Co esponding au ho : Zineb El Ayachi
Copy igh © 2025 Au ho (s) e ain he copy igh o his a icle. This a icle is published unde he e ms o he C ea i e Commons A ibu ion Liscense 4.0.
A i icial In elligence in adia ion oncology: A sys ema ic li e a u e e iew o cu en
impac and u u e di ec ions
Zineb El Ayachi *, Samia Khal i, Kaou a Soussy, Wissal Hassani, Fa ima Zah a Fa hane Zenab Alami and
Tou ia Bouha a
Depa men o Radia ion The apy, Oncology Hospi al, HASSAN II Uni e si y Hospi al, Facul y o Medicine and pha macy
Fez, Uni e si y Mohammed Ben Abdellah, Fès 30000, Mo occo.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 1330-1337
Publica ion his o y: Recei ed on 09 July 2025; e ised on 16 Augus 2025; accep ed on 18 Augus 2025
A icle DOI: h ps://doi.o g/10.30574/wja .2025.27.2.2991
Abs ac
Radia ion oncology gene a es as amoun s o da a a e e y s ep o ca e— om simula ion and con ou ing o planning,
deli e y, and ollow‑up—c ea ing e ile g ound o a i icial‑in elligence ools ha can sho en wo k lows, s anda dize
decisions, and link ea men o ou comes. We pe o med a PRISMA‑guided sys ema ic e iew o he li e a u e
(PubMed, Embase, Scopus, Web o Science, IEEE Xplo e, and a Xi ; Janua y 2000–July 2025) o iden i y s udies ha
applied machine‑ o deep‑lea ning me hods o segmen a ion, ea men ‑planning dose p edic ion, syn he ic CT o CBCT
enhancemen , quali y assu ance, mo ion acking, adiomics‑based p ognosis, o adap i e adio he apy. A e
dual‑ e iewe sc eening o 33 eco ds, 18 s udies me inclusion c i e ia o he co e syn hesis and 15 we e e ained as
con ex ual backg ound. The mos obus e idence—and he g ea es ex e nal alida ion—was ound o supe ised
au o‑segmen a ion: one mul i‑ins i u ional NSCLC s udy included mo e han 2,000 pa ien s, and a e‑analysis o RTOG
0617 showed ha deep‑lea ning hea con ou s al e ed mean hea dose and s eng hened dose‑su i al associa ions.
Deep‑lea ning dose‑p edic ion and au oplanning wo k lows achie ed plan quali y compa able o expe planne s while
ma kedly educing planning ime. Syn he ic CT and CBCT co ec ion imp o ed dose calcula ion and image egis a ion
in adap i e wo k lows, and p edic i e quali y‑assu ance models showed p omising sensi i i y and speci ici y.
Radiomics s udies equen ly epo ed high in e nal pe o mance bu seldom p o ided ex e nal alida ion o
calib a ion. O e all, a i icial in elligence is al eady clinically use ul o au o‑segmen a ion and planning assis ance;
howe e , b oad deploymen will equi e mul i‑cen e ex e nal alida ion, sys ema ic calib a ion, d i moni o ing, and
ou come‑linked p agma ic ials embedded wi hin a lea ning‑heal h‑sys em amewo k.
Keywo ds: Cone‑beam CT; Adap i e adio he apy; Radiomics; Dose p edic ion; Au o‑segmen a ion; Radia ion
oncology; Deep lea ning; A i icial in elligence
1. In oduc ion
Radia ion oncology (RT) has always been da a‑ ich, ye ea ly apid‑lea ning isions s uggled o ansla e knowledge
in o p ac ice [1]. Founda ional p oposals o link elec onic heal h eco ds o ou comes o eshadowed oday’s AI
pipelines [2]. Recen policy pape s highligh g owing en husiasm o a i icial in elligence in RT depa men s wo ldwide
[3].This op imism es s on b eak h oughs in deep lea ning ha allow compu e s o ex ac obus ea u es om images
[4] and e en mas e complex decision spaces such as he game o Go [5]. Agains his backd op, we sys ema ically
e iewed AI applica ions ac oss he RT wo k low, emphasizing s udies ha p o ide ex e nal alida ion o clinical
impac .
