Image models for the detection and characterization of a rare liver disease (Porto-Sinusoidal Vascular Disorder)

id195180


IMAGE MODELS FOR THE DETECTION AND
CHARACTERIZATION OF A RARE LIVER
DISEASE (PORTO-SINUSOIDAL VASCULAR
DISORDER)
MARTÍN SUÁREZ FERNÁNDEZ
Thesis supe iso
JUANCARLOSGARCIAPAGAN(FundacióRece caClínicBa celona-IDIBAPS)
Thesis co-supe iso
DARIOGARCÍAGASULLA(Depa men o Compu e Science)
Tu o :JAVIERBÉJARALONSO(Depa men o Compu e Science)
Deg ee
Mas e 'sDeg eeinA i icialIn elligence
Mas e 's hesis
School o Enginee ing
Uni e si a Ro i a i Vi gili (URV)
Facul y o Ma hema ics
Uni e si a de Ba celona (UB)
Ba celona School o In o ma ics (FIB)
Uni e si a Poli ècnica de Ca alunya (UPC) - Ba celonaTech

Abs ac
Po o-Sinusoidal Vascula Diso de (PSVD) is a a e and o en unde diagnosed li e condi ion wi h
signi ican clinical challenges, equen ly equi ing in asi e p ocedu es like biopsies o diagnosis. This
hesis in es iga es he use o ad anced a i icial in elligence (AI) models o imp o e he de ec ion
and cha ac e iza ion o PSVD using medical imaging da a. The esea ch add esses key challenges,
including noisy da ase s, da a imbalance ( h ough me ada a), and he sub le isual ma ke s o PSVD
ha o e lap wi h condi ions such as ci hosis.
Using s a e-o - he-a deep lea ning echniques, including 2D and 3D con olu ional neu al ne wo ks,
his s udy e alua es di e en a chi ec u es, ans e lea ning s a egies, and p ep ocessing me hods.
A de ailed pipeline o noise educ ion and image segmen a ion was de eloped o help models ocus
on ele an ana omical ea u es while educing in e e ence om i ele an a eas.
The esul s show ha AI has he po en ial o imp o e PSVD de ec ion, wi h ad ancemen s in model
obus ness and in e p e abili y. G ad-CAM was used o c ea e explainable isualiza ions, o e ing
insigh s in o model decision-making and suppo ing clinical alida ion. Despi e hese ad ancemen s,
he esea ch emphasizes he need o la ge and mo e di e se da ase s, along wi h u he e inemen
o AI me hods, o achie e highe diagnos ic accu acy.
This wo k con ibu es o he g owing ield o AI-d i en medical imaging by p o iding a ounda ion
o u u e inno a ions in diagnosing a e diseases. I also highligh s he impo ance o sus ainabili y
and e hical conside a ions in heal hca e echnology.
i
Acknowledgmen s
I would like o exp ess my since e g a i ude o my supe iso s Da ío and Ja ie o his in aluable
guidance and suppo h oughou his s udy. Special hanks a e ex ended o Juan Ca los and
o he physicians om Hospi al Clinic and IDIBAPS o p o iding expe insigh s and acili a ing
access o he medical imaging da a essen ial o his esea ch. Addi ionally, I deeply app ecia e
he con ibu ions o all HPAI eam membe s who assis ed in he de elopmen o he me hodologies
desc ibed in his wo k, especially Jaume, who, on a weekly basis helped me discuss aspec s o he
p ojec . Thei dedica ion and e o we e essen ial in he success ul comple ion o his hesis. And
las bu no leas , hank all my amily and close iends o hei suppo du ing he de elopmen o
he p ojec , especially du ing he inal s ages which supposed he bigges amoun o wo k.
iii

Table o Con en s
1 In oduc ion 1
1.1 Mo i a ion ......................................... 1
1.2 Resea chQues ions..................................... 2
2 Rela ed wo k 3
2.1 Clinicalbackg ound .................................... 3
2.2 Di icul ies in Di e en ia ing PSVD om Ci hosis . . . . . . . . . . . . . . . . . . . 4
2.3 AI T ans o ming Medical Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.4 AI o PSVDdiagnosis................................... 5
3 Da ase 7
3.1 Secu i yp o ocols ..................................... 7
3.2 Da ase sanalysis...................................... 8
3.3 Da ap ep ocessing..................................... 16
4 Me hodology 27
4.1 Inpu Con igu a ions.................................... 27
4.2 A chi ec u es ........................................ 28
4.3 T ans e Lea ning...................................... 31
4.4 Expe imen alSe up .................................... 33
4.5 E alua ion.......................................... 33
4.6 Expec edOu comes .................................... 36
5 Expe imen s 37
5.1 Fullimage.......................................... 38
5.2 UsingBoundingBoxes................................... 40
5.3 UsingSegmen a ionMasks ................................ 42
5.4 Summa y .......................................... 44
6 P oblem simpli ica ion 47
6.1 Heal hy s PSVD s Ci hosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.2 Heal hy sPSVD...................................... 49
6.3 PSVD sCi hosis..................................... 51
6.4 Summa y .......................................... 52
7 Sus ainabili y and E hical Implica ions 53
8 Conclusions 57
8.1 Resea ch Ques ions Add essed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
8.2 KeyCon ibu ions ..................................... 57
8.3 Limi a ions and Fu u e Wo k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
8.4 FinalRema ks ....................................... 58
i
Lis o Figu es
3.1 Da a sa e y ea men measu es . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Di e en ypes CT scan iews [14] ............................ 9
3.3 Class dis ibu ion ac oss da ase s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4 O e all da se s sex dis ibu ions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.5 Condi ion p e alence by sex. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.6 Age dis ibu ions o pa ien s ac oss da ase s. . . . . . . . . . . . . . . . . . . . . . . 12
3.7 Age dis ibu ion by condi ion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.8 Dis ibu ion o slice hickness among da ase s olumes. . . . . . . . . . . . . . . . . . 13
3.9 Dis ibu ion o numbe o slices pe olume . . . . . . . . . . . . . . . . . . . . . . . 13
3.10 Dis ibu ion o slice coun s pe olume by condi ion. . . . . . . . . . . . . . . . . . . 14
3.11 Dis ibu ion o da ase s samples’ s udy yea s. . . . . . . . . . . . . . . . . . . . . . . 14
3.12 S udy yea dis ibu ion by condi ion. . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.13 Dis ibu ion o scanne models used o imaging. . . . . . . . . . . . . . . . . . . . . 15
3.14 Phases o de eloped au oma ic c opping me hod . . . . . . . . . . . . . . . . . . . . 18
3.15
Closeness o ep esen a i e slices in absolu e and ela i e di e ence o indices be ween
he bes slice and subsequen bes slices . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.16 SAM example segmen a ions o li e (o ange) and spleen (g een) . . . . . . . . . . . 19
3.17 GennUNe example segmen a ions o li e (o ange) and spleen (g een) . . . . . . . . 20
3.18 MONAI example segmen a ions o spleen (g een) . . . . . . . . . . . . . . . . . . . . 20
3.19 In e ace o Slice 3D p og am o segmen a ion mask c ea ion . . . . . . . . . . . . . 21
3.20 Segmen a ion mask c ea ion. D awing (Le ) and co esponding mask (Righ ) . . . . 21
3.21 Di e ence be ween aw and clipped image . . . . . . . . . . . . . . . . . . . . . . . . 22
3.22 Example o slice a e da a augmen a ion pipeline . . . . . . . . . . . . . . . . . . . . 25
4.1 Inpu ype 1 (Full) image example slice. . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2 Inpu ype 2 (C opped) image example slice. . . . . . . . . . . . . . . . . . . . . . . 28
4.3 Inpu ype 3 (Masked) image example slice. . . . . . . . . . . . . . . . . . . . . . . . 28
4.4 Residual connec ion [30].................................. 29
4.5 Mul iclass con usion ma ix [41]. ............................. 34
4.6 Example saliency maps gene a ed using G ad-CAM. . . . . . . . . . . . . . . . . . . 35
5.1 Con usion ma ices om bes models ained on Full images . . . . . . . . . . . . . . 38
5.2 Saliency maps ob ained using G ad-CAM om bes 2D models ained on ull images 39
5.3 Saliency maps om 3D model Medical Ne ained on Full images . . . . . . . . . . 40
5.4 Con usion ma ices om bes models ained on C opped images . . . . . . . . . . . 41
5.5 Saliency maps om bes 2D models ained on c opped images . . . . . . . . . . . . 42
5.6 Con usion ma ices om bes models ained on Masked images . . . . . . . . . . . . 43
ii
Chap e 2. Rela ed wo k
Ea ly de ec ion o PSVD is c ucial o p e en disease p og ession and educe he likelihood o se e e
complica ions.
PSVD’s a iabili y and o e lap wi h o he li e condi ions make i di icul o diagnose and ea , bu
hey also highligh oppo uni ies o inno a ion. F om a compu e science pe spec i e, he challenges
posed by PSVD, such as in e p e ing medical imaging, iden i ying sub le pa e ns, and p edic ing
disease p og ession, a e well-sui ed o ad anced compu a ional echniques. Machine lea ning and
image analysis, in pa icula , ha e he po en ial o enhance diagnos ic accu acy, educe eliance on
in asi e es s, and p o ide new insigh s in o he condi ion.
In conclusion, PSVD is a a e bu impo an li e diso de ha p esen s unique challenges in bo h
medicine and echnology. I s dis inc ea u es and unp edic able clinical cou se equi e ca e ul
e alua ion and new app oaches o diagnosis and ea men . By in eg a ing compu a ional models
in o he s udy and managemen o PSVD, he e is hope o mo e accu a e diagnoses, be e pa ien
ou comes, and a deepe unde s anding o he disease’s p og ession.
2.2 Di icul ies in Di e en ia ing PSVD om Ci hosis
Po o-sinusoidal ascula disease (PSVD) and ci hosis sha e many clinical simila i ies, making i
di icul o diagnose and manage hese condi ions co ec ly. Bo h can p esen wi h complica ions
ela ed o po al hype ension (PH), such as an enla ged spleen, a iceal bleeding, luid buildup in
he abdomen (asci es), and po al ein blockages (PVT) [4,5]. Howe e , PSVD is undamen ally
di e en om ci hosis. While ci hosis in ol es ad anced sca ing and s uc u al damage o he
li e , PSVD is de ined by speci ic ascula changes wi hou signi ican ib osis o dis o ion o li e
s uc u e [6]. Di e en ia ing be ween he wo is c ucial since hei unde lying causes and ea men s
a e no he same.
One o he key ools o dis inguishing PSVD om ci hosis is analyzing li e issue unde a
mic oscope. PSVD ypically shows ascula changes such as blocked small eins (obli e a i e po al
enopa hy), pa ial ib osis, and nodula egene a i e changes, bu i does no display he ad anced
sca ing and egene a i e nodules seen in ci hosis [6,7]. Howe e , e en li e biopsies a e no always
de ini i e, especially in he ea ly s ages o PSVD, whe e he signs can be sub le o missed due o
limi ed sampling.
