GP-GOMEA is a state-of-the-art evolutionary algorithm for symbolic regression, known for discovering small and interpretable models. However, its computational cost remains substantial, limiting its applicability to larger datasets and more complex target expressions. In contrast, the rise of modern subsymbolic approaches, particularly deep learning, has been driven largely by the massive parallelism offered by GPUs. This thesis takes the first major step toward a fully GPU-accelerated GP-GOMEA by introducing a GPU-based fitness evaluation scheme. A GPU-friendly representation of GP-GOMEA's template-based individuals is designed alongside a corresponding evaluation strategy that exploits the inherent parallelism of population-based search. This substantially increases evaluation throughput, enabling orders of magnitude more evaluations within the same time budget.

Across four standard symbolic regression benchmarks, this increased evaluation capacity yields performance improvements, particularly for larger datasets and larger population sizes. Additionally, a dynamic-block strategy is introduced that further improves GPU utilization for small batch sizes, and GPU-accelerated evaluation is extended to Modular GP-GOMEA. Moreover, the ability to efficiently evaluate much larger datasets and more complex templates enables analyses that were previously infeasible, allowing a systematic investigation of what makes expressions increasingly difficult for GP-GOMEA and providing new insights into how expression structure affects search difficulty. Finally, for the first time, this expanded capability allows a problem-agnostic evolutionary algorithm to reliably regress one of the largest Feynman equations within four hours. ... Read more

Current inverse materials discovery methods face a trade-off between broad exploration of chemical space and control over chemical validity, synthesisability, and target properties. Here, we present the COMPosition Aware Search Strategy (COMPASS), a constrained multi-objective, multi-fidelity search framework for crystalline composition spaces. COMPASS introduces a discrete mixed-site encoding for material families with fixed site stoichiometries and up to two species mixed on each crystallographic site. This encoding preserves chemical identity, allowing empirical chemical rules and property constraints to be evaluated directly during optimisation. COMPASS combines fast composition-only screening with a constrained genetic algorithm, structure-based verification using machine-learning interatomic potentials, and active learning to improve the low-fidelity model. Applied to mixed-site ABX3 perovskites, COMPASS identifies 15,922 computationally promising candidates satisfying chemical, novelty, stability, and band-gap criteria. In the same constrained discovery task, COMPASS achieves an approximately two-orders-of-magnitude higher yield of desired candidates than the tested open-source MatterGen baselines [Zeni et al., Nature, 2025, 639, 624--632]. These results position COMPASS as a framework for chemically well-understood discovery problems where chemical constraints can guide search through large composition spaces. ... Read more

High-Dose-Rate brachytherapy is a critical component in the treatment of locally advanced cervical cancer. While automated treatment planning systems, such as BRIGHT, have demonstrated the ability to generate high-quality plans, their clinical adoption is hindered by the complexity of their configuration. Deploying such a system in a new hospital requires the manual definition of a clinical protocol that accurately reflects the local institution’s specific standard of care. This "cold start" problem is time-consuming for both doctors and researchers.

This thesis proposes a novel framework for the Automated Discovery of Clinical Protocols. By formulating the protocol configuration as a bi-level optimization problem, we employ the Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm to autonomously extract implicit expert knowledge from a repository of historical clinical plans. The system evolves a set of protocol parameters that, when fed into BRIGHT, reproduce radiation dose distributions as preferred by human experts.

We validate this approach using anonymized patient data from Virginia Commonwealth University. Through a series of experiments with incrementally increasing complexity, ranging from optimizing simple dose thresholds to evolving the definitions of dosimetric metrics, we demonstrate that the proposed framework can successfully identify protocols that generate treatment plans that are quantitatively similar to the clinical ground truth. This research serves as a proof-of-concept, offering a pathway to rapidly deploy automated planning systems while ensuring alignment with local clinical expertise. ... Read more

Artificial Intelligence (AI) is an idea, a set of research subfields, and, ultimately, a suite of technologies that is reshaping the world. AI systems are intended to solve problems that would otherwise require biological or human intelligence to address. Since the middle of the 2010s, breakthroughs in the AI subfield of deep learning enabled rapid progress in computer vision, natural language processing, generative modelling, and other areas.

The key idea of deep learning is to use sufficiently large quantities of data to set parameters of a neural network such that it will perform well on a target task. A neural network is a directed graph of parameterized operations that transform a numeric input into a numeric output. The parameters of a network can be optimized via gradient-based techniques, in contrast to the hyperparameters that are often manually set by an expert. Typical hyperparameter categories are the settings of the gradient-based optimizer, the choice of operations used in the network, and the structure of its computational graph. The latter two are commonly referred to as the architecture of a network, and optimizing them is called Neural Architecture Search (NAS).

