Peter A. N. Bosman

Alexander Chebykin, Tanja Alderliesten, Peter A. N. Bosman

To achieve excellent performance with modern neural networks, having the right network architecture is important. Neural Architecture Search (NAS) concerns the automatic discovery of task-specific network architectures. Modern NAS approaches leverage supernetworks whose subnetworks encode candidate neural network architectures. These subnetworks can be trained simultaneously, removing the need to train each network from scratch, thereby increasing the efficiency of NAS. A recent method called Neural Architecture Transfer (NAT) further improves the efficiency of NAS for computer vision tasks by using a multi-objective evolutionary algorithm to find high-quality subnetworks of a supernetwork pretrained on ImageNet. Building upon NAT, we introduce ENCAS - Evolutionary Neural Cascade Search. ENCAS can be used to search over multiple pretrained supernetworks to achieve a trade-off front of cascades of different neural network architectures, maximizing accuracy while minimizing FLOPs count. We test ENCAS on common computer vision benchmarks (CIFAR-10, CIFAR-100, ImageNet) and achieve Pareto dominance over previous state-of-the-art NAS models up to 1.5 GFLOPs. Additionally, applying ENCAS to a pool of 518 publicly available ImageNet classifiers leads to Pareto dominance in all computation regimes and to increasing the maximum accuracy from 88.6% to 89.0%, accompanied by an 18\% decrease in computation effort from 362 to 296 GFLOPs. Our code is available at https://github.com/AwesomeLemon/ENCAS

Alexander Chebykin, Tanja Alderliesten, Peter A. N. Bosman

Hyperparameter Optimization (HPO) can lift the burden of tuning hyperparameters (HPs) of neural networks. HPO algorithms from the Population Based Training (PBT) family are efficient thanks to dynamically adjusting HPs every few steps of the weight optimization. Recent results indicate that the number of steps between HP updates is an important meta-HP of all PBT variants that can substantially affect their performance. Yet, no method or intuition is available for efficiently setting its value. We introduce Iterated Population Based Training (IPBT), a novel PBT variant that automatically adjusts this HP via restarts that reuse weight information in a task-agnostic way and leverage time-varying Bayesian optimization to reinitialize HPs. Evaluation on 8 image classification and reinforcement learning tasks shows that, on average, our algorithm matches or outperforms 5 previous PBT variants and other HPO algorithms (random search, ASHA, SMAC3), without requiring a budget increase or any changes to its HPs. The source code is available at https://github.com/AwesomeLemon/IPBT.

Thomas Uriot, Marco Virgolin, Tanja Alderliesten et al.

Dimensionality reduction (DR) is an important technique for data exploration and knowledge discovery. However, most of the main DR methods are either linear (e.g., PCA), do not provide an explicit mapping between the original data and its lower-dimensional representation (e.g., MDS, t-SNE, isomap), or produce mappings that cannot be easily interpreted (e.g., kernel PCA, neural-based autoencoder). Recently, genetic programming (GP) has been used to evolve interpretable DR mappings in the form of symbolic expressions. There exists a number of ways in which GP can be used to this end and no study exists that performs a comparison. In this paper, we fill this gap by comparing existing GP methods as well as devising new ones. We evaluate our methods on several benchmark datasets based on predictive accuracy and on how well the original features can be reconstructed using the lower-dimensional representation only. Finally, we qualitatively assess the resulting expressions and their complexity. We find that various GP methods can be competitive with state-of-the-art DR algorithms and that they have the potential to produce interpretable DR mappings.

Georgios Andreadis, Peter A. N. Bosman, Tanja Alderliesten

Finding a realistic deformation that transforms one image into another, in case large deformations are required, is considered a key challenge in medical image analysis. Having a proper image registration approach to achieve this could unleash a number of applications requiring information to be transferred between images. Clinical adoption is currently hampered by many existing methods requiring extensive configuration effort before each use, or not being able to (realistically) capture large deformations. A recent multi-objective approach that uses the Multi-Objective Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm (MO-RV-GOMEA) and a dual-dynamic mesh transformation model has shown promise, exposing the trade-offs inherent to image registration problems and modeling large deformations in 2D. This work builds on this promise and introduces MOREA: the first evolutionary algorithm-based multi-objective approach to deformable registration of 3D images capable of tackling large deformations. MOREA includes a 3D biomechanical mesh model for physical plausibility and is fully GPU-accelerated. We compare MOREA to two state-of-the-art approaches on abdominal CT scans of 4 cervical cancer patients, with the latter two approaches configured for the best results per patient. Without requiring per-patient configuration, MOREA significantly outperforms these approaches on 3 of the 4 patients that represent the most difficult cases.

Arkadiy Dushatskiy, Alexander Chebykin, Tanja Alderliesten et al.

Population Based Training (PBT) is an efficient hyperparameter optimization algorithm. PBT is a single-objective algorithm, but many real-world hyperparameter optimization problems involve two or more conflicting objectives. In this work, we therefore introduce a multi-objective version of PBT, MO-PBT. Our experiments on diverse multi-objective hyperparameter optimization problems (Precision/Recall, Accuracy/Fairness, Accuracy/Adversarial Robustness) show that MO-PBT outperforms random search, single-objective PBT, and the state-of-the-art multi-objective hyperparameter optimization algorithm MO-ASHA.