Pa ien s wi h PSVD o en ha e ela i ely well-p ese ed li e unc ion. Key indica o s, such as
albumin, bili ubin, and blood clo ing ac o s, may appea no mal o close o no mal. On he
o he hand, ci hosis equen ly causes signi ican li e dys unc ion due o widesp ead issue damage
[5]. Tha said, ad anced cases o PSVD can esul in complica ions ha esemble decompensa ed
ci hosis, such as se e e luid e en ion o gas oin es inal bleeding, making diagnosis e en mo e
complica ed [8]. Imaging me hods like ul asound o CT scans may also s uggle o di e en ia e he
wo condi ions, as bo h can show simila ea u es, such as spleen enla gemen o abno mali ies in
he po al ein.
The di icul y in dis inguishing PSVD om ci hosis has a majo impac on ea men . T ea men s
o ci hosis o en ocus on educing ib osis and managing ela ed complica ions, while PSVD
equi es a di e en s a egy ha a ge s ascula issues. Fo ins ance, an icoagulan s o ea PVT
in PSVD a e used cau iously due o he isk o bleeding, whe eas hey migh be mo e commonly
p esc ibed in ci hosis [7]. Misdiagnosing PSVD as ci hosis can esul in inapp op ia e ea men s,
and in ex eme cases, unnecessa y li e ansplan s [9].
4

Chap e 2. Rela ed wo k
To ackle hese diagnos ic challenges, esea che s a e de eloping be e ools and me hods. These
include ad anced imaging echniques, non-in asi e bioma ke s, and imp o ed c i e ia o in e p e ing
li e biopsies [9]. Accu a e diagnosis ea ly on is essen ial o ensu e ha pa ien s ecei e he igh
ea men s and o a oid complica ions caused by mismanagemen .
2.3 AI T ans o ming Medical Imaging
A i icial in elligence (AI) has made signi ican s ides in ad ancing medical imaging and imp o ing
diagnos ic accu acy and e iciency ac oss a ious medical ields. In hepa ology, AI has been pa icu-
la ly impac ul in enhancing he de ec ion and managemen o li e diseases, such as ci hosis and
i s complica ions.
Machine lea ning (ML) models ha e shown supe io p edic i e capabili ies o adi ional diagnos ic
ools, pa icula ly in assessing li e ib osis. Fo example, Chang e al. e alua ed a ious ML models,
including logis ic eg ession, andom o es s, and a i icial neu al ne wo ks, o p edic ing ib osis
s ages in pa ien s wi h me abolic-associa ed a y li e disease (MASLD). These models consis en ly
ou pe o med con en ional me hods in accu acy and eliabili y [10].
Deep lea ning (DL) echniques, pa icula ly con olu ional neu al ne wo ks (CNNs), ha e u he
expanded he capabili ies o medical imaging. These models ha e been success ully applied o
imaging modali ies such as CT and MRI o de ec li e condi ions. Fo ins ance, Mazumde e al.
de eloped an au oma ed li e segmen a ion model combining 3D-U-Ne and DeepLab 3+ algo i hms,
accu a ely iden i ying ci hosis om CT scans [11].
Beyond li e diseases, AI has demons a ed i s e sa ili y in diagnosing a a ie y o condi ions.
Inno a i e AI models ha e been designed o analyze non- adi ional diagnos ic ma ke s, such as
ongue colo , o de ec diseases like diabe es and cance wi h high accu acy. Simila ly, in oncology,
AI-assis ed mammog aphy has gained ac ion o imp o ing b eas cance de ec ion. By le e aging
deep lea ning on ex ensi e mammog am da ase s, AI ools ha e achie ed as e and po en ially mo e
accu a e iden i ica ion o cance ous pa e ns compa ed o human adiologis s.
Despi e hese ad ancemen s, challenges emain in in eg a ing AI in o e e yday clinical p ac ice.
Key conce ns include da a p i acy, model in e p e abili y, and he lack o s anda dized alida ion
p o ocols. Ensu ing compliance wi h medical egula ions and sa egua ding pa ien in o ma ion is
c i ical o success ul adop ion [12].
In summa y, AI con inues o e olu ionize medical imaging, o e ing ans o ma i e po en ial o
diagnosing li e diseases and beyond. Ongoing esea ch is essen ial o add ess cu en challenges
and maximize he u ili y o AI-d i en ools in clinical se ings.
2.4 AI o PSVD diagnosis
Howe e , despi e all he ad ances made in he medical ield ega ding AI, he applica ion o his kind
o sys ems o he speci ic case o PSVD has no been de eloped ha much, being one o he mos
ad anced cases he p e ious wo k done wi h educed e sions o he da a used in his s udy [13].
The men ioned wo k ca e ully analyzed he da a and success ully c ea ed a deep lea ning model
able o p edic wi h high accu acy he diagnosis o he disease by ollowing a classi ica ion app oach
be ween 4 classes: Ci hosis, PSVD, Heal hy, and Miscellaneous (G oup o di e se li e diseases
o he han he ones al eady s a ed). Howe e , his model, when p esen ed wi h a new da ase
5
Chap e 2. Rela ed wo k
was no able o pe o m as good as expec ed, p o iding almos no p edic ions o he PSVD class.
Addi ionally, he explainabili y me hods used on he men ioned model showed ha i was hugely
biased by high- alue zones such as ibs o he backbone o he pa ien , leading he physicians o no
us he model.
The e o e, as s a ed be o e in his documen , one o he goals o his s udy is o de elop a model
ha consis en ly pe o ms good and eliable p edic ions.
6
Chap e 3
Da ase
This chap e cap u es he de ini ion and analysis o he da a used o he p ojec , de ailing aspec s
o bo h da ase s used as well as di e en ypes o p ocessing applied o such da ase s in o de o
ob ain a cu a ed e sion wi h educed noise.
3.1 Secu i y p o ocols
Fi s and o emos , due o he necessa y p i acy measu es ha ha e o be aken o he pe sonal
de ails o he pa ien s no o be publicly e ealed, a p o ocol had o be ollowed in o de o be able
o wo k wi h he da a.
As explained in he p e ious wo k done by Rubén [13], he da a was sa ely anspo ed om Hospi al
Clinic o he Ba celona Supe compu ing Cen e using an enc yp ed disk and we e sa ely s o ed.
Howe e , access o he da a had o be made using an HVAC clien o a se e in BSC, no allowing
access o he da a om Ma eNos um, only allowing he aining in machines wi h ewe compu ing
capabili ies.
In o de o cope wi h his p oblem, a small esea ch was pe o med o ind a secu e way o s o ing
he da a in he high-pe o mance clus e o la e access. Such esea ch led o he ollowing esul s:
•
Enc yp ion: Da a is enc yp ed in he clus e using Ad anced Enc yp ion S anda d (AES) in
Ciphe Block Chaining (CBC) mode.
•
Enc yp ion key: The key used o access enc yp ed da a is sa ed in a di e en con aine only
accessible by au ho ized use s. I is passed as a a iable in memo y each ime a job is un so
ha he e is ne e a pe sis en copy o such key.
•
Re-enc yp ion: In o de o inc ease he secu i y o he p oposed me hod, enc yp ed da a
may be e-enc yp ed wi h a newly gene a ed key a any desi ed momen .
Figu e 3.1 p o ides a simple ep esen a ion o he da a ea men necessa y o access da a o model
aining.
7
Chap e 3. Da ase
Figu e 3.1: Da a sa e y ea men measu es
3.2 Da ase s analysis
In he s udy conduc ed, wo dis inc da ase s we e p o ided by Ins i u o de In es igaciones Biomédicas
Augus Pi i Sunye (IDIBAPS), comp ised o 201 (T aining da ase ) and 208 (Tes da ase ) olumes
each. The T aining da ase is he one used in [13] and he Tes da ase is he one used o measu e
he pe o mance o he di e en models de eloped. In [13] he model de eloped, which pe o med
eally well on he alida ion pa i ion o he T aining da ase was no able o co ec ly p edic a
single ins ance o he PSVD se o pa ien s, which lead us o pe o m a ca e ul s udy o bo h da ase s
and analyze hei di e ences.
The ollowing explo a o y da a analysis (EDA) ocuses on iden i ying pa e ns and cha ac e is ics
wi hin he da ase s ha a e ele an o diagnosing Po o Sinusoidal Vascula Disease (PSVD) and
also o e eal any possible key di e ences be ween hem. Key insigh s ela ed o gende , age, imaging
slice dis ibu ion, and scanning p o ocols a e ou lined below. These indings se e as a ounda ion
o p ep ocessing, and model de elopmen aimed a PSVD diagnosis.
3.2.1 De ini ion
Bo h da ase s a e comp ised o CT (Compu ed Tomog aphy) scans which use X-Ray adia ion and
in o ma ion abou he a enua ion o such adia ion o c ea e wha we see as images. The main
p ope ies o he iles comp ising bo h da ase s a e desc ibed below.
•
Axial iew: Each and e e y one o he CT scans o he T aining da ase p o ides images in
axial iew ( iew om he bo om o he pa ien . See igu e 3.2 o unde s and he di e en
ypes o iews a ailable o CT scans) whe eas some o he scans om he Tes ing se (4
olumes) a e p esen ed in co onal iew and, gi en he di e ence in o ma , which made hem
unusable o models ained on axial iews, we e he e o e disca ded.
•
DICOM o ma : The ile o ma used o he scans is called DICOM (Digi al Imaging
and Communica ions in Medicine) which is he s anda d o ma o his ype o scans in he
8
Chap e 3. Da ase
medical ield. I p o ides no only he images ha will be used o aining bu also some
addi ional me ada a ha will be explained in de ail la e in his documen . No e ha he e is
no di e ence in he o ma ile be ween bo h da ase s.
•
Region o in e es : All o he olumes ha comp ise bo h da ase s should heo e ically be
composed o images ha con ain he li e o a leas he abdomen. The majo i y o CT scans
o he T aining da ase ill ou such equisi e, wi h one sample p o iding images om he ches
down o he pel is. Howe e , he DICOM iles p esen ed in he Tes da ase p esen slices
belonging o o he a eas apa om he abdomen, including he ches , wais , o e en legs and
head.
Figu e 3.2: Di e en ypes CT scan iews [14]
3.2.2 Class dis ibu ion
As p e iously men ioned, he whole da ase is di ided in o ou subg oups o classes:
1. Heal hy: G oup o CT scans made on people in a heal hy s a e ega ding he li e .
2. PSVD: Se o pa ien s wi h a con i med diagnosis o PSVD.
3.
Ci hosis: This g oup consis s o he CT scans made o hose pa ien s wi h li e ci hosis.
Such pa i ion is included wi h he objec i e o making he model mo e obus owa ds he
di e en ia ion be ween PSVD and Ci hosis, which is one he main di icul ies o expe s in
he ield.
4.
