is s as a ge ed image analysis
cou ses and wo kshops as well as by pa icipa ion in he la ges in e na ional con e ences.
By in ol ing a wide se o li e sciences Resea ch In as uc u es (RIs) in he communi y
ou each ac i i ies, AI4Li e will ensu e b oad pa icipa ion, con ibu ion and up ake by
imaging scien is s oo ed in a ious scien i ic disciplines, including o example s uc u al
biology, medicinal chemis y, ma ine biodi e si y and plan pheno yping. Ou each and
eaching e en s o me hod de elope s and po en ial con ibu o s o Open Calls and
Challenges a e equally impo an o c ea e a ib an communi y and p oduc i e
in e disciplina y discou se du ing he du a ion o he p oposed p ojec .
POLICY IMPLICATIONS AND RECOMMENDATIONS
RECOMMENDATION 3: Implemen Open Calls wi h cascading g an s o suppo image
analys s and ool main aine s and de elope s o ca e he needs and onboa d new use s.
In e na ional co-ope a ion o esea ch in as uc u es.
To acili a e he de elopmen o communi y s anda ds o image da a and o AI models, he AI4Li e
p ojec coo dina es in e ac ions wi h a ious esea ch in as uc u es, like EU-Opensc een, Ins uc ,
EMBRC, EMPHASIS and da a-hea y RIs like ELIXIR. Speci ic in e ac ion wi h he in e na ional
communi y is media ed h ough ne wo ks, like Global BioImaging, NEUBIAS, COMULISglobe,
olumeEM, e c. Addi ionally, we a e pa ne ing wi h esea ch eams and o ganisa ions beyond
Eu ope, such as Bioimaging No h Ame ica (BINA), Janelia (e.g. COSEM o Jan Funke) and he
BROAD ins i u e (e.g. wi h Be h Cimini). The aim o such discussions is o align on s anda ds,
a oid duplica ion o e o s, ensu e c oss-compa ibili y, and in eg a e new ools c ea ed by ex a
Eu opean g oups and agencies. Fu he dissemina ion o ac i i ies is planned h ough newly
unded p ojec s like oundingGIDE. These in e ac ions unde pin he EU policy on Open Science
suppo ing he de elopmen o image da a as a global esou ce.
Employmen and skills in esea ch in as uc u es.
AI4Li e has con ibu ed a ge ed cou se modules on he BMZ a nume ous ypes o e en s and
occasions (e.g. EMBO cou ses on image analysis, DL4MIA cou se a HT, DL o Image Analysis a
EMBL o he DL@MBL deep lea ning cou ses a he Ma ine Biology Labo a o ies in Woods Hole,
USA). Fo he compu a ional communi y, AI4Li e has o ganised hacka hons, whe e nume ous
in e na ional pa icipan s con ibu ed o shaping he AI4Li e me ada a s anda ds and he
in as uc u e behind he BioImage Model Zoo. Th ough he de elopmen and dissemina ion o
sha ed communi y s anda ds, AI4Li e suppo s and p omo es good p ac ices o he submission,
s o age and access o bo h image da a and AI models. Fu he mo e, he in as uc u e se ices
p o ided by AI4Li e acili a e enhanced image analysis, accele a ing he pace o esea ch wi hin
he imaging communi y.
RECOMMENDATION 4: Cul i a e communi y collabo a ion by con inued o ganisa ion o
aining, de elope and dissemina ion e en s o ga he communi y eedback, os e communi y
engagemen and con ibu ions, and he eby enhance he BMZ capabili ies.
G eening o esea ch in as uc u es.
By p omo ing he de elopmen o in e ope able s anda ds o p e- ained AI models AI4Li e
suppo s ep oducible and e icien science and allows he euse o scien i ic ou pu s, p olonging
he u ili y o exis ing esou ces. In addi ion, by p o iding access o sha ed compu e in as uc u e o
un he models in a use - iendly way, AI4Li e p omo es lexible usage o hese esou ces
elimina ing he need o duplica e in as uc u e in o local ins ances. Fu he mo e, s o ing anno a ed
images in he o m o AI- eady da ase s minimises he need o e un esou ce-in ensi e models o
p oduce ou pu s.
In e ac ion o esea ch in as uc u es wi h indus y.