Hyperparameters can strongly influence both the performance of a network on the target task, and its efficiency. Therefore, it is important to find good hyperparameter values. The goal of hyperparameter optimization algorithms is to automate this process, which is challenging in the deep learning context for several reasons. Firstly, to evaluate how good a set of hyperparameter values is, a network typically needs to be trained, which takes time and expensive hardware, thus restricting how many sets of hyperparameter values can be evaluated on a finite budget. Secondly, neural networks require many hyperparameters to be set, with each having many potential values that non-trivially interact with those of other hyperparameters, leading to large search spaces that may be difficult to optimize in. Finally, hyperparameter optimization is often multi-objective, that is, involving several conflicting objectives, such as maximizing performance of a network while minimizing its inference time.

Multi-objective problems are commonly addressed via Evolutionary Algorithms (EAs). In an EA, several solutions, called a population, are optimized simultaneously, making them natural candidates to search for sets of solutions that represent different trade-offs in multi-objective problems. Other advantages of EAs are their ability to tackle large search spaces and the ease with which they can be parallelized, which is important for practical usage on modern hardware. Additionally, in order to reduce the inefficiency of EAs in terms of the number of evaluations of the objective functions required to reach convergence, these algorithms can be hybridized with approaches such as Bayesian optimization that can achieve excellent results within a budget of only a few evaluations.

The main goal of this thesis is to explore how EAs can be leveraged to perform hyperparameter optimization for deep learning effectively, so that the resulting networks achieve excellent performance, and efficiently, so that minimal computational effort would be required. ... Read more

Crafting and refining high-dose-rate brachytherapy treatment plans for cervical cancer is a time-consuming process. In recent years, BRIGHT was developed, an AI-based automated treatment planning method that provides not just one, but a set of optimized, patient-specific treatment plans, each with a different trade-off between objectives of interest. BRIGHT's plans are optimized using protocols that define guidelines on the delivered doses. In this thesis, we explore an alternative approach using dose-response models. These models provide insights into the estimated outcomes and risks of a treatment plan. Additionally, they offer great adaptability by including external patient characteristics. Currently, the use of these models remains limited to a feedback role. However, we instead include these models in BRIGHT to optimize them directly. Using one Tumor Control Probability (TCP) model and five Normal-Tissue Complication Probability (NTCP) models, we designed and tested several dose-response objective formulations and optimization techniques. We found that with the current models, these new objectives are insufficient as a replacement for BRIGHT's protocol-based coverage and sparing objectives. The produced plans have greatly improved dose-response outcomes but fall short in protocol compliance. Extending the existing objectives rather than replacing them proved more favorable. Average model improvements around 0.004 for NTCP are observed among the best coverage-sparing plans satisfying the protocol. Additionally, by sacrificing some sparing in the protocol-satisfaction range, improvements around 0.005 are possible for TCP and NTCP. Moreover, these dose-response-focused plans show distinct differences in their dose distribution favoring the dose-response targets compared to regular BRIGHT. Ultimately, the improvements we obtained are only marginal, and the clinical implications of this are unclear. The covariates of the models used in this thesis mostly overlapped with BRIGHT's objectives and did not fully extract their potential. Nevertheless, this thesis proves that the concept is viable and builds a foundation for this technique for when more and better dose-response models become available. ... Read more

Multimodal machine learning models can exploit complementary information from multiple data modalities. MultiFIX (Multimodal Feature engIneering eXplainable artificial intelligence) is a framework designed to construct partially interpretable multimodal models, providing explanations for both modality-specific features and each modality its contribution to the final prediction. However, it was shown to not scale effectively for tasks with extreme joint-modality dependence.

This thesis proposes an alternative training strategy that integrates knowledge of the features to be engineered, expressed as feature targets that guide the learning process. The strategy improves upon baseline performance, even when the feature targets are non-ideal. Since ground-truth feature targets are typically unavailable in real-world settings, the feature targets are optimised using the Gene-pool Optimal Mixing Evolutionary Algorithm. The optimised feature targets, though only loosely aligned with the ground-truth features, enables the alternative training method to surpass baseline MultiFIX performance on a three-gated XOR task.

