Abstract

Recent advances in 3D scanning technology have led to a widespread use of 3D shapes in a multitude
of fields, including computer vision and medical imaging. These shapes are, however, often
contaminated by noise, which needs to be removed or attenuated in order to ensure high-quality
3D shapes for subsequent use in downstream tasks. On the other hand, the availability of largescale
pediatric hand radiographs and brain imaging benchmarks has sparked a surge of interest
in designing efficient techniques for bone age assessment and brain disease prediction, which are
fundamental problems in computer-aided diagnosis. Bone age is an effective metric for assessing
the skeletal and biological maturity of children, while understanding how the brain develops is
crucial for designing prediction models for the classification of brain disorders.
In this thesis, we present a feature-preserving framework for carpal bone surface denoising in the
graph signal processing setting. The proposed denoising framework is formulated as a constrained
optimization problem with an objective function comprised of a fidelity term specified by a noise
model and a regularization term associated with data prior. We show through experimental results
that our approach can remove noise effectively while preserving the nonlinear features of surfaces,
such as curved surface regions and fine details. Moreover, recovering high quality surfaces from
noisy carpal bone surfaces is of paramount importance to the diagnosis of wrist pathologies, such
as arthritis and carpal tunnel syndrome. We also introduce a deep learning approach to pediatric
bone age assessment using instance segmentation and ridge regression. This approach is comprised
of two intertwined stages. In the first stage, we employ an image annotation and instance
segmentation model to extract and separate different regions of interests in an image. In the second
stage, we leverage the power of transfer learning by designing a deep neural network with
a ridge regression output layer. For the classification of brain disorders, we propose an aggregator
normalization graph convolutional network by exploiting aggregation in graph sampling, skip
connections and identity mapping. We also integrate both imaging and non-imaging features into
the graph nodes and edges, respectively, with the aim of augmenting predictive capabilities. We
validate our proposed approaches through extensive experiments on various benchmark datasets,
demonstrating competitive performance in comparison with strong baseline methods.