Biomedical Image Analysis using Deep Learning: Towards Clinically Applicable Models.

Applied artificial intelligence has a huge potential to transform the way biomedical research and healthcare practice are conducted. Computer vision analysis powered by Deep Learning (DL) has already proven promising for understanding biomedical images. From microscopy to brain imaging, current DL-based solutions are more accurate, faster and easier to develop and deploy than traditional solutions. However, its application to the clinic is not ready yet. Many models present poor generalisation, lack explainability, or suffer from algorithmic biases, which may lead to overconfident results that could lead to life-threatening consequences.
In this context, the present thesis explores the current translation of Machine Learning (ML) solutions to real-world settings in biomedical research and clinical practice with a special emphasis on DL solutions for biomedical imaging. The aim of this thesis is threefold. First, to study the state-of-the-art of computer vision solutions for biomedical tasks together with their performance with noisy labelled data and capabilities to identify biomarkers of Parkinson’s Disease. Second, to investigate the factors that impact the models’ accuracy and their phenomenological fidelity. This includes understanding the effects of confounded datasets in medical imaging tasks and the overall suitability of datasets in ML tasks. Finally, to provide good practices and tools for robust DL solutions in biomedical applications.
In conclusion, this thesis offers insights to develop more robust ML solutions for biomedical and healthcare settings through a multidisciplinary approach that combines cutting-edge technologies with scientific methodologies. We hope the present thesis brings new knowledge to the area and provides new opportunities for future researchers.