Miscellaneous: Finally, a g oup wi h o he kinds o li e diseases is included wi h he pu pose
o no limi ing he possibili ies o he diagnosis, especially since a pa ien migh no p esen
PSVD o Ci hosis bu s ill no be heal hy.
Bo h da ase s a e ela i ely balanced, wi h he T aining da ase p esen ing almos a pe ec dis i-
bu ion o pa ien s ac oss all classes and he Tes da ase showing a sligh imbalance owa ds he
Ci hosis (65 samples) and PSVD (42 olumes) g oups. I is wo h no ing ha he mos impo an
class (PSVD) is he one wi h he lowes amoun o samples, meaning he is a sligh unde ep esen a-
ion o he disease. Figu e 3.3 p o ides a g aphical ep esen a ion o he class dis ibu ion o bo h
g oups.
9

Chap e 3. Da ase
(a) T ain da ase class dis ibu ion (b) Tes da ase class dis ibu ion
Figu e 3.3: Class dis ibu ion ac oss da ase s
3.2.3 Sex
One o he mos ele an aspec s o many diseases owa ds he p obabili y o he pa ien ha ing i
o no is he sex o such pa ien . Biological di e ences migh make males o emales mo e p one
o ha ing a speci ic disease, and hus, a de ailed analysis o he sex dis ibu ion ac oss da ase s is
necessa y o be e unde s and he diso de .
The da ase s show a highe ep esen a ion o male pa icipan s (58.2% in he T aining se and 62%
in he Tes se ) 3.4, wi h o e 80% o PSVD cases occu ing in males in he T aining se and a bi
less (73%) in he Tes se . This aligns wi h known dispa i ies in disease p e alence ac oss sexes,
in luenced by ac o s such as ho monal di e ences and body composi ion. Recognizing his imbalance
is c ucial o de eloping diagnos ic models ha a e ai and inclusi e, ensu ing accu a e p edic ions
o all pa ien s (Figu e 3.5). A i s sigh , i appea s ha PSVD and Ci hosis a e signi ican ly mo e
equen in males compa ed o emales, whe eas emales domina e he Miscellaneous class. This
migh be one o he ac o s making i di icul o di e en ia e be ween pa ien s p esen ing PSVD o
Ci hosis.
The o e all gende dis ibu ion is ela i ely balanced ac oss o he condi ions on bo h da ase s wi h
he Tes se being he mos i egula ly dis ibu ed se . Howe e , he e is enough balance o ensu e
model gene aliza ion and educe he likelihood o gende bias du ing aining. Addi ionally, bo h
da ase s p esen he same ype o imbalance in all aspec s excep o he Heal hy class which is
balanced in bo h cases. This leads us o belie e ha he e will be no conside able bias ega ding he
sex o he pa ien s.
10
Chap e 3. Da ase
(a) T aining da ase (b) Tes da ase
Figu e 3.4: O e all da se s sex dis ibu ions.
(a) T aining da ase (b) Tes da ase
Figu e 3.5: Condi ion p e alence by sex.
3.2.4 Age
Age is one o he mos impo an ac o s ega ding almos any disease, since he olde he human body
he weake i becomes. Specially in his case, whe e he main symp om is he po al hype ension,
age becomes a eally ele an aspec o ake in o accoun . Addi ionally, ega ding he isual aspec
o he p oblem, which is he one conce ning his s udy, i is impo an o ake in o accoun ha
he li e migh expe ience changes in i s mo phology as he body ages [15], which migh lead o
conside able di e ences ega ding he isual examina ion o he CT scan.
I can be seen in Figu e 3.6 how o he T aining se he age is equally dis ibu ed along a wide ange
o ages concen a ed a ound 30 and 40 yea s o age, consis en wi h he ypical demog aphic o
PSVD diagnosis. Howe e , he age dis ibu ion no ably a ies o he Tes da ase which p esen s a
concen a ion o pa ien s a ound he ages o 50 and 60. Such di e ences a e o be aken in o accoun
when analyzing he esul s since biases migh be p oduced by he age ac o due o he model
lea ning on a younge se o pa ien s and mos likely wi h sligh di e ences in he base mo phology
o he li e which may make i s uggle o e ec i ely gene alize o olde popula ions leading hen o
educed p edic i e accu acy
11
Chap e 3. Da ase
(a) T aining da ase (b) Tes da ase
Figu e 3.6: Age dis ibu ions o pa ien s ac oss da ase s.
When age is s a i ied by condi ion (Figu e 3.7), se e al di e ences can be obse ed be ween he
T aining and he Tes da ase s. On he T aining se , he class wi h he younges pa ien s is Heal hy
and he one wi h he oldes ones is Miscellaneous, con a y o he Tes da ase . Addi ionally,
Ci hosis and PSVD show e y simila anges o age in he Tes se , whe eas o he T aining se
he e is a clea di e ence. Again, as men ioned p e iously, di e ences in he demog aphics o he
da ase s’ pa ien s migh p oduce undesi ed esul s, especially gi en he low amoun o da a a ailable
which may no be enough o lea n he na u al di e ences p oduced in he li e as i ages.
(a) T aining da ase (b) Tes da ase
Figu e 3.7: Age dis ibu ion by condi ion.
3.2.5 Slices pe Volume
The o e all dis ibu ion o slice hickness (Figu e 3.8) shows o bo h da ase s wo main g oups o
olumes wi h 4 and 5mm slice hickness wi h an inc eased size o he 4mm g oup o he Tes se .
Such simila i ies sugges ha he slice hickness will mos likely no skew he model’s esul s on
he Tes se . Howe e , as o he slice coun s (Figu e 3.9), i can be seen how he Tes se p esen s
a much highe coun o slices pe olume when compa ed o he T aining se . This is due o he
p e iously men ioned ac ha mos i no all olumes in he Tes se con ain slices no jus o he
li e a ea bu a he o he whole body om he ches down o he beginning o he emu and in
some ex eme cases, slices whe e he b ain is shown can be seen.
12
Chap e 3. Da ase
(a) T aining da ase (b) Tes da ase
Figu e 3.8: Dis ibu ion o slice hickness among da ase s olumes.
(a) T aining da ase (b) Tes da ase
Figu e 3.9: Dis ibu ion o numbe o slices pe olume
Fo bo h da ase s, he dis ibu ion o imaging slices by condi ion (Figu e 3.10) highligh s a sligh
imbalance ega ding he Miscellaneous class, which p esen s a highe slice coun . Howe e , he
da ase p esen s an o e all balance ega ding he numbe o slices pe pa ien , meaning in heo y
ha he e is no po en ial bias ega ding his cha ac e is ic o he da ase . I mus be no ed ha
e en hough he dis ibu ion is balanced, he alues o he Tes da ase a e much highe han o
he T aining se , caused by he ac ha es olumes p esen slices ha do no show he li e .
The e o e, e en i a i s sigh i looks like he e will be no bias, he ue amoun o slices pe olume
ha p esen he li e in he Tes se is unknown.
Addi ionally, e en i ew, he e a e ou lie s in slice coun s poin ing o a iabili y in imaging p ac ices,
likely in luenced by di e ences in equipmen o ins i u ional p o ocols. P ep ocessing echniques will
be c i ical o handling hese inconsis encies.
13
Chap e 3. Da ase
Figu e 3.17: GennUNe example segmen a ions o li e (o ange) and spleen (g een)
Figu e 3.18: MONAI example segmen a ions o spleen (g een)
Se e al o he models such as Aladdin5 [18], Blackbean [19] o o he s om compe i ions abou li e
segmen a ion such as SLIVER07 [20] o FLARE22 [21] we e conside ed, bu due o limi a ions
ega ding how he da a is accessed, such op ions could no be es ed in he end.
A e ca e ully explo ing he esul s i was deduced ha , due o he ac ha he da ase con ains
no only heal hy li e s bu also li e s o pa ien s wi h speci ic and a e diseases, his app oach did
no succeed. The publicly a ailable li e segmen a ion models a e ei he ained o wo k on heal hy
li e s o li e s wi h cance ( umo s), making hem unsui able o he used da ase .
Manual segmen a ion
Finally, a e a discussion wi h he physicians om Hospi al Clinic, i was decided ha a manual
app oach o his ask could be sui able. The main eason behind his is ha he a e na u e o he
disease and he needs o he physicians call mo e o p ecision a he han speed, meaning ha e en
i he p ocess akes longe i is be e i he esul is mo e accu a e. The e o e, he whole da ase
was manually segmen ed ollowing he s eps below:
•
Tool: Since he da a is s o ed as DICOM iles, he decision o which ool o use was mainly
based on ha , making Slice 3D he mos sui able as i p o ides an easy- o-use in e ace and
makes he segmen a ion p ocess much easie .
•
Da a loading: Despi e Slice 3D being one o he mos i no he mos app op ia e p og am o
his ask, he ac ha da a canno be s o ed pe sis en ly wi hou being enc yp ed makes i
a i s impossible o use. Howe e , his p og am p o ides a Py hon console ha allows o
un cus om code, so a sc ip was c ea ed o load he da a om he sa e con aine and load i
di ec ly in o he p og am so ha he e is no pe sis en and dec yp ed copy o he da a a any
momen , hus, complying wi h he secu i y measu es.
20

Chap e 3. Da ase
•
Segmen a ion: The p og am allows he segmen a ion o 3D olumes by an in e pola ion
app oach. This is, o a gi en olume, manual segmen a ions can be made e e y N slices and
hen, using in e pola ion, segmen a ions o in e media e slices a e c ea ed. This mechanism
allows o segmen a pa ien ’s li e in a easonable amoun o ime and also wi h high p ecision
since he ool allows he edi ion o he au oma ically c ea ed segmen a ions.
Figu e 3.19: In e ace o Slice 3D p og am o segmen a ion mask c ea ion
(a) D awing (b) Mask
Figu e 3.20: Segmen a ion mask c ea ion. D awing (Le ) and co esponding mask (Righ )
3.3.2 C opping o Masking
As men ioned in sec ion 3.3.1, images a e ei he c opped o masked using he co esponding
segmen a ion, hus, educing o a la ge ex en he noise p esen in each image.
Bene i s o Clipping and Masking: These p ep ocessing echniques o e se e al ad an ages:
•
Enhanced Model T aining: By educing he inpu size and noise, hese echniques allow
he model o lea n mo e e icien ly and ocus on pa e ns ha a e uly ele an o he ask.
•
Imp o ed Gene aliza ion: By elimina ing i ele an egions, he model becomes less likely
o o e i o ex aneous ea u es, imp o ing i s pe o mance on unseen da a.
21
Chap e 3. Da ase
•
Reduced Compu a ional Load: C opped o masked images a e smalle in size, educing
memo y and compu a ional equi emen s du ing aining and in e ence.
3.3.3 Clipping
One o he bes and mos common p ac ices in medical image analysis is a echnique called windowing
also known as in ensi y clipping. This echnique adjus s he ange o pixel in ensi ies o enhance he
isualiza ion o speci ic issues o s uc u es in an image [22,23,24]. CT scans cap u e a wide ange
o in ensi ies co esponding o a ious issue densi ies, om ai o dense bone. This echnique is
especially use ul in ou case because he olumes in he da ase show pixel alue anges o ei he [0,
1024] o [0, 8192], he e o e, by clipping he image o a ce ain ange o alues, we ensu e ha a e
no maliza ion, li e alues co espond be ween di e en samples.
