Anatomical Landmark Localisation and Uncertainty Estimation

Machine learning promises transformative applications in medical image analysis. However, the black-box nature of Deep Neural Networks and data sensitivity issues hinders their clinical deployment. Addressing these challenges necessitates the development of lightweight models suitable for local deployment, accompanied by improved methods for uncertainty estimation of model predictions. Such uncertainty estimation methods could flag potentially erroneous predictions for a human-in-the-loop to review. In this thesis, we tackle these challenges, specifically focusing on the task of landmark localisation, a supervised task that involves identifying precise coordinates of anatomical structures within medical images.

Our first approach introduces PHD-Net, a lightweight, patch-based landmark localisation model that estimates prediction uncertainty heuristically. We experimentally show our approach performs exceptionally given its size and scales well with model capacity, offering an alternative perspective to landmark localisation with a unique uncertainty estimation property. Building on this foundational concept of uncertainty, we broaden its applicability to a wider range of landmark localisation models through the introduction of the Frequentist-inspired Quantile Binning framework. Our approach is general, applicable to any regression problem. Recognising the limitations of relying solely on localisation accuracy to holistically evaluate our models, we introduce evaluation metrics specifically designed for assessing binning-based uncertainty measures, enabling better model uncertainty estimation benchmarking. In our final work, we present the first application of Gaussian Processes to anatomical landmark localisation, achieving genuine Bayesian uncertainty.

Underpinning the impact of our research is a commitment to open-source accessibility. All our tools and innovations are made publicly available on Github within the low-code/no-code framework of MediMarker, or the PyKale library.