This thesis investigates CNN-based detection of active Tuberculosis (aTB) from chest X-rays using the TBX11K dataset, which includes ground-truth bounding boxes. It shows that adding more annotated data improves model performance and proposes a novel evaluation metric—ROAD-Normalised PropEng Average—to compare visual explanation methods, identifying DenseNet-121 with Eigen-CAM as the most faithful and accurate combination.

This thesis develops stateless methods for sleep-phase detection from polysomnographs (PSG), while exploring techniques to improve cross-database generalisation.

In this thesis, we developed a technique to learn vessel segmentation from retinal fundus images using weakly-supervised methods.

In this thesis, we sought at developing precise neuron segmentation in order to support future study on the morphological phenotype of amyotrophic lateral sclerosis.

Performance measures are an important component of machine learning algorithms. They are useful when it comes to evaluate the quality of a model, but also to help the algorithm improve itself. When used in small data sets, these measures may not properly express the performance of the model. That is when confidence intervals and credible regions can be useful. Expressing the performance measures in a probabilistic setting allows one to develop them as distributions. One can then use those distributions to establish credible regions.

In this study, we investigate the benefits of automatic Pulmonary Tuberculosis (PTB) detection methods based on radiological signs found on frontal chest X-Ray images.

The goal of this study is to identify whether it is possible to predict the occurrence of adverse events in patients based on their clinical data.