Tissues o en ha e o e lapping in ensi y anges; o ins ance, so issues and luids can ha e simila
in ensi ies. By na owing he window le el (cen e ) and wid h ( ange), speci ic issues, such as lungs,
bones, o so issues, can be emphasized, imp o ing diagnos ic accu acy. Addi ionally, clipping
in ensi ies ou side he selec ed ange minimizes noise om i ele an egions, such as backg ound ai
o e y dense ma e ials. This enhances he signal- o-noise a io in he a ea o in e es .
(a) Sample aw slice (b) Resul o slice clipped o (-100, 300)
Figu e 3.21: Di e ence be ween aw and clipped image
3.3.4 No maliza ion
No maliza ion is a c ucial p ep ocessing s ep o ensu ing consis ency ac oss he da ase and imp o ing
he s abili y and e iciency o model aining. In medical imaging, CT scans ypically ha e pixel
in ensi y alues ha co espond o issue densi ies. These alues can a y widely depending on he
imaging p o ocol, equipmen , and pa ien -speci ic ac o s. No maliza ion helps s anda dize hese
in ensi ies, making he da a mo e sui able o machine lea ning models.
22
Chap e 3. Da ase
Fo his da ase , pixel in ensi y alues we e scaled o a ixed ange be ween 0 and 255. This app oach
ensu ed ha all images had compa able in ensi y dis ibu ions, ega dless o he o iginal acquisi ion
pa ame e s. No malizing o his ange is pa icula ly use ul o models ha equi e s anda dized
inpu anges [25], such as con olu ional neu al ne wo ks (CNNs).
Bene i s o No maliza ion:
•
Imp o ed Con e gence: By s anda dizing he inpu da a, no maliza ion educes he isk o
la ge g adien s du ing aining, leading o as e and mo e s able con e gence o he model.
•
Reduced Sensi i i y o Va ia ions: No maliza ion minimizes he impac o a ia ions in
pixel in ensi y due o di e ences in equipmen o imaging p o ocols, making he model mo e
obus o eal-wo ld a iabili y.
•
Enhanced Compa abili y: Ensu ing consis en in ensi y anges ac oss all images allows o
ai compa isons and ensu es ha he model ocuses on meaning ul pa e ns a he han being
in luenced by in ensi y scale di e ences.
In summa y, no maliza ion is a ounda ional s ep in p ep ocessing ha ensu es he da ase is
consis en and eady o machine lea ning asks. I educes a iabili y, enhances model pe o mance,
and con ibu es o eliable and ep oducible esul s.
3.3.5 Resizing
The da ase included images o a ying sizes and esolu ions, so all images we e esized o consis en
dimensions (512x512) o ma ch he model’s inpu equi emen s. Also, no e ha e en i in o ma ion
migh be los , he di e ences in sizes be ween he di e en olumes o he da ase s a e no big
enough o cause impo an losses.
3.3.6 Da a Augmen a ion
To imp o e model obus ness and educe o e i ing, da a augmen a ion was applied. This s ep
a i icially expanded he da ase by applying a ious ans o ma ions o he images. Se e al ypes o
ans o ma ions we e applied bu only a ew o hem wo ked p ope ly:
Tes ed ans o ma ions
1.
B igh ness Adjus men : Modi ies he in ensi y o he image o simula e di e en ligh ing
condi ions, making i b igh e o da ke .
2.
Con as Adjus men : Al e s he di e ence be ween ligh and da k a eas o enhance o
supp ess speci ic ea u es.
3. Gamma Adjus men : Applies a nonlinea ans o ma ion o adjus he in ensi y o da ke
o ligh e egions, emphasizing sub le de ails.
4. Ro a ion: Ro a es he image up o a ce ain andom deg ee.
5. Flip: Randomly lips e ically o ho izon ally he image.
6.
Scaling: Reescales he da a o a ce ain ange o so ha alues o simila egions in wo
di e en images p esen simila alues. Fo example, li e should commonly p esen alues
a ound 60 bu due o di e ences in he amoun o con as p o ided o he pa ien , hese migh
a y.
23
Chap e 3. Da ase
7.
E asing: Randomly masks ou pa ches o he image o simula e occlusions and imp o e he
model’s obus ness.
8.
Zooming: Changes he scale o he image, simula ing a ia ions in objec size o dis ance by
zooming in o ou .
9. Gaussian noise: Applies a Gaussian il e o he image in o de o blu i .
10.
Elas ic T ans o m: Applies andom, localized dis o ions o he image by de o ming i
elas ically, simula ing na u al de o ma ions and enhancing he model’s abili y o handle
geome ic a iabili y.
Augmen a ions pu poses
•
Ro a ions, Flips, and Scaling: Geome ic ans o ma ions such as o a ions, ho izon al
lips, and scaling we e applied o simula e di e ences in o ien a ion, pe spec i e, and size.
Ro a ions helped he model become in a ian o he angle a which scans we e aken, while
lips accoun ed o ana omical symme y. Scaling adjus ed he size o image ea u es, e lec ing
he a iabili y in pa ien ana omy and scanning p o ocols.
•
B igh ness, Con as , and Gamma Adjus men s: Adjus men s o b igh ness, con as ,
and gamma alues we e used o eplica e a ia ions in image acquisi ion condi ions, such as
changes in scanne se ings o he applica ion o con as agen s. These augmen a ions we e
implemen ed in o de o imp o e he model’s obus ness o in ensi y di e ences, enhancing
i s abili y o gene alize ac oss da ase s wi h di e se ligh ing and con as p ope ies. Such
ans o ma ions a e pa icula ly e ec i e in medical imaging, whe e small pixel in ensi y
di e ences can highligh impo an diagnos ic ea u es [26].
•
Noise and Blu : Gaussian noise and blu we e added o simula e he noise and a i ac s ha
occu du ing CT scans. These augmen a ions we e mean o help he model handle noisy da a,
making i less sensi i e o image impe ec ions. Noise augmen a ion has been shown o educe
o e i ing and imp o e pe o mance on da ase s wi h inhe en a iabili y [27].
•
Elas ic T ans o ma ions: Elas ic ans o ma ions in oduced non-linea de o ma ions o
he images, simula ing dis o ions in so issues o o gans caused by pa ien mo emen o
scanning a iabili y. These ans o ma ions s e ched and comp essed localized egions o
he image, main aining he o e all s uc u e while in oducing ealis ic dis o ions. Elas ic
ans o ma ions a e pa icula ly help ul in medical imaging, as hey enhance he model’s
obus ness o ana omical a iabili y and posi ional shi s. This me hod is o en used in asks
like segmen a ion and classi ica ion o imp o e pe o mance on de o med o i egula da a [28].
Howe e , maybe due o he limi a ions in he a ailable da a, only a subse o hese augmen a ions
wo ked, being hese: Random lips, Random o a ions, Scaling, and he Elas ic ans o m. Figu e
3.22 shows he di e ence be ween a sample slice and he same slice a e da a augmen a ion.
24
Chap e 3. Da ase
(a) P ep ocessed unaugmen ed slice
(b) P ep ocessed slice a e da a augmen a ion is applied
Figu e 3.22: Example o slice a e da a augmen a ion pipeline
3.3.7 Label Encoding and Smoo hing
Labels we e p epa ed o supe ised lea ning by encoding hem in a machine- eadable o ma . Fo
he cu en ask, hese we e ans o med in o a one-ho ec o and addi ionally, due o he ac ha
i is no su e ha e e y slice o a olume p esen s he disease o he pa ien , labels we e smoo hed.
This is, using Gaussian noise, he one-ho ec o ep esen ing he label o a olume had i s alues
edis ibu ed in o de o add some unce ain y (see Algo i hm 1). The use o noisy o uzzy labels
helps he model a oid o e i ing and he e o e gene alize be e [29].
Algo i hm 1 Gaussian Fuzzy Labels
Requi e: spa se_label,σ
1: Ini ialize one-ho ec o : one_ho ←ze os(num_classes)
2: Se one_ho [spa se_label]←1
3: Add Gaussian noise: noise ∼ N (0, σ)
4: Compu e uzzy label: uzzy_label ←one_ho +noise
5: Clip uzzy label o ensu e non-nega i e alues: uzzy_label ←max( uzzy_label, 0)
6: No malize uzzy label o ensu e p obabili ies sum o 1:
uzzy_label ← uzzy_label
P( uzzy_label)
7: e u n uzzy_label
25

Chap e 3. Da ase
3.3.8 Spli ing he Da ase
The T aining da ase was spli in o wo subse s: aining and alida ion. A s anda d spli a io o
80%-20% was used, ensu ing each subse ep esen ed he o e all da a dis ibu ion. Addi ionally,
s a i ied sampling was applied o main ain class balance ac oss spli s. Finally, he Tes da ase was
le un ouched o be used as a blind es pa i ion o p ope ly measu e he gene aliza ion capabili ies
o he ained models.
3.3.9 Summa y
The p ep ocessing pipeline ensu ed ha he da a was clean, well-s uc u ed, and op imized o
expe imen a ion. These s eps played a i al ole in imp o ing he quali y o he inpu s and, ul ima ely,
he pe o mance o he models.
26
Chap e 4
Me hodology
This chap e ou lines he me hodology used o e alua e he e ec i eness o di e en neu al ne wo k
a chi ec u es, inpu con igu a ions, and ans e lea ning s a egies o he p oposed classi ica ion
p oblem. The goal is o iden i y he op imal combina ion o hese ac o s o maximize classi ica ion
accu acy while balancing compu a ional e iciency and main aining a easonable model explainabili y
le el. Each expe imen was designed o sys ema ically analyze he impac o hese a iables on
model pe o mance and gene aliza ion.
4.1 Inpu Con igu a ions
Th ee dis inc inpu con igu a ions we e used o explo e he ole o p ep ocessing in imp o ing model
pe o mance:
Full Image (Figu e 4.1): This con igu a ion u ilized he en i e image slice, p o iding comple e
ana omical con ex in addi ion o some ex e nal noise. This app oach allows he model o conside
all a ailable spa ial in o ma ion, which is pa icula ly use ul when ele an ea u es appea ac oss
he en i e image. Howe e , he amoun o noise included in he inal image is mo e han he ele an
in o ma ion gained, making i a weak app oach.
Figu e 4.1: Inpu ype 1 (Full) image example slice.
27
Chap e 4. Me hodology
C opped Image (Figu e 4.2): In his se up, he images we e c opped using he p e iously
explained me hod o y o educe he noise p esen in each slice as much as possible in an au oma ed
way. C opping educes i ele an backg ound noise, ocuses he model on he mos in o ma i e a eas,
and dec eases compu a ional complexi y.
Figu e 4.2: Inpu ype 2 (C opped) image example slice.
