Grafični vmesnik za računalniško-podprto analizo angiografskih slik z modeli globokega učenja

KOS, ERAZEM (author), Špiclin, Žiga (mentor), Bizjak, Žiga (comentor)

Intrakranialne anevrizme imajo visoko prevalenco, s približno eno anevrizmo na 20 do 30 oseb, in so večinoma asimptomatske. Ena četrtina anevrizem bo rupturirala tekom življenja, kar običajno povzroči hemoragično možgansko kap - resno zdravstveno stanje z visoko smrtnostjo (50%), medtem ko 66% preživelih utrpi trajno nevrološko okvaro. Pri tovrstnih krvavitvah je 20% incidenca smrti in te so nenadne in se zgodijo zunaj bolnišnice. Nevroradiologi se za diagnozo intrakranialnih anevrizem poslužujejo 3D slikovnih preiskav kot sta računalniško tomografska ali magnetno resonančna angiografija (CTA in MRA), zajete 3D slike pa prikazujejo z grafičnimi vmesniki zgolj v obliki 2D rezin, kar pomeni da običajno opazujejo naenkrat le majhen del (0,7%) celotne slike. Za zaznavanje anevrizem in vrednotenje njihove 3D oblike je zato potrebna dobra mentalna predstava, ki temelji predvsem na izkušnjah posameznega nevroradiologa, posledično pa je tudi odločanje o nujnosti posega pogosto subjektivno. V magistrski nalogi smo razvili in vrednotili računalniški program z grafičnim vmesnikom, ki na 3D MRA in CTA slikah glave omogoča odkrivanje možnih lokacij anevrizem z visoko občutljivostjo (zazna >95% vseh anevrizem) in kvantitativno analizo njihove morfologije. Ključni sestavni deli so programski moduli razviti z metodami umetne inteligence in strojnega učenja: (i) modul za razgradnjo intrakranialnega ožilja na CTA slikah, (ii) modul za razgradnjo intrakranialnega ožilja na MRA slikah in (iii) modul za zaznavanje anevrizem na razgradnji intrakranialnega ožilja. Grafični vmesnik omogoča 3D upodabljanje razgrajenega ožilja in anevrizem, 2D prikaz posameznih rezin 3D slike in ročne meritve dimenzij označenih anatomskih struktur. Slednje omogoča objektivnost vrednotenja morfologije intrakranialne anevrizme in posledično na objektivnih dejstvih temelječe odločanje o njihovem nadaljnjem zdravljenju. Neposreden učinek rabe razvitega programa z grafičnim vmesnikom je samodejno, hitrejše in izboljšano diagnosticiranje intrakranialnih anevrizem, kar bi razbremenilo zdravniško osebje in hkrati omogočilo presejalno slikovno testiranje ter tako zmanjšalo smrtnost zaradi spontane rupture neodkrite in nezdravljene intrakranialne anevrizme. Po drugi strani je vrednotenje morfologije anevrizem podlaga za oceno tveganja prihodnje rupture anevrizme in kot taka v pomoč nevroradiologu pri izbiri načina zdravljenja.

• Language: Slovenian
• Keywords: razgradnja intrakranialnega ožilja, avtomatsko zaznavanje anevrizem, globoko učenje, ročna referenčna razgradnja, objektivno vrednotenje, grafični vmesnik za prikaz 2D in 3D slik, 3D upodabljanje, analiza morfologije
• Work type: Master's thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2022
• PID: 20.500.12556/RUL-137853
• COBISS.SI-ID: 113809155
• Publication date in RUL: 04.07.2022

Secondary language

• Language: English
• Title: Graphical user interface for computer-assisted analysis of angiographic scans based on deep learning models

Abstract

Intracranial aneurysms occur in approximately one out of 20 to 30 people and are typically asymptomatic. One quarter of the aneurysms will rupture during patient's lifetime, which can cause a hemorrhagic brain stroke - a serious condition with a high mortality rate (50%), while 66% of survivors will suffer from permanent neurological deficits. Among such bleedings there 20% of sudden deaths that occur out of hospital environments. Diagnosis of intracranial aneurysms is based on computed tomography or magnetic resonance angiography (CTA or MRA, respectively) imaging, whereas neuroradiologists use programs with graphic user interfaces that mostly only render 2D slices of 3D images, thereby depicing simultaneously only a small fraction (0,7% of the whole image). Therefore, to detect the aneurysms and evaluate their 3D morphology, a good mental reconstruction of these structures is needed, which is gained from the past experiences. Consequently their resulting decisions about further treatments could be somewhat subjective. In this thesis we developed and evaluated a computer program with a graphical user interface, which allows to detect potential aneurysm locations in 3D MRA and CTA images with a high sensitivity rate (>95% of all aneurysms). Key parts of the program are modules developed using deep learning and machine learning methods: (i) module for segmentation of intracranial vessels in CTA images, (ii) module for segmentation of intracranial vessels in MRA images and (iii) module for aneurysm detection from intracranial vessel segmentations. The graphical user interface allows 3D rendering of vessel and aneurysm segmentation, 2D slice-wise display of the 3D image and manual annotation tools, which enable objective aneurysm morphology evaluation and thus an evidence-based treatment decision-making. The use of said program allows for an automatic, faster and more accurate diagnosis of intracranial aneurysms, which would relieve medical staff's workload and, at the same time, enable screening imaging tests so as to lower death rates caused by spontaneous ruptures of undiagnosed and untreated intracranial aneurysms. On the other hand, the morphology quantification is the basis for future rupture risk assessment and thus helpful in determining the optimal and timely treatment.

• Keywords: intracranial vessel segmentation, automatic aneurysm detection, deep learning, manual reference segmentation, objective evaluation, graphical user interface for 2D and 3D image visualization, 3D rendering, analysis of morphology