Alen Jose Anto

Thesis Topic: RadRecon: A Self-Supervised Framework for Radial MRI Reconstruction Using ESPIRiT-Based Sensitivity Maps

Supervisors: Erik Gösche, Prof Dr. Florian Knoll

Description

This thesis introduces RadRecon, a self-supervised deep learning framework for multi-coil Magnetic Resonance Imaging (MRI) reconstruction designed to reduce reliance on fully sampled ground-truth data. RadRecon builds on the Self-Supervised Learning via Data Undersampling (SSDU) strategy and adds a spoke-splitting method, where the acquired k-space data are divided into two separate subsets (ω,ε). Each subset helps supervise the reconstruction of the other through a two-way learning process, allowing for internal consistency without needing external references. The reconstruction objective uses a relative mixed ϑ1–ϑ2 loss to balance smoothness and structural preservation, along with additional image-domain and perceptual loss terms to support stable training and improve visual quality.