Generation of distortion-corrected BOLD fMRI images without the presence of magnetic field maps

Nikolov, Ivan (author), Demšar, Jure (mentor)

Susceptibility-induced distortions degrade EPI image quality and can misalign functional images with anatomical scans. Correcting these distortions is therefore important for reliable downstream analysis. Classical correction methods require additional acquisitions, such as field maps or reversed phase-encoding images, which are often unavailable in legacy datasets or time-constrained clinical settings. This thesis investigates whether deep learning can correct susceptibility-induced distortions from a single distorted image and a T1-weighted anatomical scan. We compare several neural network architectures using data from three Human Connectome Project cohorts. We then evaluate the selected MedNeXt v1 Medium model in a complete 4D fMRI pipeline. The proposed approach achieves the best RMSE, SSIM, and NMI across all cohorts. It outperforms the uncorrected baseline, SynBOLD-DisCo, and a topup-refined variant of itself. It performs well across all evaluated phase-encoding directions. Inference requires only a single feed-forward pass and can run on a standard CPU without a dedicated GPU.

• Language: English
• Keywords: functional MRI, BOLD, susceptibility distortion correction, deep learning, image-to-image translation, Human Connectome Project
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-189779
• Publication date in RUL: 15.10.2026

Secondary language

• Language: Slovenian
• Title: Generiranje distorzijsko popravljenih BOLD fMRI slik brez prisotnosti slik magnetnega polja

Abstract

Distorzije zaradi magnetne susceptibilnosti poslabšajo kakovost EPI posnetkov in lahko povzročijo neskladje med funkcionalnimi in anatomskimi slikami, zato je njihovo popravljanje pomembno za zanesljive nadaljnje analize. Klasične metode zahtevajo dodatne zajeme, kot so posnetki magnetnega polja ali slike z obratno smerjo faznega kodiranja, ki pogosto niso na voljo v starejših podatkovnih zbirkah ali časovno omejenih kliničnih okoljih. V magistrskem delu raziskujemo uporabo globokega učenja za popravljanje distorzij na podlagi ene slike z distorzijami in T1-utežene anatomske slike. Primerjamo več arhitektur nevronskih mrež na podatkih treh kohort projekta Human Connectome Project ter izbrani model MedNeXt v1 Medium ocenimo v celotnem postopku obdelave 4D fMRI. Predlagani pristop v vseh kohortah dosega najboljše rezultate RMSE, SSIM in NMI ter presega nepopravljeno izhodišče, SynBOLD-DisCo in svojo različico z dodatnim postopkom topup. Dobro deluje pri vseh ovrednotenih smereh faznega kodiranja. Za sklepanje potrebuje le en prehod skozi nevronsko mrežo in se lahko izvaja na običajnem procesorju brez namenske grafične kartice.

• Keywords: funkcionalno magnetnoresonančno slikanje, BOLD, popravljanje distorzij zaradi susceptibilnosti, globoko učenje, prevajanje slika-v-sliko, Human Connectome Project