Monika Grewal, Dustin van Weersel, Henrike Westerveld et al.

Deep learning models benefit from training with a large dataset (labeled or unlabeled). Following this motivation, we present an approach to learn a deep learning model for the automatic segmentation of Organs at Risk (OARs) in cervical cancer radiation treatment from a large clinically available dataset of Computed Tomography (CT) scans containing data inhomogeneity, label noise, and missing annotations. We employ simple heuristics for automatic data cleaning to minimize data inhomogeneity and label noise. Further, we develop a semi-supervised learning approach utilizing a teacher-student setup, annotation imputation, and uncertainty-guided training to learn in presence of missing annotations. Our experimental results show that learning from a large dataset with our approach yields a significant improvement in the test performance despite missing annotations in the data. Further, the contours generated from the segmentation masks predicted by our model are found to be equally clinically acceptable as manually generated contours.

Vangelis Kostoulas, Peter A. N. Bosman, Tanja Alderliesten

Segmentation of regions of interest in images of patients, is a crucial step in many medical procedures. Deep neural networks have proven to be particularly adept at this task. However, a key question is what type of deep neural network to choose, and whether making a certain choice makes a difference. In this work, we will answer this question for the task of segmentation of the Organs At Risk (OARs) in radiation treatment of cervical cancer (i.e., bladder, bowel, rectum, sigmoid) in Magnetic Resonance Imaging (MRI) scans. We compare several state-of-the-art models belonging to different architecture categories, as well as a few new models that combine aspects of several state-of-the-art models, to see if the results one gets are markedly different. We visualize model predictions, create all possible ensembles of models by averaging their output probabilities, and calculate the Dice Coefficient between predictions of models, in order to understand the differences between them and the potential of possible combinations. The results show that small improvements in metrics can be achieved by advancing and merging architectures, but the predictions of the models are quite similar (most models achieve on average more than 0.8 Dice Coefficient when compared to the outputs of other models). However, the results from the ensemble experiments indicate that the best results are obtained when the best performing models from every category of the architectures are combined.

Jasper Post, Johannes Koch, Anton Bouter et al.

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. In this work, we take the first major step toward a fully GPU-accelerated GP-GOMEA by introducing a GPU-based fitness evaluation scheme. We design a GPU-friendly representation of GP-GOMEA's template-based individuals and 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. Moreover, the ability to efficiently evaluate much larger datasets and more complex templates enables analyses that were previously infeasible, allowing us to systematically analyze what makes expressions increasingly difficult for GP-GOMEA, 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.

Alexander Chebykin, Arkadiy Dushatskiy, Tanja Alderliesten et al.

In this work, we show that simultaneously training and mixing neural networks is a promising way to conduct Neural Architecture Search (NAS). For hyperparameter optimization, reusing the partially trained weights allows for efficient search, as was previously demonstrated by the Population Based Training (PBT) algorithm. We propose PBT-NAS, an adaptation of PBT to NAS where architectures are improved during training by replacing poorly-performing networks in a population with the result of mixing well-performing ones and inheriting the weights using the shrink-perturb technique. After PBT-NAS terminates, the created networks can be directly used without retraining. PBT-NAS is highly parallelizable and effective: on challenging tasks (image generation and reinforcement learning) PBT-NAS achieves superior performance compared to baselines (random search and mutation-based PBT).

Arthur Guijt, Dirk Thierens, Ellen Kerkhof et al.

Deep learning has been shown to be very capable at performing many real-world tasks. However, this performance is often dependent on the presence of large and varied datasets. In some settings, like in the medical domain, data is often fragmented across parties, and cannot be readily shared. While federated learning addresses this situation, it is a solution that requires synchronicity of parties training a single model together, exchanging information about model weights. We investigate how asynchronous collaboration, where only already trained models are shared (e.g. as part of a publication), affects performance, and propose to use stitching as a method for combining models. Through taking a multi-objective perspective, where performance on each parties' data is viewed independently, we find that training solely on a single parties' data results in similar performance when merging with another parties' data, when considering performance on that single parties' data, while performance on other parties' data is notably worse. Moreover, while an ensemble of such individually trained networks generalizes better, performance on each parties' own dataset suffers. We find that combining intermediate representations in individually trained models with a well placed pair of stitching layers allows this performance to recover to a competitive degree while maintaining improved generalization, showing that asynchronous collaboration can yield competitive results.