The same approach was evaluated on simpler tasks, such as the single XOR and AND problems, where it achieved slightly lower but still comparable performance to the already strong baselines. Results indicate that this computationally intensive approach is most beneficial for problems characterised by high joint-modality dependence and complex feature interactions. Interestingly, closer alignment between the optimised and ground-truth feature targets did not consistently lead to higher MultiFIX performance. Consequently, future improvements are likely to stem from refining how feature targets are integrated into the training process, rather than from further optimisation of the targets themselves. ... Read more

This thesis utilizes Evolutionary Algorithms (EAs) within the BRIGHT framework for developing breast cancer brachytherapy treatment plans. We use expert knowledge and state-of-the-art EAs to formulate treatment planning as a multi-objective optimization problem whose solutions can be applied to actual patients. We propose four novel 2- and 3-objective formulations of this problem, which we implement within BRIGHT and empirically validate against anonymized data from 9 real-world patient cases. We demonstrate that all four formulations, under reasonable computational and time budgets, are capable of generating plans that match or exceed the properties of reference treatment plans. To verify the clinical relevance of our contributions, we rely on the expertise of a clinical expert who assesses whether the generated plans can be used in clinical treatment planning. The results of our empirical analysis show that 17 of the 18 plans presented to the expert are clinically acceptable and of immediate value to practitioners within the field of breast brachytherapy today. ... Read more

Cervical cancer affects about half a million women globally every year. The treatment of cervical cancer with the aim of healing mainly consists of surgery, radiation treatment, or a combination of radiation treatment with chemotherapy or hyperthermia. Radiation treatment is a type of treatment wherein a high dose of ionizing radiation is used to kill the tumor cells. The radiation dose is usually delivered in the form of External Beam Radiation Treatment (EBRT) with a linear accelerator followed by internal radiation treatment (brachytherapy) during which a small radioactive source is passed through an applicator and needles that are placed temporarily nearby the cervix. EBRT typically spans several weeks with daily sessions (often referred to as fractions), whereas brachytherapy typically consists of three or four fractions based on one to three implantations. The aim of the radiation treatment is to provide effective radiation to kill the tumor cells while sparing the nearby healthy tissue or Organs At Risk (OARs) as much as possible. This is achieved by treatment planning following the contouring of target volumes and OARs, on medical imaging scans, which typically are Computed Tomography (CT) and/orMagnetic Resonance Imaging (MRI).... ... Read more

Counterfactual explanations are a useful tool to explain trained models. They are based on counterfactual thoughts, which are a natural human thought process that helps us reason about the past. When applied to trained models they show how to make minimal changes to a data point in order to obtain a desired output.
Most methods find these counterfactuals by optimizing a set of objectives. Previously these objectives were often combined into a loss function using an aggregation operator. This operator implicitly decides the priority between the objectives, but this ordering is not always in line with the user’s preferences.
To mitigate this the Multi-Objective Counterfactuals (MOC) method was introduced. MOC turns counterfactual generation into a multi-objective optimization problem and presents the user with a diverse set of counterfactuals that have different trade-offs for the objectives. It optimizes the set of objectives with an evolutionary algorithm called Nondominated Sorting Genetic Algorithm II.
In this thesis we optimize this problem using Multi-Objective Real-Valued Gene-Pool Optimal Mixing Evolutionary Algorithm, which is a different evolutionary algorithm. We present a single-modal method and two multi-modal methods. We compare the performance of our methods to a counterfactual generation method named Diverse Counterfactual Explanations (DiCE), which focusses on feasibility and diversity within a set of generated counterfactuals. Additionally, we also present a visualization tool for sets of counterfactuals.
The single-modal method generates counterfactuals that are realistic, but do not consistently perform well in other areas. The first multi-modal method generates diverse sets of counterfactuals, but overall performs worse. The second multi-modal method generates counterfactuals that perform similarly to the single-modal method, but are more diverse. ... Read more

Machine learning (ML) models are used increasingly in high-stakes areas such as health and finance because of their strong performance. However, having good performance in metrics such as accuracy or the f1 score alone is not all that is important as trust is also essential in these areas. TalkToModel is a system that addresses this challenge, using a large language model (LLM), by letting the users interact with their models through natural language. However, this system only allows a user to ask black-box questions. Another way to achieve trust is through models that are constructed with interpretability in mind that a human can understand. Genetic programming (GP) models are such models that have the potential to be interpreted. This thesis investigates if GP models can be made even more interpretable
using TalkToModel. To do this an enhanced version of TalkToModel called TalkToGP is created with three main contributions: 1) Integration of GP Models into TalkToModel, 2) the ability to ask GP modelspecific questions and 3) the ability to do a comparative analysis between multiple GP models. This system is built using the feedback from GP users who gave insights on their experience with GP as well as their wishes for this system. In the end, the system is evaluated by GP users in an experiment. The experiments showed that the enhanced version of TalkToModel shows a strong indication that it increases the interpretability of GP models. This means the system could be a useful tool for anyone working with GP models. ... Read more