AI4Li e main ains in e ac ions wi h key imaging and pha ma indus y ep esen a i es such as
Baye , Zeiss and Leica. Th ough hese con ac s, AI4Li e aims o p omo e open s anda ds,
compa ibili y and in e ope abili y a all le els and o engage indus y pa ne s o con ibu e o and
eely consume all a ailable analysis models (AI-based ools and me hods) in he BMZ se ice
po olio. A p ime example o hese collabo a ions is he in eg a ion o BMZ models in o Leica’s
AIVIA mic oscopy image analysis so wa e.
Technology de elopmen , da a and digi al se ices, digi alisa ion.
Th ough he BMZ in as uc u e se ices, he ela ed s anda disa ion o AI models and me hods
and he mul i ace ed aining and ou each ac i i ies, AI-based me hods will become an essen ial
and easily accessible echnology o cu ing-edge, imaging-hea y li e sciences esea ch. New
AI4Li e – Policy B ie – 27.10.2023
s a e-o - he-a me hods will immedia ely p opaga e o he communi y and connec mic oscope and
me hod de elope s o p ac i ione s in esea ch labo a o ies and acili ies.
RECOMMENDATION 5: Alloca e esou ces o suppo da a euse o p omo e imp o ed,
accele a ed and sus ainable echnology and se ice de elopmen s.
Le el o connec ion o you RI o EOSC.
AI4Li e suppo s making image da a FAIR and AI- eady h ough a communi y-d i en p ocess.
Cu en ly, such da ase s a e openly a ailable h ough he BioImage A chi e, while he guidelines,
oge he wi h o he p ojec ou pu s a e indexed in he EOSC esou ces h ough OpenAIRE.
Besides FAIR da a, AI4Li e is o e ing digi al se ices (e.g. he BMZ) and will in he u u e o e
web-based human-in- he-loop AI model execu ion, da a labelling, model e aining, e c. The p ojec
also in e ac s wi h o he c oss-domain da a-d i en and EOSC p ojec s, like iMagine, ANERIS,
EUCanImage, and EOSC4Cance , h ough i s ex ensi e ne wo k o RI pa ne s.
RECOMMENDATION 6: Recognize he bene i s o di ec use suppo o making image da a
and analysis me hods FAIR h ough da a s ewa dship se ices. Hence, bo h unding and policy
suppo o ecognise hese e o s will be c ucial o unde pin EU policy on Open Science.
RECOMMENDATION 7: Es ablishing mechanisms o accessing on-demand compu e
esou ces oge he wi h echnical suppo and compu e oppo uni ies ia he EOSC is impo an o
sus ain he AI4Li e in as uc u e ha uses esou ces in ensi ely, bu o en in bu s s.
Implemen a ion o access beyond a basic compu e con ingen on a pe -use basis ia exis ing
ag eemen s o h ough compu e okens will need o be enabled.
Sus ainabili y o esea ch in as uc u es.
The key ou pu s o he p ojec ha suppo he sus ainabili y o he in as uc u e as a whole
include:
- BioImage Model Zoo: This web se ice makes anno a ed p e- ained ML/DL models
a ailable o esea che s and in e aces wi h he backend compu e esou ces. AI4Li e is
ex ending he communi y pa ne s and usabili y o his esou ce, making key con ibu ions
o inc ease i s usage and communi y up ake.
- BioEngine: This backend in as uc u e empowe s he BMZ wi h he capabili y o un he
p e- ained models on any image da ase . Tha way, he AI4Li e p ojec p o ides access o
ex ended compu e esou ces, inc easing i s capaci y o accommoda e a g owing use load.
- Model, da a and anno a ions s anda ds: To p omo e he wide use o BMZ models he AI4li e
p ojec is de eloping s anda ds o p omo e in e ope abili y o he in as uc u e. These
s anda ds a e collabo a i ely de eloped wi hin he communi y, which ensu es he long- e m
sus ainabili y o he ou pu s.
- Open Calls: While he di ec suppo o he use p o ided ia he Open Calls encou ages
he use o he BMZ in as uc u e, he open call playbook o execu ion and p ojec selec ion
will be adop ed in u u e p ojec s o un Open calls.
AI4Li e has ecei ed unding om he Eu opean Union’s Ho izon Eu ope esea ch and
inno a ion p og amme unde g an ag eemen numbe 101057970. Views and opinions
exp essed a e howe e hose o he au ho (s) only and do no necessa ily eflec hose o he
Eu opean Union o he Eu opean Resea ch Execu i e Agency (REA). Nei he he Eu opean
Union no he g an ing au ho i y can be held esponsible o hem.
AI4Li e – Policy B ie – 27.10.2023
