After standard surgery for neck hernias, about 25% of patients report low satisfaction. This thesis applied inferential statistics to clinical data and Machine Learning to medical imaging, with the... Show moreAfter standard surgery for neck hernias, about 25% of patients report low satisfaction. This thesis applied inferential statistics to clinical data and Machine Learning to medical imaging, with the goal of finding out where differences in functional outcomes after surgery come from and how artificial intelligence can improve the diagnostic and prognostic process. The initial idea that differences in functional recovery were due to surgical technique was refuted by an RCT from this dissertation. The differences in functional recovery between three different surgical groups (removal of the intervertebral disc without artificial material, placement of intervertebral disc prosthesis, and fusion of vertebrae with a cage) were found to be minimal. It was striking that not surgical technique, but patients' mental health and preoperative, radiological imaging were found to be predictive of clinical recovery after surgery. Although the intervertebral disc prosthesis did not deliver on the promise of preserving mobility and thus could not prevent degeneration at adjacent levels, using Deep Learning based solely on the preoperative MRI of the neck, researchers were able to predict, among other things, which patients would require reoperation after surgery for that adjacent degeneration. The Deep Learning model did that significantly better than an experienced neuroradiologist and neurosurgeon. Such Deep Learning models eliminate the need for time-consuming questionnaires and are thus more cost-effective and less stressful for the patient, while they can be used to identify radiological features important for predicting the postoperative course. After validation with larger radiological datasets, these models can support clinical decision-making and help physicians develop personalized treatment strategies. Challenges within image analysis research for the spine lies in integrating different models into one automated process, preferably built into the electronic health record. Show less
Quantum hardware comes with a different computing paradigm and new ways to tackle applications. Much effort has to be put into understanding how to leverage this technology to give real-world... Show moreQuantum hardware comes with a different computing paradigm and new ways to tackle applications. Much effort has to be put into understanding how to leverage this technology to give real-world advantages in areas of interest for industries such as combinatorial optimization or machine learning. Variational quantum algorithms (VQAs) have been introduced as a way to work within the scope of reach of the current imperfect and unstable hardware. A VQA boils down to a parameterized quantum circuit, which is a quantum circuit with adjustable real-valued parameters. These parameters are generally tweaked by a classical optimization algorithm to optimize a quantity of interest. As VQAs are most often used as heuristics, it is not clear whether they genuinely outperform current classical state-of-the-art algorithms in relevant domains of application. Plus, a VQA can come with many components (which can also be calledhyperparameters) making it a (growing) complex system to analyze performances on many considered tasks. Consequently, we will face the issues of algorithm selection and configuration, which we tackle through this thesis on many examples relevant to industrial applications. In this thesis, VQAs for combinatorial optimization, chemistry/material science, and machine learning problems were considered. Show less
Uncertainty and noise are frequently-encountered obstacles in real-world applications of numerical optimization. The practice of optimization that deals with uncertainties and noise is commonly... Show moreUncertainty and noise are frequently-encountered obstacles in real-world applications of numerical optimization. The practice of optimization that deals with uncertainties and noise is commonly referred to as robust optimization. This thesis concentrates on robust optimization w.r.t the parametric uncertaintiesin the search variables. These parametric uncertainties are assumed to be structurally symmetric, additive in nature, and can be modeled in a deterministic or aprobabilistic fashion. This dissertation empirically studies the models, algorithms, and techniques utilized for surrogate-assisted robust optimization in this context. Based on the studies performed in the dissertation, we conclude that Kriging, SVM, and Polynomial Regression are useful modeling techniques to solve robust optimization problems. We also validate the applicability of Autoencoders and PCA for addressing high-dimensional problems. Lastly, we find that mini-max robustness is the most efficient robustness formulation technique in practical scenarios. Show less