Alexander Chebykin, Peter A. N. Bosman, Tanja Alderliesten

Sharing synthetic medical images is a promising alternative to sharing real images that can improve patient privacy and data security. To get good results, existing methods for medical image synthesis must be manually adjusted when they are applied to unseen data. To remove this manual burden, we introduce a Hyperparameter-Free distributed learning method for automatic medical image Synthesis, Sharing, and Segmentation called HyFree-S3. For three diverse segmentation settings (pelvic MRIs, lung X-rays, polyp photos), the use of HyFree-S3 results in improved performance over training only with site-specific data (in the majority of cases). The hyperparameter-free nature of the method should make data synthesis and sharing easier, potentially leading to an increase in the quantity of available data and consequently the quality of the models trained that may ultimately be applied in the clinic. Our code is available at https://github.com/AwesomeLemon/HyFree-S3

Monika Grewal, Jan Wiersma, Henrike Westerveld et al.

Purpose: Deformable Image Registration (DIR) can benefit from additional guidance using corresponding landmarks in the images. However, the benefits thereof are largely understudied, especially due to the lack of automatic landmark detection methods for three-dimensional (3D) medical images. Approach: We present a Deep Convolutional Neural Network (DCNN), called DCNN-Match, that learns to predict landmark correspondences in 3D images in a self-supervised manner. We trained DCNN-Match on pairs of Computed Tomography (CT) scans containing simulated deformations. We explored five variants of DCNN-Match that use different loss functions and assessed their effect on the spatial density of predicted landmarks and the associated matching errors. We also tested DCNN-Match variants in combination with the open-source registration software Elastix to assess the impact of predicted landmarks in providing additional guidance to DIR. Results: We tested our approach on lower-abdominal CT scans from cervical cancer patients: 121 pairs containing simulated deformations and 11 pairs demonstrating clinical deformations. The results showed significant improvement in DIR performance when landmark correspondences predicted by DCNN-Match were used in the case of simulated (p = $0e^0$) as well as clinical deformations (p = 0.030). We also observed that the spatial density of the automatic landmarks with respect to the underlying deformation affect the extent of improvement in DIR. Finally, DCNN-Match was found to generalize to Magnetic Resonance Imaging (MRI) scans without requiring retraining, indicating easy applicability to other datasets. Conclusions: DCNN-Match learns to predict landmark correspondences in 3D medical images in a self-supervised manner, which can improve DIR performance.

Arkadiy Dushatskiy, Adriënne M. Mendrik, Peter A. N. Bosman et al.

There has recently been great progress in automatic segmentation of medical images with deep learning algorithms. In most works observer variation is acknowledged to be a problem as it makes training data heterogeneous but so far no attempts have been made to explicitly capture this variation. Here, we propose an approach capable of mimicking different styles of segmentation, which potentially can improve quality and clinical acceptance of automatic segmentation methods. In this work, instead of training one neural network on all available data, we train several neural networks on subgroups of data belonging to different segmentation variations separately. Because a priori it may be unclear what styles of segmentation exist in the data and because different styles do not necessarily map one-on-one to different observers, the subgroups should be automatically determined. We achieve this by searching for the best data partition with a genetic algorithm. Therefore, each network can learn a specific style of segmentation from grouped training data. We provide proof of principle results for open-sourced prostate segmentation MRI data with simulated observer variations. Our approach provides an improvement of up to 23% (depending on simulated variations) in terms of Dice and surface Dice coefficients compared to one network trained on all data.

Monika Grewal, Henrike Westerveld, Peter A. N. Bosman et al.

Deformable image registration (DIR) involves optimization of multiple conflicting objectives, however, not many existing DIR algorithms are multi-objective (MO). Further, while there has been progress in the design of deep learning algorithms for DIR, there is no work in the direction of MO DIR using deep learning. In this paper, we fill this gap by combining a recently proposed approach for MO training of neural networks with a well-known deep neural network for DIR and create a deep learning based MO DIR approach. We evaluate the proposed approach for DIR of pelvic magnetic resonance imaging (MRI) scans. We experimentally demonstrate that the proposed MO DIR approach -- providing multiple registration outputs for each patient that each correspond to a different trade-off between the objectives -- has additional desirable properties from a clinical use point-of-view as compared to providing a single DIR output. The experiments also show that the proposed MO DIR approach provides a better spread of DIR outputs across the entire trade-off front than simply training multiple neural networks with weights for each objective sampled from a grid of possible values.

Mafalda Malafaia, Thalea Schlender, Peter A. N. Bosman et al.

In the health domain, decisions are often based on different data modalities. Thus, when creating prediction models, multimodal fusion approaches that can extract and combine relevant features from different data modalities, can be highly beneficial. Furthermore, it is important to understand how each modality impacts the final prediction, especially in high-stake domains, so that these models can be used in a trustworthy and responsible manner. We propose MultiFIX: a new interpretability-focused multimodal data fusion pipeline that explicitly induces separate features from different data types that can subsequently be combined to make a final prediction. An end-to-end deep learning architecture is used to train a predictive model and extract representative features of each modality. Each part of the model is then explained using explainable artificial intelligence techniques. Attention maps are used to highlight important regions in image inputs. Inherently interpretable symbolic expressions, learned with GP-GOMEA, are used to describe the contribution of tabular inputs. The fusion of the extracted features to predict the target label is also replaced by a symbolic expression, learned with GP-GOMEA. Results on synthetic problems demonstrate the strengths and limitations of MultiFIX. Lastly, we apply MultiFIX to a publicly available dataset for the detection of malignant skin lesions.