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 1330-1337
1331
2. Me hods
This sys ema ic e iew ollowed he PRISMA 2020 amewo k. A p o ocol was d a ed be o e he sea ch began bu was
no egis e ed. We conside ed human adio he apy s udies o clinically ancho ed phan om and challenge e alua ions
ha applied a i icial‑, machine‑, o deep‑lea ning o: (i) segmen a ion o a ge s o o gans a isk; (ii) ea men
planning and dose p edic ion; (iii) syn he ic‑CT o cone‑beam‑CT image‑quali y enhancemen ; (i ) quali y assu ance
and e o de ec ion; ( ) mo ion modeling o ma ke ‑less acking; ( i) adiomics‑based p ognos ic modeling; and ( ii)
adap i e o MR‑linac wo k lows. Eligible compa a o s included ou ine clinical s anda ds o expe eade s, al hough
single‑a m echnical s udies wi hou a compa a o we e also allowed. P ima y ou comes we e ask‑speci ic (e.g.,
Dice/HD95 and edi ing ime o segmen a ion; mean absolu e e o , DVH del as, and plan‑accep ance o dose
p edic ion; HU o dose‑calcula ion e o s and egis a ion accu acy o image‑quali y; sensi i i y/speci ici y o QA;
la ency and 3D e o o mo ion; and AUC/C‑index, calib a ion, and ex e nal alida ion s a us o adiomics). We
sea ched PubMed, Embase, Scopus, Web o Science, IEEE Xplo e, and a Xi om 1 Janua y 2000 h ough 27 July 2025
using combined AI and adio he apy keywo ds, and scanned ClinicalT ials.go . Two e iewe s independen ly sc eened
i les, abs ac s, and ull ex s and ex ac ed s udy design, umo si e, da ase sizes and spli s, alida ion ype,
quan i a i e me ics, wo k low ime‑sa ings, code/da a a ailabili y, and unding o con lic ‑o ‑in e es s a emen s,
esol ing disag eemen s by consensus. Risk o bias was judged wi h QUADAS‑2 o echnical/diagnos ic asks and
PROBAST o p ognos ic models; epo ing quali y was benchma ked agains TRIPOD and CLAIM i ems. Owing o
he e ogenei y, we used s uc u ed na a i e syn hesis wi h quan i a i e summa ies (Figu e 1).
Figu e 1 PRISMA 2020 low diag am
3. Resul s and Discussion
3.1. Segmen a ion o Ta ge s and O gans a Risk
The la ges ex e nally alida ed s udy—2,208 NSCLC pa ien s ac oss eigh cen e s—achie ed a median olume ic Dice
o 0.91 and su ace Dice 0.86 [6]. Au oma ed hea con ou s e ospec i ely e‑analyzed RTOG 0617 and al e ed mean
hea dose as well as dose‑su i al co ela ions [7]. Sel -con igu ing pipelines, such as nnU-Ne , p o ide compe i i e
pe o mance wi h minimal enginee ing on he HECKTOR head-and-neck challenge es se , nnU-Ne achie ed a Dice
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 1330-1337
1332
sco e o 0.747 (Isensee e al.,2021 [8]; Sa jani e al., 2021 [9]). Compa a i e clinical e alua ions show educed in e -
obse e a iabili y and subs an ial ime sa ings, al hough expe edi s emain necessa y o small o pos ope a i e
s uc u es (Wong e al., 2020 [10]; an Dijk e al., 2020 [11]; Cos ea e al., 2022 [12]). Table 1
A ecen MRI-guided adap i e‑ he apy s udy in p os a e cance showed ha AI con ou ing educed online adap a ion
ime o unde six minu es (Nachba e al., 2023 [13]).