Mixed-integer optimization problems, incorporating both discrete and continuous variables, present unique challenges across various domains such as computer science, finance, logistics, and healthcare. Evolutionary Algorithms (EAs) have emerged as powerful optimization techniques capable of tackling such complex problems in either the discrete or continuous domain. Model-based EAs, integrating machine learning techniques, have further improved the efficiency and scalability of these algorithms. The first algorithm to combine discrete and continuous model-based EAs is the Mixed-Integer Hybrid Evolutionary Algorithm (MIHEA). Despite its potential, MIHEA remains relatively underutilized in research endeavours. This thesis seeks to bridge this gap in research by applying the algorithm to a novel context: structure learning of Bayesian Networks (BNs).

BNs offer a transparent framework for probabilistic reasoning, making them well-suited for various applications. However, learning the structure and discretizations of BNs from data is a challenging task. It is not uncommon for the datasets to contain values of continuous nodes as well. Assuming no normality, these nodes in the data need to be discretized, since BNs are designed for discrete data. The optimal discretizations of these nodes depend on the structure of the BN, meaning that the discretization optimization should happen simultaneously with the structure learning. MIHEA holds promise for addressing this challenge by leveraging its mixed-integer optimization capabilities.

The investigation performed in this thesis starts with a reproduction study. The description of the code for MIHEA is shown to be inconsistent with the experiment results, which prompts a reproduction study resulting in a version of the code that more accurately fits the results.
Subsequently, MIHEA is applied to the structure learning of BNs, where discrete variables represent network structure and continuous variables encode discretizations of continuous nodes. Several solution representations are explored. The experiments show that MIHEA achieves similar or better performance than the state-of-the-art DBN-GOMEA approach on the task of recreating randomly generated BNs from data, at the cost of increased execution time.
These results demonstrate the potential of model-based mixed-integer EAs, particularly MIHEA, for BN structure learning from continuous data. The findings encourage further exploration and utilization of mixed-integer EAs in solving (real-world) problems involving BNs and continuous data.
... Read more

With the advances in General-Purpose computing on Graphics Processing Units (GPGPU), it is worthwhile to explore whether other areas in the field of Artificial Intelligence (AI) can reap the benefits. One such area is Evolutionary Algorithms (EAs), which—among other processes—involves the repetitive exchange of genes among individuals. This repetitive nature aligns with our intuition for parallel optimisation, precisely what GPGPU is designed for. Currently, the state-of-the-art approach in EA is known as Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA), which capitalises on the information embedded within the population by computing the linkage between genes across the entire population. However, when it comes to parallelising the exchange of complete linkage sets, particularly in the context of our specific problem of interest, the challenge becomes more intricate.

In the case of our problem, known as Max-cut, there are dependencies between genes that must be considered when constructing parallel sets of linkage sets, referred to as packages. We propose three solutions: contamination, revision, and association. Contamination fully utilises parallel capabilities but deviates from the concept of linkage sets. Revision constructs the linkage sets as described by GOMEA, but keeps the dependencies between linkage sets within a package untouched. Association on the other hand attempts to resolve the dependencies by generating a dependency graph to create the set of packages.

From our experiments, we can conclude that parallel acceleration using GPGPU is roughly on par with—and sometimes even outperforms—its non-parallelised counterpart. Out of the three solutions, it is evident that association demonstrates the most promising performance profile in terms of approaching the optimal solution. However, the performance falls significantly short of matching the capabilities exhibited by GOMEA. Furthermore, all of the solutions face a significant burden when evaluating the fitness for each exchanged linkage set. An option to consider as an extension to the current setup is known as partial evaluation, although the performance exhibited by contamination implies that simplicity could be the key to success. Further exploration of the acceleration process using widely employed parallel operators—such as those found in linear algebra—has the potential to yield valuable insights for enhancing performance. ... Read more

Deformable Image Registration (DIR) is a medical imaging process involving the spatial alignment of two or more images using a transformation model that can account for non-rigid deformations. B-spline-based transformation models have emerged as a common approach to express such spatial alignments. However, without additional measures, their flexibility can lead to physically implausible deformations. This flexibility has motivated the inclusion of penalty terms to improve the smoothness and regularity of the transformation. Determining an appropriate weight for this penalty term is difficult, as each registration problem requires a different trade-off between this penalty term and the quality of the transformed image.