Thalea Schlender, Catharina J. A. Romme, Yvette M. van der Linden et al.

Survival analysis is central to clinical research, informing patient prognoses, guiding treatment decisions, and optimising resource allocation. Accurate time-to-event predictions not only improve quality of life but also reveal risk factors that shape clinical practice. For these models to be relevant in healthcare, interpretability is critical: predictions must be traceable to patient-specific characteristics, and risk factors should be identifiable to generate actionable insights for both clinicians and researchers. Traditional survival models often fail to capture non-linear interactions, while modern deep learning approaches, though powerful, are limited by poor interpretability. We propose a Pipeline for Interpretable Survival Analysis (PISA) - a pipeline that provides multiple survival analysis models that trade off complexity and performance. Using multiple-feature, multi-objective feature engineering, PISA transforms patient characteristics and time-to-event data into multiple survival analysis models, providing valuable insights into the survival prediction task. Crucially, every model is converted into simple patient stratification flowcharts supported by Kaplan-Meier curves, whilst not compromising on performance. While PISA is model-agnostic, we illustrate its flexibility through applications of Cox regression and shallow survival trees, the latter avoiding proportional hazards assumptions. Applied to two clinical benchmark datasets, PISA produced interpretable survival models and intuitive stratification flowcharts whilst achieving state-of-the-art performances. Revisiting a prior departmental study further demonstrated its capacity to automate survival analysis workflows in real-world clinical research.

Vangelis Kostoulas, Arthur Guijt, Ellen M. Kerkhof et al.

Brachytherapy involves bringing a radioactive source near tumor tissue using implanted needles. Image-guided brachytherapy planning requires amongst others, the reconstruction of the needles. Manually annotating these needles on patient images can be a challenging and time-consuming task for medical professionals. For automatic needle reconstruction, a two-stage pipeline is commonly adopted, comprising a segmentation stage followed by a post-processing stage. While deep learning models are effective for segmentation, their results often contain errors. No currently existing post-processing technique is robust to all possible segmentation errors. We therefore propose adaptations to existing post-processing techniques mainly aimed at dealing with segmentation errors and thereby improving the reconstruction accuracy. Experiments on a prostate cancer dataset, based on MRI scans annotated by medical professionals, demonstrate that our proposed adaptations can help to effectively manage segmentation errors, with the best adapted post-processing technique achieving median needle-tip and needle-bottom point localization errors of $1.07$ (IQR $\pm 1.04$) mm and $0.43$ (IQR $\pm 0.46$) mm, respectively, and median shaft error of $0.75$ (IQR $\pm 0.69$) mm with 0 false positive and 0 false negative needles on a test set of 261 needles.

Georgios Andreadis, Eduard Ruiz Munné, Thomas H. W. Bäck et al.

When choosing a deformable image registration (DIR) approach for images with large deformations and content mismatch, the realism of found transformations often needs to be traded off against the required runtime. DIR approaches using deep learning (DL) techniques have shown remarkable promise in instantly predicting a transformation. However, on difficult registration problems, the realism of these transformations can fall short. DIR approaches using biomechanical, finite element modeling (FEM) techniques can find more realistic transformations, but tend to require much longer runtimes. This work proposes the first hybrid approach to combine them, with the aim of getting the best of both worlds. This hybrid approach, called DL-MOREA, combines a recently introduced multi-objective DL-based DIR approach which leverages the VoxelMorph framework, called DL-MODIR, with MOREA, an evolutionary algorithm-based, multi-objective DIR approach in which a FEM-like biomechanical mesh transformation model is used. In our proposed hybrid approach, the DL results are used to smartly initialize MOREA, with the aim of more efficiently optimizing its mesh transformation model. We empirically compare DL-MOREA against its components, DL-MODIR and MOREA, on CT scan pairs capturing large bladder filling differences of 15 cervical cancer patients. While MOREA requires a median runtime of 45 minutes, DL-MOREA can already find high-quality transformations after 5 minutes. Compared to the DL-MODIR transformations, the transformations found by DL-MOREA exhibit far less folding and improve or preserve the bladder contour distance error.

E. M. C. Sijben, J. C. Jansen, M. de Ridder et al.

Volume measurement of a paraganglioma (a rare neuroendocrine tumor that typically forms along major blood vessels and nerve pathways in the head and neck region) is crucial for monitoring and modeling tumor growth in the long term. However, in clinical practice, using available tools to do these measurements is time-consuming and suffers from tumor-shape assumptions and observer-to-observer variation. Growth modeling could play a significant role in solving a decades-old dilemma (stemming from uncertainty regarding how the tumor will develop over time). By giving paraganglioma patients treatment, severe symptoms can be prevented. However, treating patients who do not actually need it, comes at the cost of unnecessary possible side effects and complications. Improved measurement techniques could enable growth model studies with a large amount of tumor volume data, possibly giving valuable insights into how these tumors develop over time. Therefore, we propose an automated tumor volume measurement method based on a deep learning segmentation model using no-new-UNnet (nnUNet). We assess the performance of the model based on visual inspection by a senior otorhinolaryngologist and several quantitative metrics by comparing model outputs with manual delineations, including a comparison with variation in manual delineation by multiple observers. Our findings indicate that the automatic method performs (at least) equal to manual delineation. Finally, using the created model, and a linking procedure that we propose to track the tumor over time, we show how additional volume measurements affect the fit of known growth functions.