Figu e 2 illus a es ep esen a i e segmen a ion accu acy epo ed by ecen ex e nally alida ed s udies
Figu e 2 Rep esen a i e segmen a ion pe o mance
Table 1 Ex e nal‑ alida ion pe o mance o au o‑segmen a ion models.
S udy
Si e/Task
N
Valida ion
Pe o mance
(Dice/HD95)
Clinical impac / Time
Hosny
2020
NSCLC
2,208
Mul i-ins i u ion
ex e nal
VD 0.91 “0.83–0.92” ;
SD
0.86 0.71–0.91]
Func ional alida ion;
end‑use es ing
Tho 2021
Hea
(RTOG0617)
442
T ial Coho
MHD 15 s 12 Gy
(p=5.8×10^-16)
DL dose s onge OS p edic o
(median p
2.8×10^-5 s 2.0×10^-4)
Isensee
2021
Sa jani
2021
H&N
(HECKTOR)
201
Challenge es
Dice 0.747
Benchma k ; minimal
enginee ing
Wong
2020
OAR,
mul i‑si e
NR
Clinical
High Dice/low HD95
Time educed; a iabili y
dec eased
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 1330-1337
1333
3.2. T ea men Planning and Dose P edic ion (Table 2)
Table 2 Pe o mance o dose‑p edic ion s udies.
S udy
Si e
Design
Key me ics
No es
Nguyen 2019
P os a e
Dose p edic ion
MAE ≲2–3 Gy; DVH del as
small
Au oplanning
easible
Nguyen 2019
Head & Neck
HD U‑Ne
MAE ≲3 Gy
NR
Kajikawa
2019
P os a e
IMRT
CNN dose
MAE epo ed
NR
Fan 2019
Va ious
Au oplanning om 3D
dose
Clinical accep abili y
NR
Zhou 2020
Rec al IMRT
3D dose
MAE; DVH
NR
Voxel-wise U-Ne models p edic p os a e dose wi h a mean absolu e e o (MAE) o 2 Gy [14] and head-and neck dose
wi h MAE o app oxima ely 3 Gy [15]. O he CNN a chi ec u es yield simila pe o mance in p os a e IMRT [16] and
enable au oma ic VMAT planning a e dose p edic ion [17]. Lung IMRT s udies demons a e gene alisabili y ac oss
beam a angemen s [18] and bowel-spa ing ec al plans [19].
A helical omo he apy ne wo k epo ed MAE o app oxima ely 2 Gy in i s -in-human es ing [20], while he ully
con olu ional DoseNe a chi ec u e achie ed planne -le el quali y on 120 pel ic cases [21]. No el loss unc ions
con inue o imp o e ho spo con ol and DVH ideli y [22] (Figu e 3).
Figu e 3 Mean absolu e e o (Gy) epo ed by dose-p edic ion s udies
3.3. Image Quali y: Syn he ic CT (sCT) and Cone‑Beam CT (CBCT) Enhancemen
Low- ield MR-linac p og ammes ely on syn he ic CT wi h HU e o s < 40 HU o pel ic dose calcula ion [23]. Low-dose
cone-beam CT (CBCT) can be es o ed wi h ad e sa ial sca e co ec ion [24] o CycleGAN enhancemen , hal ing dose-
calcula ion e o [25]. Physics-in o med sca e -ke nel supe posi ion emains a benchma k classical app oach [26] and
is now being e-implemen ed wi h GPU accele a ion [27]. Quali y Assu ance and E o De ec ion P edic i e QA models
ained on deli e y logs o gamma maps can lag ailing plans and de ec deli e y e o s wi h high sensi i i y and
speci ici y, enabling p io i ized human e iew. P ospec i e deploymen equi es independen da ase s, p e‑speci ied
h esholds, and con inuous pe o mance moni o ing.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 1330-1337
1334
3.3.1. Quali y Assu ance and E o De ec ion
P edic i e QA models ained on deli e y o gamma maps can lag ailing plans and de ec deli e y e o s wi h high
sensi i i y/speci ici y, enabling p io i ized human e iew [27,28]. P ospec i e deploymen equi es independen
da ase s, p e‑speci ied h esholds and con inuous pe o mance moni o ing.
