% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@MASTERSTHESIS{Rathore:1038863,
author = {Rathore, Omini},
title = {{U}ncertainty {E}stimation in the {S}egmentation of {B}rain
{M}agnetic {R}esonance {A}ngiograms},
school = {RWTH},
type = {Masterarbeit},
reportid = {FZJ-2025-01679},
pages = {52},
year = {2025},
note = {Masterarbeit, RWTH, 2024},
abstract = {Abstract:The analysis of brain vessel scans is a critical
step in detecting cerebrovascular diseases.Manual
segmentation of cerebrovascular structures is a
labour-intensive process susceptibleto inter-observer
variability, which has led to extensive research into
automatic segmentationusing deep learning techniques. These
methods have yielded promising models;however, their opaque
nature and methodological complexity hinder understanding,
interpretability,and adoption by clinicians in practical
settings. Furthermore, MagneticResonance Angiogram data is
inherently complex, as cerebrovascular structures
exhibitsignificant variability across individuals and
scanners, further challenging the reliability ofautomated
models.This thesis seeks to bridge this gap by incorporating
uncertainty quantification techniquesinto deep
learning-based cerebrovascular segmentation models.
Uncertainty quantificationmeasures the model’s confidence
in its predictions, providing insights that enhance
interpretability.Moreover, uncertainty maps can identify
regions with elevated uncertainty,thereby increasing
clinicians’ trust in the model. In this work, we implement
an efficientensemble model utilizing Bayesian Approximation
to estimate uncertainty. The alignmentbetween areas of high
uncertainty and erroneous predictions demonstrates the
effectivenessof the ensemble model in generating reliable
uncertainty maps. The proposed model outperformsthe baseline
point-estimate model, with further improvements achieved
throughthe integration of uncertainty estimates.The
model’s robustness is further validated on
out-of-distribution datasets, including MRAscans from
diverse scanners and populations of both healthy and
unhealthy patients. Theresulting uncertainty maps elucidate
the higher error rates in these datasets, reinforcingthe
model’s ability to explain its limitations and maintain
trustworthiness.iii},
cin = {IAS-8},
cid = {I:(DE-Juel1)IAS-8-20210421},
pnm = {5112 - Cross-Domain Algorithms, Tools, Methods Labs (ATMLs)
and Research Groups (POF4-511)},
pid = {G:(DE-HGF)POF4-5112},
typ = {PUB:(DE-HGF)19},
url = {https://juser.fz-juelich.de/record/1038863},
}