Damy M. F. Ha, Tanja Alderliesten, Peter A. N. Bosman

Bayesian networks model relationships between random variables under uncertainty and can be used to predict the likelihood of events and outcomes while incorporating observed evidence. From an eXplainable AI (XAI) perspective, such models are interesting as they tend to be compact. Moreover, captured relations can be directly inspected by domain experts. In practice, data is often real-valued. Unless assumptions of normality can be made, discretization is often required. The optimal discretization, however, depends on the relations modelled between the variables. This complicates learning Bayesian networks from data. For this reason, most literature focuses on learning conditional dependencies between sets of variables, called structure learning. In this work, we extend an existing state-of-the-art structure learning approach based on the Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA) to jointly learn variable discretizations. The proposed Discretized Bayesian Network GOMEA (DBN-GOMEA) obtains similar or better results than the current state-of-the-art when tasked to retrieve randomly generated ground-truth networks. Moreover, leveraging a key strength of evolutionary algorithms, we can straightforwardly perform DBN learning multi-objectively. We show how this enables incorporating expert knowledge in a uniquely insightful fashion, finding multiple DBNs that trade-off complexity, accuracy, and the difference with a pre-determined expert network.

Georgios Andreadis, Joas I. Mulder, Anton Bouter et al.

The transformation model is an essential component of any deformable image registration approach. It provides a representation of physical deformations between images, thereby defining the range and realism of registrations that can be found. Two types of transformation models have emerged as popular choices: B-spline models and mesh models. Although both models have been investigated in detail, a direct comparison has not yet been made, since the models are optimized using very different optimization methods in practice. B-spline models are predominantly optimized using gradient-descent methods, while mesh models are typically optimized using finite-element method solvers or evolutionary algorithms. Multi-objective optimization methods, which aim to find a diverse set of high-quality trade-off registrations, are increasingly acknowledged to be important in deformable image registration. Since these methods search for a diverse set of registrations, they can provide a more complete picture of the capabilities of different transformation models, making them suitable for a comparison of models. In this work, we conduct the first direct comparison between B-spline and mesh transformation models, by optimizing both models with the same state-of-the-art multi-objective optimization method, the Multi-Objective Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm (MO-RV-GOMEA). The combination with B-spline transformation models, moreover, is novel. We experimentally compare both models on two different registration problems that are both based on pelvic CT scans of cervical cancer patients, featuring large deformations. Our results, on three cervical cancer patients, indicate that the choice of transformation model can have a profound impact on the diversity and quality of achieved registration outcomes.

Arkadiy Dushatskiy, Gerry Lowe, Peter A. N. Bosman et al.

Deep learning algorithms have become the golden standard for segmentation of medical imaging data. In most works, the variability and heterogeneity of real clinical data is acknowledged to still be a problem. One way to automatically overcome this is to capture and exploit this variation explicitly. Here, we propose an approach that improves on our previous work in this area and explain how it potentially can improve clinical acceptance of (semi-)automatic segmentation methods. In contrast to a standard neural network that produces one segmentation, we propose to use a multi-pathUnet network that produces multiple segmentation variants, presumably corresponding to the variations that reside in the dataset. Different paths of the network are trained on disjoint data subsets. Because a priori it may be unclear what variations exist in the data, the subsets should be automatically determined. This is achieved by searching for the best data partitioning with an evolutionary optimization algorithm. Because each network path can become more specialized when trained on a more homogeneous data subset, better segmentation quality can be achieved. In practical usage, various automatically produced segmentations can be presented to a medical expert, from which the preferred segmentation can be selected. In experiments with a real clinical dataset of CT scans with prostate segmentations, our approach provides an improvement of several percentage points in terms of Dice and surface Dice coefficients compared to when all network paths are trained on all training data. Noticeably, the largest improvement occurs in the upper part of the prostate that is known to be most prone to inter-observer segmentation variation.

Martijn M. A. Bosma, Arkadiy Dushatskiy, Monika Grewal et al.

Deep Neural Networks (DNNs) have the potential for making various clinical procedures more time-efficient by automating medical image segmentation. Due to their strong, in some cases human-level, performance, they have become the standard approach in this field. The design of the best possible medical image segmentation DNNs, 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 for various tasks. However, the existing NAS methods for medical image segmentation have explored a quite limited range of types of DNN architectures that can be discovered. In this work, we propose 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 proven to have a strong performance in image classification tasks. The search is performed by looking for the best topology of multiple cells simultaneously with the configuration of each cell within, allowing for interactions between topology and cell-level attributes. From experiments on two publicly available datasets, we find that the networks discovered by our proposed NAS method have better performance 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.

