Accurately quantifying in-vivo knee kinematics is essential for understanding the biomechanical factors contributing to the development and progression of osteoarthritis. While optical motion capture systems are widely used for this purpose, their accuracy is limited by soft tissue artefacts, particularly in overweight individuals, a demographic at increased risk for knee osteoarthritis. Fluoroscopic imaging offers a more direct measurement of bone motion, but requires reliable 2D–3D registration techniques to reconstruct the 6 degree of freedom joint kinematics. Most existing 2D–3D registration workflows rely on CT-based bone models and intensity-based similarity measures. These are impractical for clinical studies due to radiation exposure concerns. MRI-based bone models are considered an alternative but the suitability of intensity-based methods for MRI-based models is limited. Because of this, alternate MRI-based registration methods need to be considered. This thesis explores the feasability of a registration workflow that registers MRI-based models to segmented fluoroscopic images of the knee. Three feature-based similarity measures, the edge matching Score (EMS), shape matching (SM) and normalized edge distance (NED), were implemented. Ten fluoroscopic images of the same healthy participant were registered with the three developed registration pipelines. Validation was performed by comparing the six degree-of-freedom pose to the pose obtained from a manually registered reference standard. The results show that the accuracy of the three similarity measures is comparable, but the shape matching offered superior computational efficiency. Registration was about twice as fast as the other similarity measures with an average registration time of 104.6 seconds per frame. The mean absolute error (MAE) was less than 0.45 mm for translation parallel to the fluoroscope image plane. The translation perpendicular to the image plane remained a significant source of error, with a MAE of less than 4.56 mm. Similarly, the MAE for rotation about the axis perpendicular to the image plane was less than 0.52◦. Rotation about the other two axes was less accurate with a MAE of less than 1.84◦. This study demonstrates the feasibility of using MRI-based bone models and segmented fluoroscopic images for 2D–3D registration of the knee in-vivo. While limitations remain. Particularly in achieving sub millimeter accuracy for the translation perpendicular to the image plane, and sub-degree accuracy for rotation about the axes that are parallel to the image plane, and generalizing to lower-quality images. Despite this, the results are broadly in line with existing literature. Future work should focus on performing more rigorous validation against a more accurate silver- or even gold standard, testing the registration pipeline on fluoroscopic images that are a better representation of clinical use and developing an automated segmentation algorithm to replace the manual segmentation. Additionally, integrating calibrated optical motion capture data to provide an initial pose estimate during optimization could further improve accuracy and reduce the need for human intervention

Image registration combines data from multiple imaging modalities or time points by aligning images. Existing 2D-3D registration methods often require manual intervention, introducing variability and increasing processing time, and rely on Computed Tomography (CT) imaging, which exposes patients to radiation. This study aims to develop an automated registration pipeline that ensures accurate registration between 3D Magnetic Resonance Imaging (MRI) and 2D fluoroscopic images by preprocessing techniques.

This study involved two registration experiments to evaluate the pipeline. The DiffDRR package was used with normalized cross-correlation (NCC) as similarity metric and gradient-based Adam optimizer for optimization. The first experiment validated the accuracy of registration between 3D MRI models (solid bone, empty shell, and filled shell) and 2D CT-derived digitally reconstructed radiographs (DRRs) with known rotations and translations as reference standard. The second experiment aimed to determine the most effective fluoroscopy preprocessing technique for aligning 3D MRI models with 2D fluoroscopic images. Preprocessing techniques for the fluoroscopic images included contrast enhancement via thresholding and normalization, and region of interest selection or segmentation.

Validation results showed smaller rotational and translational errors for both shell models compared to the solid bone model. In-plane errors for the shell models were 0.1-3.3mm and out-of-plane 15.8-55.0mm, while solid bone reported in-plane errors of 1.0-4.7mm and out-of-plane of 68.3-282.7mm. Registration between empty shell models and differently preprocessed fluoroscopic images showed that segmentation of the bones in fluoroscopic images yielded the highest NCC values. However, visual assessment revealed inconsistencies in plausible bone positioning.

This automated 2D-3D registration method shows potential for registration between 3D MRI and 2D fluoroscopic images of the knee joint. The current method requires further refinement, particularly in integrating soft tissue data, possibly via synthetic CTs, and adopting more anatomically sensitive metrics, such as mutual information or gradient correlation.

Recent years have shown a tremendous increase in the application of Artificial Intelligence to the field of radiology, often through the extraction and analysis of large numbers of quantitative features from medical images. These applications increase the demand for machine learning models to extract information from these images. To provide these models, improve their performance and reduce the time that experts have to spend on manually tuning them, the field of Automated Machine Learning (AutoML) aims to automate the design process of machine learning models by optimizing the selection of algorithms and their hyperparameters for each application. This work applies an AutoML approach to medical image classification, using a Bayesian optimization strategy to automatically optimize the selection of preprocessing and classification algorithms and their hyperparameters. Its performance is compared with the performance of a random search optimization strategy, evaluated on three datasets from three different clinical applications. The results show that the Bayesian optimization and the random search return models that achieve similar performance on the unseen test sets. We show that a random search with relatively few evaluations and a simple ensemble strategy is sufficient to achieve performance comparable to a more sophisticated and more computationally demanding Bayesian optimization approach, therefore validating the use of a random search optimization strategy in this medical image classification setting. All found models generalize poorly, with average F1-scores on the validation sets used for optimizing the models being at least 20\% lower than the average F1-scores on the unseen test sets. Finally, we further emphasize the difficulty to generalize in this setting, by showing that the differences between subsets of the evaluated datasets are large and that increasing the computation time of the optimization does not benefit the test set performance of the final solution.