From Rule-Based Algorithms to Deep Learning: Advancing Automated Valve Segmentation in 2D Phase-Contrast MRI

Two-dimensional phase-contrast magnetic resonance imaging (2D PC-MRI) is the reference standard for non-invasive blood flow quantification across cardiac valves and vessels. Despite its clinical relevance, current workflows remain semi-manual: initialization of the flow centre, contour propagation, and manual correction are still required in software such as Medis Suite MR (QFlow). This process is time-consuming, operator-dependent, and limits reproducibility. The objective of this thesis was to automate valve segmentation in 2D PC-MRI, combining classical image processing and deep learning, and to assess integration into clinical software. First, a classical algorithm for flow centre detection was developed by combining temporal standard deviation, local adaptive thresholding, and temporal gradient analysis. Tested on 30 anonymized cases, it achieved 93.3% accuracy, offering a lightweight and interpretable alternative to manual initialization. Second, deep learning models were trained on a multi-centre dataset of 155 cases, including U-Net, Dense U-Net, and UNetR. The best model, a Dense U-Net with 5-frame input, achieved a Dice coefficient of 0.92 on the validation set and was integrated into QFlow through ONNX conversion and C++ implementation, demonstrating clinical feasibility. Finally, a learning curve analysis using 60 annotated cases showed that robust performance (Dice ≈ 0.89) could already be obtained with 20% of the data, with performance plateauing beyond 80%. This highlights efficient annotation strategies and the potential for scalable clinical adoption.