Georgios Andreadis, Peter A. N. Bosman, Tanja Alderliesten

Reliably and physically accurately transferring information between images through deformable image registration with large anatomical differences is an open challenge in medical image analysis. Most existing methods have two key shortcomings: first, they require extensive up-front parameter tuning to each specific registration problem, and second, they have difficulty capturing large deformations and content mismatches between images. There have however been developments that have laid the foundation for potential solutions to both shortcomings. Towards the first shortcoming, a multi-objective optimization approach using the Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm (RV-GOMEA) has been shown to be capable of producing a diverse set of registrations for 2D images in one run of the algorithm, representing different trade-offs between conflicting objectives in the registration problem. This allows the user to select a registration afterwards and removes the need for up-front tuning. Towards the second shortcoming, a dual-dynamic grid transformation model has proven effective at capturing large differences in 2D images. These two developments have recently been accelerated through GPU parallelization, delivering large speed-ups. Based on this accelerated version, it is now possible to extend the approach to 3D images. Concordantly, this work introduces the first method for multi-objective 3D deformable image registration, using a 3D dual-dynamic grid transformation model based on simplex meshes while still supporting the incorporation of annotated guidance information and multi-resolution schemes. Our proof-of-concept prototype shows promising results on synthetic and clinical 3D registration problems, forming the foundation for a new, insightful method that can include bio-mechanical properties in the registration.

Dazhuang Liu, Marco Virgolin, Tanja Alderliesten et al.

Genetic programming (GP) is one of the best approaches today to discover symbolic regression models. To find models that trade off accuracy and complexity, the non-dominated sorting genetic algorithm II (NSGA-II) is widely used. Unfortunately, it has been shown that NSGA-II can be inefficient: in early generations, low-complexity models over-replicate and take over most of the population. Consequently, studies have proposed different approaches to promote diversity. Here, we study the root of this problem, in order to design a superior approach. We find that the over-replication of low complexity-models is due to a lack of evolvability, i.e., the inability to produce offspring with improved accuracy. We therefore extend NSGA-II to track, over time, the evolvability of models of different levels of complexity. With this information, we limit how many models of each complexity level are allowed to survive the generation. We compare this new version of NSGA-II, evoNSGA-II, with the use of seven existing multi-objective GP approaches on ten widely-used data sets, and find that evoNSGA-II is equal or superior to using these approaches in almost all comparisons. Furthermore, our results confirm that evoNSGA-II behaves as intended: models that are more evolvable form the majority of the population.

Marco Virgolin, Andrea De Lorenzo, Tanja Alderliesten et al.

Emotion recognition in children can help the early identification of, and intervention on, psychological complications that arise in stressful situations such as cancer treatment. Though deep learning models are increasingly being adopted, data scarcity is often an issue in pediatric medicine, including for facial emotion recognition in children. In this paper, we study the application of data augmentation-based contrastive learning to overcome data scarcity in facial emotion recognition for children. We explore the idea of ignoring generational gaps, by adding abundantly available adult data to pediatric data, to learn better representations. We investigate different ways by which adult facial expression images can be used alongside those of children. In particular, we propose to explicitly incorporate within each mini-batch adult images as augmentations for children's. Out of $84$ combinations of learning approaches and training set sizes, we find that supervised contrastive learning with the proposed training scheme performs best, reaching a test accuracy that typically surpasses the one of the second-best approach by 2% to 3%. Our results indicate that adult data can be considered to be a meaningful augmentation of pediatric data for the recognition of emotional facial expression in children, and open up the possibility for other applications of contrastive learning to improve pediatric care by complementing data of children with that of adults.

Arkadiy Dushatskiy, Tanja Alderliesten, Peter A. N. Bosman

Neural Architecture Search (NAS) has recently become a topic of great interest. However, there is a potentially impactful issue within NAS that remains largely unrecognized: noise. Due to stochastic factors in neural network initialization, training, and the chosen train/validation dataset split, the performance evaluation of a neural network architecture, which is often based on a single learning run, is also stochastic. This may have a particularly large impact if a dataset is small. We therefore propose to reduce this noise by evaluating architectures based on average performance over multiple network training runs using different random seeds and cross-validation. We perform experiments for a combinatorial optimization formulation of NAS in which we vary noise reduction levels. We use the same computational budget for each noise level in terms of network training runs, i.e., we allow less architecture evaluations when averaging over more training runs. Multiple search algorithms are considered, including evolutionary algorithms which generally perform well for NAS. We use two publicly available datasets from the medical image segmentation domain where datasets are often limited and variability among samples is often high. Our results show that reducing noise in architecture evaluations enables finding better architectures by all considered search algorithms.

Arkadiy Dushatskiy, Marco Virgolin, Anton Bouter et al.