3.4. Mo ion Modelling and Ma ke ‑less T acking
Respi a o y acking e o s in obo ic adiosu ge y all below 1.5 mm when ke nel-densi y p edic ion is combined wi h
s e eo x- ay imaging [29]. A ecen o hogonal-kV s udy epo ed 1 mm accu acy and < 100 ms la ency in phan om and
olun ee es ing [30].
3.5. Radiomics and P ognos ic/Toxici y Modelling (Table 3)
Table 3 Pe o mance and calib a ion o adiomics-based p ognos ic models
S udy
Si e
Endpoin
Design/Valida ion
Pe o mance
Calib a ion
Zhang
2020
LA‑NSCLC
(PET/CT)
2‑yea PFS
T ain/Tes (41/41)
C‑index 0.77–0.79; isk
g oups 61.9% s 33.2%
( ain) and 43.8% s
22.6% ( es )
Repo ed
Chen
2022
LA‑NSCLC
(CT
+ TOE)
OS
298 (2:1 spli )
AUC 0.965 ( ain); 0.869
(in e nal alida ion)
No ex e nal
Coho
Pa ma
2015
H&N
P ognos ic
In e nal
AUCs epo ed
NR
Tang
2021
HNSCC
P ognosis/ ecu ence
In e nal/boo s apped
AUCs epo ed
NR
A PET/CT signa u e s a i ied locally ad anced NSCLC in o high- and low- isk g oups wi h C-index 0.77-0.79 on an
independen coho [31]. A 298-pa ien CT s udy eached AUC 0.869 in in e nal alida ion bu lacked ex e nal es ing
[32]. Ea lie adiomic classi ie s in head-and-neck cance p o ed di icul o ep oduce [33, 34], echoing b oade
conce ns abou o e i ing in oncology ML su eys [35, 36]. (Figu e 4).
Figu e 4 Radiomics p ognos ic pe o mance exp essed as AUC o Cindex.
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 1330-1337
1335
4. Discussion
Ac oss he RT wo k low, he mos ma u e AI applica ions a e supe ised au o‑segmen a ion and
dose‑p edic ion‑enabled au oplanning. These ools can educe con ou ing a iabili y and planning ime while
main aining plan quali y. Image enhancemen echniques (sCT and CBCT co ec ion) suppo adap i e wo k lows,
whe eas p edic i e QA and ma ke ‑less acking a e poised o p agma ic e alua ion. Radiomics emains p omising bu
equi es consis en ex e nal alida ion, calib a ion, and decision‑cu e analysis be o e ou ine adop ion. Fu u e wo k
should p io i ize mul i‑cen e s udies, p e‑ egis e ed analysis plans, and obus pos ‑deploymen moni o ing o
mi iga e d i and bias.
5. Conclusion
A i icial in elligence is al eady imp o ing e iciency and consis ency ac oss adio he apy—mos no ably in au o-
segmen a ion and dose-p edic ion-enabled planning. Image-quali y me hods (sCT and CBCT co ec ion) acili a e
adap i e wo k lows, while p edic i e QA and ma ke -less acking a e p omising bu equi e p ospec i e e alua ion.
B oad clinical adop ion should p io i ize mul i-cen e ex e nal alida ion, calib a ion and d i moni o ing, and
anspa en epo ing. Embedding AI ools wi hin lea ning-heal h-sys em in as uc u es will suppo con inuous
pe o mance o e sigh and equi able bene i . These s eps will help ansla e echnical gains in o measu able
imp o emen s in plan quali y, ea men imes, and pa ien ou comes.
Compliance wi h e hical s anda ds
Acknowledgmen s
The au ho s hank colleagues in he Depa men o Radia ion Oncology o help ul discussions.