Masked Image (Figu e 4.3): Again, as explained in p e ious sec ions, segmen a ion masks
we e p oduced manually, which we e hen applied o he co esponding slice o e e y image in he
da ase . Addi ionally, slices whe e he co esponding mask does no p o ide any alue, a e disca ded.
This me hod helps us ensu e ha he model does no ge a ec ed by noise in he image and ha
he inal p edic ion is solely based on in o ma ion ga he ed om he li e .
Figu e 4.3: Inpu ype 3 (Masked) image example slice.
4.2 A chi ec u es
Fou ypes o neu al ne wo k a chi ec u es we e es ed, each o e ing di e en le els o complexi y
and in o ma ion p ocessing. All o hese a e based on he ResNe a chi ec u e, ResNe 50 o be
speci ic, which is widely used in he medical ield. I uses esidual connec ions o help he model
28
Chap e 4. Me hodology
lea n be e and a oid p oblems like anishing g adien s. The i s me hod, a 2D ResNe p e ained
on RadImageNe , is he one used in p io esea ch [13], while he o he h ee a chi ec u es in oduce
no el app oaches designed o add ess speci ic challenges in p ocessing 3D medical imaging da a.
Below, each a chi ec u e is desc ibed in de ail.
2D ResNe
ResNe , sho o Residual Ne wo k, was in oduced by [30] in 2015 o sol e he anishing g adien
p oblem ha made i ha d o ain e y deep neu al ne wo ks. ResNe uses esidual connec ions 4.4,
o "sho cu connec ions" which le g adien s pass h ough he ne wo k mo e easily by skipping one
o mo e laye s. This design makes i possible o ain ne wo ks wi h hund eds o e en housands o
laye s, imp o ing pe o mance in image ecogni ion asks.
Figu e 4.4: Residual connec ion [30]
This 2D con olu ional neu al ne wo k p ocesses one image slice a a ime, p o iding one p edic ion
o each slice o a olume. Howe e , since each olume has o be gi en a single p edic ion, he
ou pu s o he ne wo k o each slice ha e o be agg ega ed and in his case, he majo i y o ing
s a egy has been used, he idea is o gi e he olume he p edic ion ha is epea ed mos along he
whole olume. This me hod is an exac eplica o wha was de eloped in [13] and will be used o
wo easons, i s o all, because i was a wo king me hod on he T aining se , and second, o show
can da a p ocessing and noise educ ion s a egies can imp o e a model’s pe o mance, answe ing
RQ1.
Mo eo e , hese ypes o models p o ide highly de ailed saliency maps when G ad-CAM is applied,
which a o s i s usage o he p oposed ask, ha includes conce ns abou he eliabili y o he
models.
CNN + LSTM
Nowadays, mos 3D app oaches o medical imaging p ocessing a e aimed a segmen a ion and hese
a e mos commonly based on he use o U-Ne . Howe e , o he small subse o 3D app oaches
whose ask is classi ica ion, a widesp ead app oach is o combine a Con olu ional Neu al Ne wo k
(CNN) and a Long Sho -Te m Memo y (LSTM) [31] ne wo k, which is a ype o ecu en neu al
ne wo k [32,33,34]. This hyb id model combines he ea u e ex ac ion capabili ies o 2D CNNs o
ex ac ing ea u es om single slices wi h LSTMs o analyzing pa e ns ac oss mul iple slices. This
se up wo ks well when con ex om neighbo ing slices is impo an .
The speci ic implemen a ion o his a chi ec u e consis s o a combina ion o ResNe 50 p e ained
29
Chap e 4. Me hodology
Toge he , hese me hods ensu ed ha he e alua ion was igo ous, alida ing he models no
only in e ms o nume ical pe o mance bu also in e ms o clinical ele ance and eliabili y. This
combina ion o me ics and in e p e abili y ools suppo s he de elopmen o us wo hy AI sys ems
o medical imaging, ensu ing ha hey align wi h expe unde s anding and deli e meaning ul
p edic ions.
Visualiza ion: Pe o mance me ics we e isualized using:
•Line g aphs o ack pe o mance o e aining epochs.
•Hea maps o show how he models pe o med ac oss di e en classes.
4.6 Expec ed Ou comes
The expe imen s aim o:
•
Find he bes combina ion o inpu ype, a chi ec u e, and ans e lea ning s a egy o
classi ica ion.
•
Show he ad an ages o using RadImageNe weigh s compa ed o ImageNe weigh s and
aining om sc a ch.
•
S udy how noise educ ion me hods (Full, C opped, Masked) a ec model pe o mance and
gene aliza ion in o de o answe RQ1 and RQ2.
•
De e mine i he aining se is enough o c ea e a model able o co ec ly diagnose PSVD
obus ly.
•
P o ide use ul hea maps ha help iden i y he laws in each model and ha aid physicians in
he diagnos ic p ocess.
•
S udy he e ec s o educing he ask’s complexi y by emo ing classes, answe ing he e o e
RQ3.
36

Chap e 5
Expe imen s
This chap e de ails he expe imen s conduc ed o e alua e he pe o mance o di e en app oaches
o he ask. Th ee main s a egies we e explo ed: using ull images, using bounding boxes, and
using segmen a ion masks. Fo each app oach, he esul s a e analyzed, and explainable AI (XAI)
echniques a e applied o be e unde s and how he model makes decisions.
Fo each sec ion, bo h quan i a i e and quali a i e esul s a e p o ided o analyze in de ail he impac
o di e en app oaches o he sugges ed p oblem. In he quan i a i e sec ions, ables p o iding
mac o F1 sco es a e aged h ough he i e uns o each model on he h ee pa i ions a e shown.
These p esen se en di e en ows, each co esponding o a di e en app oach ega ding he model
used:
•
Resne Sc a ch: Re e s o he 2D model using he ResNe a chi ec u e wi h andom weigh
ini ializa ion.
•
Resne Imagene : Uses he same a chi ec u e as he p e ious one bu is ini ialized wi h
ImageNe weigh s.
•
Resne Rad: Ano he ResNe model bu ini ialized wi h weigh s o a model p e ained on
RAD ImageNe .
•
CNN + LSTM: Re e s o he model ha combines a CNN o ex ac ea u es and an LSTM
o combine such ea u es in a sequence-like way.
•
CNN + Pool: Simila app oach as CNN + LSTM bu eplaces LSTM wi h an a e age pooling
laye o simpli y he agg ega ion p ocess.
•
Medical Ne 1 GPU: Co esponds o he MedicalNe -based model p esen ed in sec ion 4.2 bu
using only 1 GPU, which limi s he ba ch size used du ing he aining p ocess o 2 samples
pe s ep.
•
Medical Ne 4 GPU: Same model as be o e bu using 4 GPUs du ing he aining p ocess
which allows o aise he ba ch size up o 8, allowing he model o lea n om di e en classes
on each s ep, and he e o e ha ing a mo e egula ized lea ning p ocess.
37
Chap e 5. Expe imen s
5.1 Full image
In his me hod, he en i e image was used as inpu o he model wi hou any noise educ ion me hods
applied like c opping o segmen a ion. The goal was o see how well he model could gene alize
di ec ly om he aw image da a.
5.1.1 Quan i a i e esul s
The esul s in Table 5.1 show ha models p e ained on adiology-speci ic da a (Resne Rad)
pe o med bes ac oss all me ics, achie ing he highes alida ion and es F1 sco es. The Resne
Sc a ch model, ained om sc a ch on his da ase , had a e y high aining F1 sco e (99.88 ±0.3)
bu s uggled wi h alida ion and es pe o mance, sugges ing o e i ing. Sequen ial a chi ec u es
like CNN + LSTM and pooling-based models showed weake pe o mance, possibly because hey
ailed o cap u e he spa ial ea u es essen ial o 3D medical imaging asks due o he huge amoun
o noise p esen in he inpu s. On he o he hand, he 3D con olu ion-based model Medical Ne
managed o achie e F1 sco es simila o he 2D app oaches showing ha 3D app oaches a e easible.
Model A g T ain F1 A g Valida ion F1 A g Tes F1
Resne Sc a ch 99.88 ±0.3 50.24 ±11.0 44.34 ±4.2
Resne Imagene 100.0 ±0.0 62.44 ±7.4 47.68 ±5.1
Resne Rad 100.0 ±0.0 69.27 ±3.7 48.28 ±2.7
CNN + LSTM 49.25 ±3.6 42.44 ±4.4 23.84 ±2.7
CNN + Pooling 37.50 ±4.6 37.56 ±6.8 31.43 ±6.5
Medical Ne 1 GPU 31.50 ±4.5 61.82 ±17.5 24.31 ±0.9
Medical Ne 4 GPU 84.62 ±7.7 66.36 ±5.4 41.18 ±1.0
Table 5.1: A g Mac o F1 sco es o models ained on Full images
(a) Resne Imagene (b) Resne Rad
Figu e 5.1: Con usion ma ices om bes models ained on Full images
38
Chap e 5. Expe imen s
Howe e , e en i he mac o a e age F1 o he bes models eaches almos 50%, i can be seen in he
con usion ma ices o he bes -pe o ming models ou o all he uns conduc ed ha such me ic
alues a e aised by he abili ies o he model o iden i y ci ho ic and heal hy pa ien s, a he han
hose p esen ing he PSVD diso de , a class o which he bes F1 sco e is 34.5%, e ealing he poo
pe o mance on he class we a e in e es ed in. This limi a ion shows he need o ind ways o help
he models ocus on impo an pa s o he images while igno ing noise. The nex sec ion looks a
me hods o imp o e he inpu da a by picking ou he key a eas o in e es . By cu ing down noise
and guiding he models o ocus on he mos impo an pa s, hese me hods aim o make i easie
o ind small ea u es linked o PSVD, imp o ing how well he models wo k wi h his ha d- o-de ec
class.
5.1.2 Quali a i e esul s
In addi ion o he p e ious esul s, saliency maps o he bes -pe o ming models we e ob ained
using G ad-CAM om PSVD pa ien s om he Tes se o see i he p edic ions made by such
models a e o be us ed o no .
The saliency maps om he 2D models (Figu e 5.2 show ha hey do no ocus on speci ic impo an
a eas in he images. Ins ead, he models ha e a sca e ed ac i a ion pa e n, which sugges s hey
a e no inding he key ea u es linked o he a ge condi ions. This could mean he models a e
elying on unimpo an o andom pa s o he image, a he han lea ning use ul pa e ns ha
sepa a e he classes.
(a) T ained om sc a ch (b) P e ained on RAD Imagene
Figu e 5.2: Saliency maps ob ained using G ad-CAM om bes 2D models ained on ull images
On he o he hand, Medical Ne shows mo e de ailed and speci ic pa e ns, meaning ha i does
ocus on ce ain pa s o he image such as he li e and spleen (Figu e 5.3. Howe e , such saliency
maps a e homogeneous h oughou he espec i e o gans and do no e lec any speci ic de ail.
Mo eo e , hea maps show high- alue ac i a ions on he able o he ou e pa o he abdomen.