When it comes to solving optimization problems with evolutionary algorithms (EAs) in a reliable and scalable manner, detecting and exploiting linkage information, i.e., dependencies between variables, can be key. In this article, we present the latest version of, and propose substantial enhancements to, the Gene-pool Optimal Mixing Evoutionary Algorithm (GOMEA): an EA explicitly designed to estimate and exploit linkage information. We begin by performing a large-scale search over several GOMEA design choices, to understand what matters most and obtain a generally best-performing version of the algorithm. Next, we introduce a novel version of GOMEA, called CGOMEA, where linkage-based variation is further improved by filtering solution mating based on conditional dependencies. We compare our latest version of GOMEA, the newly introduced CGOMEA, and another contending linkage-aware EA DSMGA-II in an extensive experimental evaluation, involving a benchmark set of 9 black-box problems that can only be solved efficiently if their inherent dependency structure is unveiled and exploited. Finally, in an attempt to make EAs more usable and resilient to parameter choices, we investigate the performance of different automatic population management schemes for GOMEA and CGOMEA, de facto making the EAs parameterless. Our results show that GOMEA and CGOMEA significantly outperform the original GOMEA and DSMGA-II on most problems, setting a new state of the art for the field.

Arkadiy Dushatskiy, Tanja Alderliesten, Peter A. N. Bosman

We propose a novel surrogate-assisted Evolutionary Algorithm for solving expensive combinatorial optimization problems. We integrate a surrogate model, which is used for fitness value estimation, into a state-of-the-art P3-like variant of the Gene-Pool Optimal Mixing Algorithm (GOMEA) and adapt the resulting algorithm for solving non-binary combinatorial problems. We test the proposed algorithm on an ensemble learning problem. Ensembling several models is a common Machine Learning technique to achieve better performance. We consider ensembles of several models trained on disjoint subsets of a dataset. Finding the best dataset partitioning is naturally a combinatorial non-binary optimization problem. Fitness function evaluations can be extremely expensive if complex models, such as Deep Neural Networks, are used as learners in an ensemble. Therefore, the number of fitness function evaluations is typically limited, necessitating expensive optimization techniques. In our experiments we use five classification datasets from the OpenML-CC18 benchmark and Support-vector Machines as learners in an ensemble. The proposed algorithm demonstrates better performance than alternative approaches, including Bayesian optimization algorithms. It manages to find better solutions using just several thousand fitness function evaluations for an ensemble learning problem with up to 500 variables.

Timo M. Deist, Monika Grewal, Frank J. W. M. Dankers et al.

Real-world problems are often multi-objective with decision-makers unable to specify a priori which trade-off between the conflicting objectives is preferable. Intuitively, building machine learning solutions in such cases would entail providing multiple predictions that span and uniformly cover the Pareto front of all optimal trade-off solutions. We propose a novel approach for multi-objective training of neural networks to approximate the Pareto front during inference. In our approach, the neural networks are trained multi-objectively using a dynamic loss function, wherein each network's losses (corresponding to multiple objectives) are weighted by their hypervolume maximizing gradients. We discuss and illustrate why training processes to approximate Pareto fronts need to optimize on fronts of individual training samples instead of on only the front of average losses. Experiments on three multi-objective problems show that our approach returns outputs that are well-spread across different trade-offs on the approximated Pareto front without requiring the trade-off vectors to be specified a priori. Further, results of comparisons with the state-of-the-art approaches highlight the added value of our proposed approach, especially in asymmetric Pareto fronts.

S. C. Maree, T. Alderliesten, P. A. N. Bosman

In model-based evolutionary algorithms (EAs), the underlying search distribution is adapted to the problem at hand, for example based on dependencies between decision variables. Hill-valley clustering is an adaptive niching method in which a set of solutions is clustered such that each cluster corresponds to a single mode in the fitness landscape. This can be used to adapt the search distribution of an EA to the number of modes, exploring each mode separately. Especially in a black-box setting, where the number of modes is a priori unknown, an adaptive approach is essential for good performance. In this work, we introduce multi-objective hill-valley clustering and combine it with MAMaLGaM, a multi-objective EA, into the multi-objective hill-valley EA (MO-HillVallEA). We empirically show that MO-HillVallEA outperforms MAMaLGaM and other well-known multi-objective optimization algorithms on a set of benchmark functions. Furthermore, and perhaps most important, we show that MO-HillVallEA is capable of obtaining and maintaining multiple approximation sets simultaneously over time.

S. C. Maree, T. Alderliesten, P. A. N. Bosman

The aim of bi-objective optimization is to obtain an approximation set of (near) Pareto optimal solutions. A decision maker then navigates this set to select a final desired solution, often using a visualization of the approximation front. The front provides a navigational ordering of solutions to traverse, but this ordering does not necessarily map to a smooth trajectory through decision space. This forces the decision maker to inspect the decision variables of each solution individually, potentially making navigation of the approximation set unintuitive. In this work, we aim to improve approximation set navigability by enforcing a form of smoothness or continuity between solutions in terms of their decision variables. Imposing smoothness as a restriction upon common domination-based multi-objective evolutionary algorithms is not straightforward. Therefore, we use the recently introduced uncrowded hypervolume (UHV) to reformulate the multi-objective optimization problem as a single-objective problem in which parameterized approximation sets are directly optimized. We study here the case of parameterizing approximation sets as smooth Bezier curves in decision space. We approach the resulting single-objective problem with the gene-pool optimal mixing evolutionary algorithm (GOMEA), and we call the resulting algorithm BezEA. We analyze the behavior of BezEA and compare it to optimization of the UHV with GOMEA as well as the domination-based multi-objective GOMEA. We show that high-quality approximation sets can be obtained with BezEA, sometimes even outperforming the domination- and UHV-based algorithms, while smoothness of the navigation trajectory through decision space is guaranteed.