Disclosu e o con lic o in e es
The au ho s decla e no compe ing in e es s
Funding
No speci ic unding was ecei ed o his wo k.
Au ho con ibu ions
All au ho s con ibu ed o he concep ion, li e a u e sea ch, analysis, and manusc ip d a ing. All au ho s app o ed he
inal manusc ip .
Da a a ailabili y
All da a a e con ained wi hin he a icle and i s e e ences.
Re e ences
[1] Bibaul JE, Bu gun A, Gi aud P. A i icial in elligence applied o adio he apy. Cance /Radio hé apie.
2017;21:239-243. doi:10.1016/j.can ad.2016.09.021
[2] Abe ne hy AP, E he edge LM. Rapid-lea ning sys em o cance ca e. J Clin Oncol. 2010;28:4268-4274.
doi:10.1200/JCO.2010.28.5478
[3] Kawamu a M, Kamomae T. Re olu ionizing adia ion he apy: he ole o AI in clinical p ac ice. J Radia Res.
2024;65:1-9. doi:10.1093/j / ad090
[4] LeCun Y, Bengio Y, Hin on G. Deep lea ning. Na u e. 2015;521:436-444. doi:10.1038/na u e14539
[5] Sil e D, Huang A, Maddison CJ, e al. Mas e ing he game o Go wi h deep neu al ne wo ks and ee sea ch. Na u e.
2016;529:484-489. doi:10.1038/na u e16961
[6] Hosny A, Pa ma C. Clinical alida ion o deep lea ning algo i hms o adio he apy a ge ing o NSCLC. Lance
Digi Heal h. 2022;4:586. doi:10.1016/S2589-7500(22)00129-7
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 1330-1337
1336
[7] Tho M, Deasy JO. Using au o-segmen a ion o educe con ou ing inconsis ency in RTOG 0617. In J Radia Oncol
Biol Phys. 2021;111:1304-1312. doi:10.1016/j.ij obp.2020.11.011
[8] Isensee F, Jaege PF, Kohl SAA, e al. nnU-Ne : a sel -con igu ing me hod o deep lea ning-based biomedical
image segmen a ion. Na Me hods. 2021;18:203-211. doi:10.1038/s41592-020-01008-z
[9] Sa jani R. nnU-Ne : Fu he au oma ing biomedical image au osegmen a ion. Radiology: Imaging Cance .
2021;3:200225. doi:10.1148/ ycan.2021209039
[10] Wong J, Fong A. Compa ing deep lea ning au o-segmen a ion o in e obse e a iabili y. Radio he Oncol.
2020;146:63-70. doi:10.1016/j. adonc.2019.10.019
[11] an Dijk LV, Van den Bosch L. Imp o ing head-and-neck OAR delinea ion by deep lea ning. Radio he Oncol.
2020;153:34-41. doi:10.1016/j. adonc.2019.09.022
[12] Cos ea M. A las-based e sus deep lea ning OAR delinea ion wi h au oma ed planning. Radio he Oncol.
2022;170:192-199. doi:10.1016/j. adonc.2022.10.029
[13] Nachba M. Au oma ic AI-based con ou ing o p os a e MRI o online ART. Z Med Phys. 2023;33:32-43.
doi:10.1016/j.zemedi.2023.05.001
[14] Nguyen D, Long T. Feasibili y o op imal dose p edic ion o p os a e om ana omy. Sci Rep. 2019;9:12282.
doi:10.1038/s41598-019-49074-4
[15] Nguyen D, Jia X, She D, e al. 3D dose p edic ion o head and neck wi h hie a chically densely connec ed U-Ne .
