Abstract
Conventional registration methods are known to be mathematically robust and inter- pretable. However, their pairwise iterative optimization process is often computationally intensive. Therefore, they usually require long registration times, especially for complex deformable registration tasks, making them unsuitable for real-time applications such as tumor tracking during radiation therapy.Deep Learning-based registration methods, on the other hand, rely on Deep Neural Networks, which automatically learn a general registration function from the data. Once trained, the model can then be used to register new unseen image pairs in the order of milliseconds.
Hence, this study aimed to leverage the registration speeds of Deep Learning-based methods to improve the registration times of a conventional registration framework implemented for the automated motion analysis of bony joint structures, specifically focusing on the knee.
A thorough overview of the state-of-the-art Deep Learning-based registration frame- works and components was conducted to identify a suitable method for the task of rigid intra-patient knee CT registration. MONAI, the open-source Artificial Intelligence plat- form focused on medical images, was the framework of choice due to its versatility and the availability of adaptable registration networks.
Different models were trained on consecutive and non-consecutive image pairs from dynamic knee CT scans. The corresponding segmentations of the femur, tibia and patella, segmented by a UNet-like segmentation network, were employed to mask the images with a single bone of interest, enabling a piecewise rigid registration approach in which each bone was registered independently.
The capabilities and generalizability of the models trained on femur-masked pairs were tested in different scenarios representing different registration complexities. The Dice Score, False Positive and False Negative volume fractions served as registration evalua- tion metrics, exclusively based on segmentation overlap.
When tested on the task of registering consecutive pairs, the models achieved post- registration Dice scores up to 0.910 ± 0.071 (Mean ± SD) and proved to reduce the conventional registration times from tens of minutes to milliseconds for a single pair.
| Date of Award | Jun 2024 |
|---|---|
| Original language | English |
| Supervisor | Benyameen Keelson (Co-promotor) & Jef Vandemeulebroucke (Promotor) |