Gradient-based methods are commonly used as optimization methods in medical image registration toolboxes due to their computational efficiency and fast convergence rates. However, due to their gradient-based approach, they may converge prematurely in local minima. In this thesis, we investigate the efficacy of a gradient-less alternative: the Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm (RV-GOMEA), a population-based method that can exploit the problem structure of optimization problems through explicit mappings of dependencies between problem variables. To improve the computational efficiency of RV-GOMEA when applied to DIR, we show how to apply partial evaluations for common image similarity metrics and penalty terms when using B-spline-based transformation models.

We test RV-GOMEA on a synthetic registration problem to better understand its performance in the context of DIR. Based on our findings, we propose several methods that hybridize RV-GOMEA with a gradient-based method and impose specific constraints on the B-spline-based transformation model. We validate the performance of these methods on clinical registration problems and find that RV-GOMEA with a gradient-based local search operator can provide significant benefits over purely gradient-based methods for DIR problems. Additionally, placing specific constraints on the transformation model can increase the regularity of transformations without requiring a penalty term. ... Read more

The Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm (RV-GOMEA) is a state-of-the-art algorithm for single-objective, real-valued optimization. As many practical applications are inherently constrained, evolutionary algorithms are equipped with constraint handling techniques to allow optimizing constrained problems. The approach currently in use with RV-GOMEA prioritizes solution feasibility over the objective value in all cases, pressuring the algorithm to find feasible solutions. However, this can be inefficient if the constrained optimum is located at the constraint boundary, as search is discouraged from exploring the search space close to infeasible solutions.

In this thesis, several well-known constraint handling techniques from literature are adapted for use with RV-GOMEA and evaluated on different benchmark problems, identifying the strengths and limitations of the various techniques. Furthermore, the inefficiency of the current technique is investigated in detail. Based on the insights gained, modifications to the existing techniques are proposed, leading to promising preliminary results. ... Read more

Even if a Multi-modal Multi-Objective Evolutionary Algorithm (MMOEA) is designed to find all locally optimal approximation sets of a Multi-modal Multi-objective Optimization Problem (MMOP), there is a risk that the found approximation sets are not smoothly navigable because the solutions belong to various niches, which reduces the insight for decision makers. Moreover, when the multi-modality of MMOPs increases, this risk grows and the trackability of finding all locally optimal approximation sets decreases. One example where this issue occurs is that of High-Dose-Rate (HDR) brachytherapy for prostate cancer. In HDR brachytherapy a treatment plan is to be optimized that irradiates a tumour with a prescribed dose, whilst sparing all of the healthy organs surrounding the tumour. The radiation is administered through a radioactive source that is stopped at certain dwell positions within a set of hollow catheters that have been implanted into the patient. In a treatment plan, each of the dwell positions is given a specific dwell time for which the source is kept at that location in order to irradiate the surrounding tissue.
To tackle the navigability issues, two new MMOEAs are proposed: Multi-Modal Bézier Evolutionary Algorithm (MM-BezEA) and Set Bézier Evolutionary Algorithm (Set-BezEA). Both MMOEAs produce approximation sets that cover individual niches and exhibit inherent decision-space smoothness as they are parameterized by Bézier curves. MM-BezEA combines the concepts behind the recently introduced BezEA and MO-HillVallEA to find all locally optimal approximation sets. Set-BezEA employs a novel multi-objective fitness function formulation to find limited numbers of diverse, locally optimal, approximation sets for MMOPs of high multi-modality.
Both algorithms, but especially MM-BezEA, are found to outperform the MMOEAs MO_Ring_PSO_SCD and MO-HillVallEA on MMOPs of moderate multi-modality with linear Pareto sets. Moreover, for MMOPs of high multi-modality, Set-BezEA is found to indeed be able to produce high-quality approximation sets, each pertaining to a single niche. Set-BezEA is also shown to be comparable to the current BRIGHT approach used in the Amsterdam UMC for the optimization of treatment plans for prostate cancer HDR brachytherapy, which opens the way for it to be introduced in the clinical practice in the future. ... Read more