This p oblem shows he need o p ep ocessing me hods ha help he models ocus on he mos
impo an pa s o he image. Me hods like bounding boxes and segmen a ion masks, which highligh
he key a eas, a e discussed in he nex sec ions o sol e his issue and imp o e bo h he accu acy
and in e p e abili y o he models.
39
Chap e 5. Expe imen s
Figu e 5.3: Saliency maps om 3D model Medical Ne ained on Full images
The saliency maps indica e ha models ained on ull images o en ocus on i ele an egions,
sugges ing as ini ially s a ed in RQ1 ha noise impac s he models’ abili y o ocus on diagnos ic
ea u es.
5.2 Using Bounding Boxes
In his app oach, bounding boxes we e applied using he c opping me hod men ioned in Sec ion
3.3.1 o isola e egions o in e es in he images. The idea was o educe noise om i ele an pa s
o he image and make he model ocus on he c i ical a eas.
5.2.1 Quan i a i e esul s
Bounding box-based p ep ocessing imp o ed pe o mance o some models by ocusing on egions o
in e es and emo ing i ele an da a. As shown in Table 5.2, Resne Sc a ch achie ed he highes
es F1 sco e (49.95 ±2.8), while Resne Rad and Resne Imagene showed sligh ly lowe sco es.
This sugges s ha bounding boxes help educe noise bu no as much as necessa y, indica ing ha
mos o he noise is p oduced by elemen s inside he abdomen and hose slices whe e he li e is no
p esen . The e o e, such noise has o be u he educed o inc ease he gene aliza ion capabili ies o
he models.
Again, as in he p e ious sec ion, con usion ma ices o he bes -pe o ming models we e e ie ed
5.4 o assess he pe o mance o hese in he PSVD g oup. While he o e all mac o F1 sco es sligh ly
imp o ed wi h noise- educ ion echniques, he speci ic F1 sco e o PSVD on he bes model only
eached 38.9%. This is a clea imp o emen compa ed o he p e ious sec ion, showing ha using
bounding boxes o educe he amoun o noise in he image helps he models ocus on ea u es ha
a e mo e use ul o de ec ing PSVD. Howe e , he pe o mance is s ill a om ideal, likely because
PSVD has sub le ea u es ha o e lap wi h o he classes, making i ha d o dis inguish.
40
Chap e 5. Expe imen s
Model A g T ain F1 A g Valida ion F1 A g Tes F1
Resne Sc a ch 94.1 ±1.4 60.49 ±7.8 49.95 ±2.8
Resne Imagene 100.0 ±0.0 69.27 ±7.4 46.40 ±4.5
Resne Rad 100.0 ±0.0 69.27 ±2.2 46.80 ±3.9
CNN + LSTM 52.12 ±5.3 33.17 ±5.6 24.03 ±3.4
CNN + Pooling 44.25 ±1.2 39.51 ±6.5 28.76 ±3.6
Medical Ne 1 GPU 38.50 ±6.7 70.90 ±11.8 27.84 ±5.7
Medical Ne 4 GPU 85.99 ±6.0 68.18 ±4.5 32.64 ±7.4
Table 5.2: A g Mac o F1 sco es o models ained on C opped images
(a) Resne Sc a ch (b) Resne Rad
Figu e 5.4: Con usion ma ices om bes models ained on C opped images
Despi e his p og ess, he con usion ma ices show ha PSVD cases a e s ill o en misclassi ied,
usually as ei he heal hy o ci ho ic. While bounding boxes educed noise and imp o ed models’
pe o mances, indica ing ha image noise educes he quali y o he esul s; slices wi hou dis inc i e
li e ea u es s ill led o equen misclassi ica ions, indica ing ha u he e inemen in p ep ocessing
o model aining is equi ed.
5.2.2 Quali a i e esul s
The saliency maps o models using bounding boxes as a noise- educ ion s a egy 5.5 show a change
in ac i a ion pa e ns, wi h a en ion now mo e ocused on ce ain a eas o he images. Howe e ,
hese ocused a eas a e s ill a om he li e , which is he key egion o diagnosing PSVD and
ela ed condi ions. This misplaced ocus sugges s ha , al hough he models ha e imp o ed in
na owing hei a en ion, hey s ill ail o iden i y he impo an ana omical ea u es needed o
accu a e diagnosis.
41

Chap e 5. Expe imen s
This issue shows ha he p ep ocessing echniques es ed so a may no be enough o guide he
models o clinically ele an egions. I highligh s he need o addi ional me hods, such as he use
o segmen a ion masks o di ec he models explici ly o he impo an a eas. The esul s ob ained
o such an app oach will be p o ided in he nex sec ion,
(a) T ained om sc a ch (b) P e ained on RAD Imagene
Figu e 5.5: Saliency maps om bes 2D models ained on c opped images
5.3 Using Segmen a ion Masks
This app oach used segmen a ion masks o isola e he li e om he es o he image, ensu ing ha
he model was p o ided wi h only in o ma ion ele an o he diagnosis o he disease.
5.3.1 Quan i a i e esul s
Using segmen a ion masks p oduced he bes o e all pe o mance, wi h Resne Rad achie ing he
highes es F1 sco e (53.36 ±2.8) as shown in Table 5.3. Segmen a ion e ec i ely emo ed i ele an
in o ma ion, allowing models o ocus on diagnos ic ea u es. Resne Imagene and MedicalNe also
pe o med well bu ell sligh ly sho o Resne Rad. CNN + LSTM showed imp o ed pe o mance
compa ed o he ull-image app oach, indica ing ha segmen a ion e ec i ely emo es noise om
each olume, as no only images a e limi ed o he li e bu also, slices no con aining he o gan a e
disca ded.
Model A g T ain F1 A g Valida ion F1 A g Tes F1
Resne Sc a ch 79.12 ±10.3 55.6 ±7.6 50.50 ±2.8
Resne Imagene 98.38 ±1.3 61.46 ±18.3 48.71 ±1.6
Resne Rad 100.0 ±0.0 67.80 ±5.3 53.36 ±2.8
CNN + LSTM 43.25 ±5.4 47.31 ±2.1 48.11 ±2.7
CNN + Pooling 26.12 ±1.6 35.12 ±8.3 26.03 ±2.4
Medical Ne 1 GPU 55.25 ±10.9 60.00 ±6.1 48.01 ±5.2
Medical Ne 4 GPU 85.25 ±4.9 65.45 ±5.8 46.17 ±3.9
Table 5.3: A g Mac o F1 sco es o models ained on Masked images
42
Chap e 5. Expe imen s
The esul s show ha while segmen a ion masks imp o ed he o e all pe o mance o some classes,
he F1 sco e o he PSVD class s ayed low a 36.1%. This is sligh ly lowe han he sco e achie ed
wi h he bounding box app oach, sugges ing ha e en wi h isola ed egions, he models s ill s uggle
o co ec ly iden i y PSVD. The con usion ma ices highligh his issue, showing ha PSVD cases
a e o en misclassi ied as ei he heal hy o ci ho ic.
The low F1 sco e o PSVD, e en wi h ocused p ep ocessing, shows how di icul i is o de ec his
condi ion. I sugges s ha he isual ma ke s o PSVD may no be clea enough in he isola ed li e
egion, o ha he models a e no p ope ly lea ning he pa e ns needed o iden i y hese ma ke s.
(a) Resne Sc a ch (b) Resne Rad
Figu e 5.6: Con usion ma ices om bes models ained on Masked images
5.3.2 Quali a i e esul s
The saliency maps o he segmen a ion mask models (Figu es 5.7 and 5.8 show ine i ably imp o ed
ocus wi hin he li e egion, which aligns wi h he pu pose o he a ge ed p ep ocessing. Howe e ,
he ac i a ions emain sca e ed and do no concen a e on speci ic ea u es ha a e likely ela ed
o PSVD. This sca e ed a en ion sugges s ha while segmen a ion masks help educe noise, hey
do no au oma ically guide he models o lea n which ea u es a e impo an o de ec ing PSVD.
(a) T ained om sc a ch (b) P e ained on RAD Imagene
Figu e 5.7: Saliency maps om bes 2D models on Masked images
Addi ionally, Figu e 5.8 shows low ac i a ion alues o ming a silhoue e o he li e as i would
appea in slices close o he one being depic ed. Such esul in he saliency maps is due o he
43
Chap e 5. Expe imen s
ac ha he Medical Ne -based app oach comp esses he spa ial dimensions oo much o p o ide
de ailed in o ma ion when he inpu olumes consis o a low amoun o slices, as is he case on he
masked se s. A solu ion o mo e de ailed in o ma ion would be o cap u e in o ma ion om ea lie
laye s in he ne wo k bu ha would esul in less class-o ien ed hea maps.
Figu e 5.8: Saliency maps om 3D model Medical Ne ained on Masked images
5.4 Summa y
The expe imen s conduc ed in his s udy p o ide answe s o he esea ch ques ions (RQs) s a ed in
1.2 by analyzing he pe o mance o imaging models o de ec ing and cha ac e izing PSVD.
RQ1: Impac o Noise in Imaging Da a: Noise in imaging da a, such as non-body elemen s,
i ele an o gans, and i ele an slices, had a signi ican nega i e impac on model pe o mance. The
ull-image app oach s uggled wi h o e i ing and poo gene aliza ion due o he high le el o noise.
Using bounding boxes and segmen a ion masks helped educe noise and imp o ed he model’s ocus
on impo an a eas, leading o be e esul s. Howe e , e en wi h segmen a ion masks, models ound
i di icul o iden i y PSVD-speci ic ea u es. This sugges s ha e en hough noise educ ion has
been p o en o imp o e he models’ pe o mances such p ep ocessing alone is no enough o add ess
he challenges o de ec ing PSVD.
RQ2: Handling Slices Wi hou Visual Cues: Gi en he esul s, i is ha d o con i m whe he
o no he amoun o slices wi hou isual cues is big enough o e en i hese can g ea ly a ec he
ou come. One way o measu e he impac o such slices is o compa e he me hods used o agg ega e
slices’ ea u es wi hin a olume. Since he bes -pe o ming models in gene al a e he ones ha only
used a CNN and agg ega e h ough max o ing i would be logical o hink ha he disease can be
iden i ied in all o mos o he slices o a olume. On he o he hand, he pooling me hod pe o med
conside ably wo se han he LSTM app oach, meaning ha he disease is mos likely no p esen in
all o he slices and is no shown wi h he same deg ee o in ensi y in e e y slice.
RQ3: E ec o O he Classes (Hypo hesis: Remo ing one class om he classi ica ion ask
can help imp o e he F1 sco e o PSVD by making he ask simple and educing he complexi y
he model needs o handle. In a mul i-class se up, he model has o di e en ia e be ween all classes,
which inc eases he chances o con usion, especially when some classes ha e o e lapping ea u es.
By emo ing one class, he model has ewe ca ego ies o lea n, allowing i o ocus mo e on he
44
Chap e 5. Expe imen s
emaining classes, including PSVD.