Phys Med Biol. 2019;64:135016. doi:10.1088/1361-6560/ab039b
[16] Kajikawa T. Con olu ional neu al ne wo k o dose p edic ion in p os a e IMRT. J Radia Res. 2019;60:685-693.
doi:10.1093/j / z051
[17] Fan J, Wang J. Au oma ic ea men planning based on p edic ed 3D dose dis ibu ion. Med Phys. 2019;46:370-
381. doi:10.1002/mp.13271
[18] Ba agán-Mon e o AM. 3D dose p edic ion o lung IMRT wi h deep neu al ne wo ks. Med Phys. 2019;46:367-
379. doi:10.1002/mp.13398
[19] Zhou J. Deep lea ning o p edic 3D dose dis ibu ion o ec al IMRT. J Appl Clin Med Phys. 2020;21:200-209.
doi:10.1002/mp.13963
[20] Liu Z. Deep lea ning-based 3D dose p edic ion o helical omo he apy. Med Phys. 2019;46:414-423.
doi:10.1002/mp.13275
[21] Kea ney V. DoseNe : a 3D ully con olu ional ne wo k o dose p edic ion. Phys Med Biol. 2018;63:235022.
doi:10.1088/1361-6560/aae50
[22] Ma M. A new loss unc ion o adio he apy dose p edic ion. F on Oncol. 2021;11:676190. doi:10.1088/0031-
9155/55/5/004
[23] Cusumano D. Syn he ic CT in low- ield MR-guided adap i e adio he apy. Radio he Oncol. 2020;152:157-162.
doi:10.1016/j. adonc.2020.10.018
[24] Ku z C. Deep lea ning-based pos -p ocessed low-dose CBCT. F on A i In ell. 2020;3:15.
doi:10.3389/ ai.2020.00015
[25] Kida S. Low-dose CBCT enhancemen using CycleGAN. Med Phys. 2020;47:998-1010. doi:10.7150/ hno.43847
[26] Sun M, S a -Lack JM. Imp o ed sca e co ec ion using adap i e sca e ke nel supe posi ion. Phys Med Biol.
2010;55:6695-6720. doi:10.1088/0031-9155/55/22/007
[27] Spadea MF. Deep lea ning-based syn he ic CT in adio he apy and PET: a e iew. Med Phys. 2021;48:978-992.
doi:10.1002/mp.15150
[28] Ny lo MJ. Deep lea ning o pa ien -speci ic IMRT QA p edic ion. Med Phys. 2019;46:3847-3856.
doi:10.1002/mp.13338
[29] Inoue M. Fac o s a ec ing accu acy in espi a o y acking wi h obo ic adiosu ge y. Jpn J Radiol. 2019;37:620-
627. doi:10.1007/s11604-019-00859-7
[30] Zhou D. Deep lea ning-based ma ke less eal- ime lung umou acking wi h o hogonal kV. J Appl Clin Med
Phys. 2023;24:13838. doi:10.1002/acm2.13838
Wo ld Jou nal o Ad anced Resea ch and Re iews, 2025, 27(02), 1330-1337
1337
[31] Zhang N, Liang R, Gensheime MF, e al. Ea ly PET/CT adiomics esponse o LA-NSCLC. The anos ics.
2020;10:11707-11718. PMID: PMC11450131
[32] Chen N-B. CT adiomics-based long- e m su i al a e CCRT in LA-NSCLC. Radia Oncol. 2022;17:184.
doi:10.1186/s13014-022-02136-w
[33] Pa ma C, G ossmann P, Rie eld D, e al. Radiomic machine-lea ning classi ie s o p ognos ic bioma ke s o
head and neck cance . F on Oncol. 2015;5:272. doi:10.3389/ onc.2015.00272
[34] Tang FH, Chu CYW, Cheung EYW. Radiomics AI p edic ion o HNSCC p ognosis and ecu ence. BJR Open.
2021;3:20200073. doi:10.1259/bj o.20200073
[35] C uz JA, Wisha DS. Applica ions o machine lea ning in cance p edic ion and p ognosis. Cance In o m.
2006;2:59-77.
[36] Kou ou K, Exa chos TP, Exa chos KP, Ka amouzis MV, Fo iadis DI. Machine lea ning applica ions in cance
p ognosis and p edic ion. Compu S uc Bio echnol J. 2015;13:8-17. doi:10.1016/j.csbj.2014.11.005
