Artificial intelligence based analysis of cerebral angiograms for aneurysm diagnosis and prognosis

BIZJAK, ŽIGA (author), Špiclin, Žiga (mentor)

Vascular diseases represent a significant global health concern, ranking as the foremost cause of disability and mortality worldwide. Accounting for approximately 32 % of all deaths, these diseases predominantly impact the cardiac and cerebral vasculatures. Among the spectrum of cerebrovascular pathologies, intracranial aneurysms (IA) emerge as a prevalent concern, manifesting as balloon-like protrusions from weakened vessel segments. IA affect 2 % to 8 % of the global population and, if ruptured, can result in stroke, a condition that is both serious and life-threatening. The management of IA encompasses crucial steps such as detection, isolation, morphological measurements, rupture risk assessment, and growth evaluation. Presently, skilled radiologists manually perform these tasks in clinical settings, introducing the potential for intra- and inter-rater variability. The automation of these processes holds promise for enhancing IA management and furnishing valuable insights for deciding between surgical intervention and follow-up imaging, a determination typically based on rupture risk assessment. This doctoral thesis focuses on the creation and validation of modality-independent computer-aided methods for IA management. Our initial step involves utilizing deep learning to segment intracranial vessels across various angiographic modalities. Following that, we transformed these segmented vessels into trigonometric meshes, a format that is modality-independent. Subsequent steps, including detection, isolation, and rupture prediction, are executed on these meshes using innovative deep-learning algorithms to attain state-of-the-art outcomes throughout the entire process. Moreover, in the case of a follow-up approach, we have devised an algorithm for evaluating intracranial growth between two consecutive imaging sessions. The development and evaluation of these methodologies were carried out using a substantial dataset comprising over 2000 CTA and 1000 MRA images sourced from multiple hospitals. Our algorithms furnish clinicians with critical information at each stage of IA management, thereby mitigating intra- and inter-rater variability and elevating the standard of patient care.

• Language: English
• Keywords: Intracranial aneurysm, Segmentation, Deep learning, Intracranial vessels, Rupture risk prediction
• Work type: Doctoral dissertation
• Organization: FE - Faculty of Electrical Engineering
• Year: 2024
• Publication date in RUL: 01.07.2024

Secondary language

• Language: Slovenian
• Title: Analiza možganskih angiogramov z metodami umetne inteligence za diagnozo in prognozo anevrizem

Abstract

Eden večjih svetovnih zdravstvenih problemov so žilne bolezni, ki predstavljajo približno 32 % vseh smrti. Žilne bolezni prizadenejo predvsem srčno in možgansko ožilje. Med slednjimi patologijami prevladujejo znotrajlobanjske anevrizme, ki se kažejo kot balonom podobni izrastki iz oslabljenih segmentov žil. Znotrajlobanjske anevrizme prizadenejo od 2 % do 8 % svetovnega prebivalstva, njihova ruptura pa lahko vodi v življenjsko nevarno možgansko kap. Obvladovanje znotrajlobanjskih anevrizem zajema ključne korake, kot so odkrivanje, izolacija, morfološke meritve, ocena tveganja za rupturo in ocena rasti. Trenutno v kliničnih okoljih usposobljeni radiologi izvajajo te naloge ročno, kar doprinese k intra- in interindividualni variabilnosti. Avtomatizacija teh postopkov je nujna za izboljšanje obvladovanja znotrajlobanjskih anevrizem in zagotavljanje točnih informacij za izbiro načina zdravljenja, kjer odločitev običajno temelji na oceni tveganja za rupturo. Pričujoča doktorska disertacija se osredotoča na zasnovo in validacijo računalniško podprtih metod za upravljanje znotrajlobanjskih anevrizem, ki so neodvisne od modalitete slikanja. Naš začetni korak vključuje uporabo globokega učenja za segmentacijo znotrajlobanjskih žil v različnih angiografskih modalitetah in njihovo pretvorbo v trigonometrično mrežo, ki je modalitetno neodvisen tip podatka. Nadaljnji koraki, vključno z odkrivanjem, izolacijo in napovedovanjem rupture, se izvajajo na trigonometričnih mrežah z uporabo inovativnih in sodobnih algoritmov globokega učenja, ki za vsak korak dosežejo najbolj optimalne rezultate. Poleg tega smo zasnovali algoritem za ocenjevanje rasti med zaporednima slikanjema znotrajlobanjskih anevrizem. Razvoj in vrednotenje teh metodologij sta bila izvedena z uporabo zelo obsežnega nabora slik, pridobljenih iz več bolnišnic. Naši algoritmi zagotavljajo zdravnikom ključne informacije na vsakem koraku obvladovanja znotrajlobanjskih anevrizem, s čimer zmanjšujejo intra- in interindividualno variabilnost ter dvigujejo standard oskrbe bolnikov.

• Keywords: znotrajlobanjska anevrizma, segmentacija, globoko učenje, znotrajlobanjsko ožilje, napoved rupture anevrizme