Fo ins ance, i one class has non-speci ic ea u es such as Miscellaneous, which p esen s se e al
diseases and he e o e di e en cha ac e is ics, such class can con ibu e hea ily o misclassi ica ion
e o s, and emo ing i can help he model o ocus on he es o he classes and be e ecognize
he unique ea u es o PSVD.
This app oach wo ks well when he emo ed class is no impo an o he s udy’s goals o when i s
emo al does no signi ican ly a ec he p ac ical use o he model. By simpli ying he classi ica ion
ask, he model can ocus on he sub le pa e ns linked o PSVD, which may lead o be e de ec ion
and highe F1 sco es o his condi ion.
Chap e 8 will ocus on he s udy o such hypo hesis and will con i m whe he o no he emo al o
one o se e al classes can inc ease he pe o mance o a model on he PSVD class.
45
Chap e 6. P oblem simpli ica ion
As o e e y o he expe imen , models s uggle o e ec i ely iden i y he cha ac e is ics ha
di e en ia e PSVD om Ci hosis, and as o he p e ious expe imen , i looks like he e is a subse
o pa ien s ha is pa icula ly di icul o classi y.
(a) Resne Rad (b) CNN + LSTM
Figu e 6.5: Con usion ma ices o bes -pe o ming models ained in PSVD s Ci hosis expe imen
6.3.2 Quali a i e esul s
Figu e 6.6 shows saliency maps o his bina y classi ica ion ask. Resne Rad and Medical Ne
showed mino imp o emen s in a en ion, wi h ac i a ions sligh ly mo e sca e ed wi hin he li e
egion. Addi ionally, he same p oblem ega ding he spa ial dimensions comp ession o he Medical
Ne model emained p esen which migh make i less eliable o physicians gi en he unce ain y
in he explicabili y.
(a) P e ained on RAD Imagene (b) Medical Ne
Figu e 6.6: Saliency maps om bes models ained in PSVD s Ci hosis expe imen
6.4 Summa y
Answe ing esea ch ques ion 3, he expe imen s in his chap e demons a e ha educing class
complexi y can imp o e model pe o mance, bo h o dis inguishing PSVD om heal hy cases and
PSVD om ci ho ic pa ien s. Howe e , e en wi h bina y classi ica ion, he models s uggle o
isola e PSVD-speci ic ea u es. These indings highligh he need o u he e inemen in da ase
c ea ion and model de elopmen .
52

Chap e 7
Sus ainabili y and E hical Implica ions
En i onmen al Sus ainabili y
De elopmen o he P ojec
•
Ene gy Consump ion: The p ojec in ol es p ocessing a low amoun o da a bu many
imes due o he numbe o expe imen s (6 dis inc expe imen s) and models es ed (7 models
each an 5 imes), leading o an es ima ed inal consump ion o 84 KWh du ing de elopmen
o e 210 uns o 0.4 KWh on a e age pe aining.
•
Ma e ials Used: The de elopmen o he p ojec was ca ied ou using he acili ies o he
Ba celona Supe compu ing Cen e , which makes use o N idia H100 g aphics ca ds.
•
Reduc ion Measu es: Measu es, such as op imizing compu a ional e iciency, ha e been
implemen ed o minimize ene gy consump ion. Addi ionally, measu es applied du ing he
p ojec de elopmen o educe he noise in he da ase esul ed in he educ ion o he da ase
size, leading o educed aining imes and educed powe consump ion. Fo ins ance, he
powe consump ion o one un on expe imen one was on a e age 0.5 KWh whe eas he
consump ion o one aining on he las expe imen (bina ized) was on a e age 0.25 KWh.
Execu ion o he P ojec
•
Resou ce U iliza ion: All o he models de eloped can be execu ed on compu e s ha do
no e en ha e a GPU (no using one leads o longe execu ion imes, bu s ill hese make i s
use easible).
Economic Sus ainabili y
De elopmen o he P ojec
•
Cos s: P ima y cos s in ol e compu a ional esou ces and pe sonnel hou s which sum up o
mo e han 11k €.
•
E iciency Measu es: As men ioned be o e, he op imiza ion o he da a helps educe he
compu ing powe needed and he e o e i s cos .
53
Chap e 7. Sus ainabili y and E hical Implica ions
Execu ion o he P ojec
•
Viabili y: Any a ailable decen compu e in a hospi al could execu e hese models, so no cos
would be assumed in ega ds o new ha dwa e.
Social Sus ainabili y
De elopmen o he P ojec
•
Skill De elopmen : De elopmen o he p ojec in a p o essional en i onmen imp o ed he
echnical and analy ical skills o he s uden .
Execu ion o he P ojec
•
Impac on S akeholde s: Po en ial bene icia ies include esea che s and clinicians. Addi-
ionally, i inally applied in a eal-li e scena io, all o he pa ien s whose diagnosis is in luenced
by esul s shown by he de eloped model will be a ec ed oo.
Risks and Limi a ions
•
E hical Conce ns: Pa ien s’ da a p i acy and po en ial biases in AI algo i hms need
con inuous assessmen o make su e ha none o he da a is unsa ely eleased o leaked and
ha he models main ain nondisc imina o y beha io .
E hical Implica ions
•
Responsi eness o Needs: The p ojec add esses a speci ic gap in medical imaging analysis,
aiming o imp o e heal hca e ou comes by bo h p o iding insigh ul ad ances in he esea ch
o such no el disease and imp o ing he decision p ocess o he diagnosis.
•
An icipa ed Consequences: Risks such as da a misuse o biased ou pu s a e acknowledged
and moni o ed.
•
Model’s ole: The p ojec aims o p o ide a ool o physicians o be e unde s and a a e
disease and a ool o aid in he diagnosis p ocess a he han doing such a diagnosis alone. In
he sho e m, i is a decision suppo sys em.
Alignmen wi h Sus ainable De elopmen Goals (SDGs)
Rele an SDG Con ibu ion
SDG 3 (Good Heal h and Well-
being)
Enhancing diagnos ic ools o be e heal hca e ou -
comes.
SDG 9 (Indus y, Inno a ion, and
In as uc u e)
P omo ing sus ainable inno a ion in medical imaging.
Table 7.1: Suis ainable De elopmen Goals
54
Chap e 7. Sus ainabili y and E hical Implica ions
Conclusion
The p ojec shows a clea e o o e alua e i s en i onmen al, economic, and social impac s while
conside ing e hical issues. Ongoing imp o emen s and in ol emen o s akeholde s a e needed o
ensu e long- e m sus ainabili y and alignmen wi h e hical s anda ds.
55
Chap e 8
Conclusions
This hesis explo ed he use o ad anced imaging models o de ec ing and cha ac e izing Po o-
Sinusoidal Vascula Diso de (PSVD), add essing he challenges o diagnosing his a e li e disease.
By applying deep lea ning echniques o medical imaging da a, he esea ch aimed o imp o e
diagnos ic accu acy while ackling issues such as da a noise, o e lapping class ea u es, and he
complexi y o ea u es cha ac e izing PSVD.
8.1 Resea ch Ques ions Add essed
RQ1: Impac o Noise in Imaging Da a: The esea ch showed ha noise, such as non-li e
egions and i ele an ana omical s uc u es, signi ican ly a ec ed model pe o mance. P ep ocessing
me hods like c opping and segmen a ion masks educed noise and helped he models ocus on ele an
egions. Howe e , e en wi h hese me hods, he models s uggled o isola e PSVD-speci ic ea u es,
highligh ing he sub le and complex na u e o he diso de ’s isual ma ke s.
RQ2: Handling Slices Wi hou Visual Cues: The analysis e ealed ha slices wi hou clea
isual indica o s o PSVD s ill con ibu ed o p edic ions when p ocessed oge he . Agg ega ion
s a egies, such as majo i y o ing o slice-le el p edic ions, imp o ed o e all accu acy and demon-
s a ed ha he disease is isible in mos o he slices a leas . O he s a egies such as pooling,
showed by ailing in compa ison o o he s ha he disease is p esen clea e in some slices han
o he s. Howe e , hese indings show he need o he de elopmen o be e s a egies o make ull
use o he da a.
RQ3: E ec o O he Classes: Reducing he numbe o classes in he classi ica ion ask imp o ed
pe o mance. Bina y classi ica ions, such as PSVD s. Heal hy o PSVD s. Ci hosis, helped he
models ocus on sub le ea u es unique o PSVD. Howe e , he expe imen s also showed he di icul y
o dis inguishing PSVD om condi ions wi h o e lapping imaging cha ac e is ics, like ci hosis.
Addi ionally, as ag eed wi h physicians om IDIBAPS, bina y classi ica ion p oblems would be mo e
app op ia e in u u e wo k, especially asks we e PSVD is o be disce ned om Ci hosis.
8.2 Key Con ibu ions
•
De eloped a p ep ocessing pipeline ha includes noise educ ion, da a augmen a ion, and
no maliza ion o imp o e da ase quali y and consis ency.
57

Chap e 8. Conclusions
•
E alua ed mul iple deep lea ning a chi ec u es (2D, 3D, and hyb id models) o iden i y op imal
con igu a ions o medical imaging asks.
•
Demons a ed he alue o domain-speci ic p e ained models (e.g., RadImageNe ) o imp o ing
diagnos ic accu acy and obus ness.
•
Highligh ed he impo ance o explainable AI ools, like G ad-CAM, o p o ide insigh s in o
model decisions, building us and u ili y in clinical p ac ice.
8.3 Limi a ions and Fu u e Wo k
Despi e he p og ess made, he esea ch aced se e al limi a ions:
•
Da ase Size and Di e si y: The small and imbalanced (in e ms o me ada a, no classes)
da ase limi ed he models’ abili y o gene alize ac oss di e en popula ions and imaging
condi ions.
•
Sub le Fea u es o PSVD: The lack o clea isual ma ke s o PSVD educed model
pe o mance, equi ing u he esea ch in o ea u e enginee ing and in e p e abili y.
•
Clinical Valida ion: Addi ional alida ion wi h ex e nal da ase s and clinical ials is needed
o ensu e he models a e eliable in eal-wo ld applica ions.
Fu u e wo k should ocus on:
•Inc easing da ase size and di e si y o imp o e gene aliza ion.
•
Fu he imp o ing he p ep ocessing pipeline o ensu e all images p esen he same condi ions,
like o example, s anda dized li e alues, in a ian o he amoun o ype o con as supplied
o he pa ien .
•
Explo ing new a chi ec u es be e sui ed o a e disease de ec ion ha makes be e use o
he hi d dimension o samples.
•
Discuss wi h medical p o essionals o e ine he models’ u ili y and in e ope abili y, ensu ing
also e hical and esponsible de elopmen .