Neural networks (NNs) have, in recent years, become a major part of modern pattern recognition, and both theoretical and applied research evolve at an astounding pace. NNs are usually trained via gradient descent (GD), but research has shown that GD is not always capable of training very small networks. As a result, networks trained via GD are often significantly larger than necessary, demanding more computing power and energy to evaluate, and thereby hampering adoption on lower- power devices. This thesis investigates whether evolutionary algorithms (EAs) can successfully train NNs and if these networks can be smaller than those required by GD. Four algorithms, namely the GD-based Adam, Adam with cold restarts, and the EAs GOMEA and BIPOP-CMA-ES, were used to train various configurations of multilayer perceptrons (MLPs) with one hidden layer for the Exclusive-OR (XOR) problem. Their relative performance was gauged by comparing the rates at which each algorithm attained an acceptable loss for a given number of hidden nodes. The main findings are that EAs could find smaller ReLU-activated networks than GD could, while, oppositely, the GD could generally find smaller networks for sigmoid- activated networks. However, GOMEA in particular was able to successfully train XOR networks using the highly discrete Heaviside activation function, whereas GD could not due gradient erasure. Problem-specific knowledge in the form of a soft symmetry breaking constraint was found to be effective to increase success rates in a limited number of cases. These findings indicate that using EAs is a viable strategy for training NNs, yielding smaller, and therefore more efficient networks than those trained via GD, provided that the network topology is suitable for the chosen EA. This opens up future research into, including but not limited to, the many ways EAs can be scaled up to training larger NNs, hybridization with GD, and niche network topologies. ... Read more

Solutions to many real-life optimization problems take a long time to evaluate. This limits the number of solutions we can evaluate. When optimizing with an Evolutionary Algorithm (EA) a frequently used approach is to approximate the objective using a surrogate function, replacing the time-consuming real evaluation. This surrogate model is combined with a so-called acquisition function, to select promising candidate solutions. The acquisition function balances the trade-off between exploration of parameter space and the exploitation of the surrogate. These candidates are subject to an expensive evaluation with the true objective function and are used to update the surrogate model. Iteratively applying this process can effectively optimize global optimization problems. In this work, we propose a new multi-objective optimization algorithm with inverse distance weighting as surrogate function, which we call IDW-SAEA (inverse distance weighting surrogate assisted evolutionary algorithm). We introduce a new objective to the optimization problem to improve exploration and reduce the complexity of the acquisition function. We show this algorithm is competitive with state-of-the-art kriging-based surrogate-assisted EAs on certain benchmark problems. Additionally, we use the algorithm to optimize a practical problem: a Finite Element Method simulation of the cervix region with applications in radiotherapy for cervical cancer. ... Read more

Deep Neural Networks (DNNs) have the potential to make various clinical procedures more time-efficient by automating medical image segmentation; largely due to their strong, in some cases human-level, performance. The design of the best possible medical image segmentation DNN, however, is task-specific. Neural Architecture Search (NAS), i.e., the automation of neural network design, has been shown to have the capability to outperform manually designed networks. However, the existing NAS methods for medical image segmentation have explored a quite limited range of types of DNN architectures that can be discovered, and, more importantly, do not evaluate the accuracy of performance estimation methods.

In this thesis, various performance estimation methods for DNNs are analysed for medical image segmentation tasks. Due to the use of different metrics, small datasets, and inter-physician variability, DNN performance values are susceptible to considerable noise. Through experiments on multiple datasets, it is shown that performance estimation needs to be more elaborate than proposed in previous literature on NAS for medical image segmentation. Only then can the noise induced by the problem be overcome. Based on evaluations of NAS performance with different levels of noise, a method is put forward to evaluate this noise, such that a more informed decision on performance estimation can be made.

The second contribution of this thesis, is the proposal of a novel NAS search space for medical image segmentation networks. This search space combines the strength of a generalised encoder-decoder structure, well known from U-Net, with network blocks that have been proven to have a strong performance in image classification tasks. The search is performed by looking for the best topology of the network, and simultaneously searching the configuration of each cell. This allows for interactions between topology- and cell-level attributes to be found. Experiments were performed on two publicly available datasets. The networks discovered by the proposed NAS method perform better than well-known handcrafted segmentation networks, and outperform networks found with other NAS approaches that perform only topology search, and topology-level search followed by cell-level search.

Finally, three search algorithms are compared for different performance estimation methods on a realistic clinical medical image segmentation task. The results show that the performance of these algorithms is very similar in noisy environments for initial runs, and show deterioration of performance for all algorithms when correlation with the validation performance values are low. This supports the findings that not adapting performance estimation to the task at hand will lead to poor NAS performance, no matter the chosen search algorithm. ... Read more