Many clinically relevant knee disorders are functional rather than purely static. Yet routine assessment
and even advanced image analysis often rely on incomplete information from, static images, or
partially observed motion. In contrast, 4D CT data can capture the dynamic behaviour of the femur,
tibia and patella over time, offering a rich basis for studying joint kinematics.
A central challenge addressed in ELEVATE is how to enhance sparse, incomplete, or lower-dimensional
musculoskeletal data into richer subject-specific 3D and 4D representations that remain anatomically
plausible and clinically meaningful. In the context of dynamic knee imaging, one concrete instance of
this broader challenge is motion completion: given only a sparse subset of temporal observations, can
the missing motion states be reconstructed in a plausible and quantitatively accurate way? Addressing
this problem could support downstream applications such as motion analysis, dynamic modelling,
temporal super-resolution, and more acquisition-efficient imaging workflows in which only a limited
number of frames is available.
This thesis will focus on existing retrospective knee datasets available within ELEVATE. The student
will work with already available 4D CT sequences and derived kinematic descriptors, and will
investigate whether deep learning methods can reconstruct a full knee motion trajectory from sparse
temporal observations more accurately and plausibly than conventional interpolation approaches.

The goal of this thesis is to develop and evaluate a deep learning framework for sparse-to-dense
completion of knee motion from retrospective 4D CT-derived kinematics. The work will investigate
how accurately missing temporal states can be reconstructed from a limited number of observed
frames, how performance changes as a function of the number and placement of those frames, and
whether learning-based methods outperform classical baselines such as linear or spline interpolation.

The developments will be performed using existing retrospective knee data and associated
segmentations or registrations available within the project. The student will derive compact motion
representations from the data, such as rigid bone transformations, relative joint kinematics, or
related low-dimensional descriptors, and use these as input to a motion completion framework. The
emphasis of the thesis is on the research component: defining the completion problem, comparing
modelling strategies, quantifying the effect of temporal sparsity, and analysing the plausibility and
accuracy of the reconstructed motion.
Given the nature of the project, this thesis will expose the student to a broad set of tasks and
competences, covering both hardware/ sensor exploration, experimental work and algorithm
development and testing. The following tasks can be distinguished:
• Literature study on dynamic musculoskeletal imaging, temporal completion, motion
representation, and deep learning for sequence modelling.
• Preparation of a retrospective knee dataset subset and extraction of compact
kinematic representations from existing 4D CT data
• Design of sparse-observation experiments by masking sequences to retain only
selected temporal frames
• Implementation and training of deep learning models for motion completion, with
comparison against interpolation
• Validation using kinematic error metrics, temporal consistency measures, and
plausibility analysis of reconstructed femur-tibia-patella motion
• Implementation: Python, PyTorch, MONAI/simpleITK, and standard scientific
computing tools