8.4 Final Rema ks
P e ious wo k in he ield has been imp o ed, especially since he wo k p esen ed by [13] was no
able o gene alize o he Tes da ase conce ning he PSVD g oup o pa ien s, he e o e showing ha
imp o emen s in he PSVD esea ch line can be ob ained by using AI. This esea ch demons a es he
po en ial o a i icial in elligence in diagnosing a e diseases like PSVD. By add essing key challenges
and p oposing clinically ele an solu ions, his hesis lays a s ong ounda ion o de eloping
mo e obus , in e p e able, and impac ul diagnos ic ools. The indings no only imp o e he
unde s anding o medical da ase s e e ed o PSVD bu can also be ex apola ed o he b oade
ield o AI-based medical image analysis.
58
Bibliog aphy
[1]
De Go a di A. e al. “Po o-sinusoidal ascula disease: p oposal and desc ip ion o a no el
en i y.” en. In: Lance Gas oen e ol. Hepa ol. 4.5 (May 2019), pp. 399–411.
[2]
And ea De Go a di, Ch is ine Sempoux, and Annalisa Be zigo i. “Po o-sinusoidal ascula
diso de ”. en. In: J. Hepa ol. 77.4 (Oc . 2022), pp. 1124–1135.
[3]
Misla Ba isic-Jaman e al. “Po o-sinusoidal ascula disease: a new de ini ion o an old
clinical en i y”. en. In: Clin. Exp. Hepa ol. 9.4 (Dec. 2023), pp. 297–306.
[4]
Elk ie L. e al. “Po o-sinusoidal ascula diso de : Diagnos ic challenges and managemen ”.
In: Jou nal o Hepa ology (2023). u l:h ps://pubmed.ncbi.nlm.nih.go /30957754/.
[5]
D. e al. Cazals-Ha em. “His opa hology o PSVD”. In: Hepa ology In e na ional (2019). u l:
h ps://pubmed.ncbi.nlm.nih.go /35690264/.
[6]
J. N. Schou en e al. “Idiopa hic non-ci ho ic po al hype ension: A e iew o i s pa ho-
physiological mechanisms”. In: Hepa ology (2011). doi:
10.1002/hep.24445
.u l:
h ps:
//aasldpubs.onlinelib a y.wiley.com/doi/ ull/10.1002/hep.24445.
[7]
Sunda am V. e al. “Eme ging diagnos ic ools in PSVD”. In: Li e In e na ional (2020). u l:
h ps://pmc.ncbi.nlm.nih.go /a icles/PMC11103802/.
[8]
D. T ipa hi e al. “Managemen o po al hype ension: UK consensus guidelines”. In: Gu
(2020). doi:
10.1136/gu jnl-2020-320402
.u l:
h ps://gu .bmj.com/con en /69/11/
2116.
[9]
Ve heij J. e al. “Ad ances in nonin asi e diagnos ics o PSVD”. In: Li e Resea ch (2021).
u l:
h ps://www.else ie .es/en- e is a-medicina-clinica-english-edi ion--
462-a iculo-po osinusoidal- ascula -diso de -a-pa adigm-S2387020624001475.
[10]
Fahad Mu lih Alshaga h h and Mowa a Said Househ. “A i icial In elligence o De ec ing
and Quan i ying Fa y Li e in Ul asound Images: A Sys ema ic Re iew”. In: Bioenginee ing
(2022). u l:h ps://www.mdpi.com/2306-5354/9/12/748.
[11]
Zhou LQ e al. “A i icial in elligence in medical imaging o he li e ”. In: Wo ld Jou nal o
Gas oen e ology (2019).
59
Bibliog aphy
[12]
Naoshi Nishida. “Ad ancemen s in A i icial In elligence-Enhanced Imaging Diagnos ics o
he Managemen o Li e Disease—Applica ions and Challenges in Pe sonalized Ca e”. In:
Bioenginee ing (2024). u l:h ps://www.mdpi.com/2306-5354/11/12/1243.
[13]
Ruben Cue o Nogue a. “Deep lea ning app oaches o de ec ing and explaining hepa ic
diso de s om CT scans”. PhD hesis. UPC, Facul a d’In o mà ica de Ba celona, Depa amen
de Ciències de la Compu ació, Jan. 2024. u l:h p://hdl.handle.ne /2117/406465.
[14]
Da ia Ke n and And e Mas meye . “3D bounding box de ec ion in olume ic medical image
da a: A sys ema ic li e a u e e iew”. In: J. Image G aph. 10.1 (2022).
[15]
Ni upama Cha e jee e al. “Is he li e esilien o he p ocess o ageing?” en. In: Ann. Hepa ol.
30.2 (Sep . 2024), p. 101580.
[16]
Nicolás Ál a ez Llopis. “F om di e se CT scans o gene aliza ion: owa ds obus abdominal
o gan segmen a ion”. PhD hesis. UPC, Facul a d’In o mà ica de Ba celona, Depa amen de
Ciències de la Compu ació, June 2024. u l:h p://hdl.handle.ne /2117/413967.
[17]
M. Jo ge Ca doso and Wenqi Li e al. MONAI: An open-sou ce amewo k o deep lea ning in
heal hca e. 2022. a Xi : 2211.02701 [cs.LG].u l:h ps://a xi .o g/abs/2211.02701.
[18]
E shuai Wang, Yaliang Zhao, and Yajun Wu. “Cascade Dual-decode s Ne wo k o Abdominal
O gans Segmen a ion”. In: Fas and Low-Resou ce Semi-supe ised Abdominal O gan Segmen-
a ion. Ed. by Jun Ma and Bo Wang. Cham: Sp inge Na u e Swi ze land, 2022, pp. 202–213.
isbn: 978-3-031-23911-3.
[19]
Ziyan Huang e al. “Re isi ing nnU-Ne o I e a i e Pseudo Labeling and E icien Sliding
Window In e ence”. In: Fas and Low-Resou ce Semi-supe ised Abdominal O gan Segmen a ion.
Ed. by Jun Ma and Bo Wang. Cham: Sp inge Na u e Swi ze land, 2022, pp. 178–189. isbn:
978-3-031-23911-3.
[20]
Tobias e al Heimann. “Compa ison and E alua ion o Me hods o Li e Segmen a ion F om
CT Da ase s”. In: IEEE T ansac ions on Medical Imaging 28.8 (2009), pp. 1251–1265. doi:
10.1109/TMI.2009.2013851.
[21]
Jun Ma e al. “Unleashing he s eng hs o unlabelled da a in deep lea ning-assis ed pan-cance
abdominal o gan quan i ica ion: he FLARE22 challenge”. In: The Lance Digi al Heal h 6.11
(2024), e815–e826. doi:h ps://doi.o g/10.1016/S2589-7500(24)00154-7.
[22]
Saha Sabou i e al. “Adding Li e Window Se ing o he S anda d Abdominal CT Scan
P o ocol: Is I Use ul?” In: I anian Jou nal o Radiology (June 2008).
[23]
Muhammad Waheed Sabi e al. “Segmen a ion o Li e Tumo in CT Scan Using ResU-
Ne ”. In: Applied Sciences 12.17 (2022). issn: 2076-3417. doi:
10.3390/app12178650
.u l:
h ps://www.mdpi.com/2076-3417/12/17/8650.
60
Bibliog aphy
[24]
Muhammad Islam, Kaleem Nawaz Khan, and Muhammad Salman Khan. E alua ion o
P ep ocessing Techniques o U-Ne Based Au oma ed Li e Segmen a ion. 2021. a Xi :
2103.
14301 [eess.IV].u l:h ps://a xi .o g/abs/2103.14301.
[25]
Yann A. LeCun e al. “E icien BackP op”. In: Neu al Ne wo ks: T icks o he T ade: Second
Edi ion. Be lin, Heidelbe g: Sp inge Be lin Heidelbe g, 2012, pp. 9–48. isbn: 978-3-642-35289-8.
doi:
10.1007/978-3-642-35289-8_3
.u l:
h ps://doi.o g/10.1007/978-3-642-35289-
8_3.
[26]
C. Sho en and T. M. Khoshgo aa . “A su ey on image da a augmen a ion o deep lea ning”.
In: Jou nal o Big Da a 6.1 (2019), pp. 1–48. doi:
10.1186/s40537- 019- 0197- 0
.u l:
h ps://doi.o g/10.1186/s40537-019-0197-0.
[27]
S. C. Wong e al. “Unde s anding da a augmen a ion o classi ica ion: When o wa p?” In:
2016 In e na ional Con e ence on Digi al Image Compu ing: Techniques and Applica ions
(DICTA). 2016, pp. 1–6. doi:
10.1109/DICTA.2016.7797091
.u l:
h ps://doi.o g/10.
1109/DICTA.2016.7797091.
[28]
P. Y. Sima d, D. S eink aus, and J. C. Pla . “Bes p ac ices o con olu ional neu al ne wo ks
applied o isual documen analysis”. In: Se en h In e na ional Con e ence on Documen
Analysis and Recogni ion, 2003. P oceedings. 2003, pp. 958–963. doi:
10.1109/ICDAR.2003.
1227801.u l:h ps://doi.o g/10.1109/ICDAR.2003.1227801.
[29]
Sco Reed e al. T aining Deep Neu al Ne wo ks on Noisy Labels wi h Boo s apping. 2015.
a Xi : 1412.6596 [cs.CV].u l:h ps://a xi .o g/abs/1412.6596.
[30]
Kaiming He e al. “Deep Residual Lea ning o Image Recogni ion”. In: CoRR abs/1512.03385
(2015). a Xi : 1512.03385.u l:h p://a xi .o g/abs/1512.03385.
[31]
Sepp Hoch ei e and Jü gen Schmidhube . “Long Sho -Te m Memo y”. In: Neu al Compu a ion
9.8 (1997), pp. 1735–1780. doi:10.1162/neco.1997.9.8.1735.
[32]
Sh eyasi Roy Chowdhu y, Yash Kha e, and Susmi a Mazumda . “Chap e 12 - Classi ica ion
o diseases om CT images using LSTM-based CNN”. In: Diagnos ic Biomedical Signal
and Image P ocessing Applica ions wi h Deep Lea ning Me hods. Ed. by Kemal Pola and
Saban Öz ü k. In elligen Da a-Cen ic Sys ems. Academic P ess, 2023, pp. 235–249. isbn:
978-0-323-96129-5. doi:
h ps://doi.o g/10.1016/B978-0-323-96129-5.00008-1
.u l:
h ps://www.sciencedi ec .com/science/a icle/pii/B9780323961295000081.
[33]
Sunil Kuma e al. “CNN-BO-LSTM: an ensemble amewo k o p ognosis o li e cance ”. en.
In: In . J. In . Technol. (Sep . 2024). doi:
h ps://doi.o g/10.1007/s41870-024-02190-5
.
[34]
Dong Liang e al. “Combining Con olu ional and Recu en Neu al Ne wo ks o Classi ica ion
o Focal Li e Lesions in Mul i-phase CT Images”. In: Medical Image Compu ing and Compu e
Assis ed In e en ion – MICCAI 2018. Ed. by Alejand o F. F angi e al. Cham: Sp inge
In e na ional Publishing, 2018, pp. 666–675. isbn: 978-3-030-00934-2. doi:
10.1007/978-3-
030-00934-2_74.
61