Development and Evaluation of a nnU-Net-Based Convolutional Neural Network for Automated Bi-Class Segmentation of Cardiac Atria from Late Gadolinium-Enhanced Magnetic Resonance Imaging

Thesis Proposal Details

Supervisor: Giacomo Cappon

Co-supervisor: Ali Gharaviri, Edoardo Bori

Co-supervisor Department/Company: Heart Rhythm Research Brussels, UZ Brussel, ECAM Brussel

Creation Date: 06/02/2026 10:12

Description

The thesis project aims to develope AI-based deep learning model for medical image

segmentation, with a primary focus on obtaining highly accurate cardiac atrial anatomy

for advanced clinical applications. The work involves refining architectures such as U-

Net and transformer-based networks through improved training strategies, larger

datasets, and organ-specific optimisation. The resulting segmented data will support

two main objectives: the creation of realistic 3D-printed organ models for surgical

training and the generation of patient-specific heart surface reconstructions for high-

resolution cardiac mapping.

Dataset and methods

Dataset type: Already acquired data

Dataset description: 3D Late Gadolinium-Enhanced MRIs and related ground truth segmentation masks

List of Methods: Deep learning, convolutional neural network, computational infrastructure, Linux, model training, testing, and evaluation

Preparatory Courses

Deep learning

Tags
#erasmus
Back to proposals list