T. Den Ottelander, A. Dushatskiy, M. Virgolin et al.

Neural Architecture Search (NAS), i.e., the automation of neural network design, has gained much popularity in recent years with increasingly complex search algorithms being proposed. Yet, solid comparisons with simple baselines are often missing. At the same time, recent retrospective studies have found many new algorithms to be no better than random search (RS). In this work we consider, for the first time, a simple Local Search (LS) algorithm for NAS. We particularly consider a multi-objective NAS formulation, with network accuracy and network complexity as two objectives, as understanding the trade-off between these two objectives is arguably the most interesting aspect of NAS. The proposed LS algorithm is compared with RS and two evolutionary algorithms (EAs), as these are often heralded as being ideal for multi-objective optimization. To promote reproducibility, we create and release two benchmark datasets, named MacroNAS-C10 and MacroNAS-C100, containing 200K saved network evaluations for two established image classification tasks, CIFAR-10 and CIFAR-100. Our benchmarks are designed to be complementary to existing benchmarks, especially in that they are better suited for multi-objective search. We additionally consider a version of the problem with a much larger architecture space. While we find and show that the considered algorithms explore the search space in fundamentally different ways, we also find that LS substantially outperforms RS and even performs nearly as good as state-of-the-art EAs. We believe that this provides strong evidence that LS is truly a competitive baseline for NAS against which new NAS algorithms should be benchmarked.

S. C. Maree, T. Alderliesten, P. A. N. Bosman

Domination-based multi-objective (MO) evolutionary algorithms (EAs) are today arguably the most frequently used type of MOEA. These methods however stagnate when the majority of the population becomes non-dominated, preventing convergence to the Pareto set. Hypervolume-based MO optimization has shown promising results to overcome this. Direct use of the hypervolume however results in no selection pressure for dominated solutions. The recently introduced Sofomore framework overcomes this by solving multiple interleaved single-objective dynamic problems that iteratively improve a single approximation set, based on the uncrowded hypervolume improvement (UHVI). It thereby however loses many advantages of population-based MO optimization, such as handling multimodality. Here, we reformulate the UHVI as a quality measure for approximation sets, called the uncrowded hypervolume (UHV), which can be used to directly solve MO optimization problems with a single-objective optimizer. We use the state-of-the-art gene-pool optimal mixing evolutionary algorithm (GOMEA) that is capable of efficiently exploiting the intrinsically available grey-box properties of this problem. The resulting algorithm, UHV-GOMEA, is compared to Sofomore equipped with GOMEA, and the domination-based MO-GOMEA. In doing so, we investigate in which scenarios either domination-based or hypervolume-based methods are preferred. Finally, we construct a simple hybrid approach that combines MO-GOMEA with UHV-GOMEA and outperforms both.

M. Virgolin, Z. Wang, B. V. Balgobind et al.

To study radiotherapy-related adverse effects, detailed dose information (3D distribution) is needed for accurate dose-effect modeling. For childhood cancer survivors who underwent radiotherapy in the pre-CT era, only 2D radiographs were acquired, thus 3D dose distributions must be reconstructed from limited information. State-of-the-art methods achieve this by using 3D surrogate anatomies. These can lack personalization and lead to coarse reconstructions. We present and validate a surrogate-free dose reconstruction method based on Machine Learning (ML). Abdominal planning CTs ($n$=142) of recently-treated childhood cancer patients were gathered, their organs at risk were segmented, and 300 artificial Wilms' tumor plans were sampled automatically. Each artificial plan was automatically emulated on the 142 CTs, resulting in 42,600 3D dose distributions from which dose-volume metrics were derived. Anatomical features were extracted from digitally reconstructed radiographs simulated from the CTs to resemble historical radiographs. Further, patient and radiotherapy plan features typically available from historical treatment records were collected. An evolutionary ML algorithm was then used to link features to dose-volume metrics. Besides 5-fold cross-validation, a further evaluation was done on an independent dataset of five CTs each associated with two clinical plans. Cross-validation resulted in Mean Absolute Errors (MAEs) $\leq$0.6 Gy for organs completely inside or outside the field. For organs positioned at the edge of the field, MAEs $\leq$1.7 Gy for D$_{mean}$, $\leq$2.9 Gy for D$_{2cc}$, and $\leq$13% for V$_{5Gy}$ and V$_{10Gy}$, were obtained, without systematic bias. Similar results were found for the independent dataset. Our novel, ML-based organ dose reconstruction method is not only accurate but also efficient, as the setup of a surrogate is no longer needed.

Monika Grewal, Timo M. Deist, Jan Wiersma et al.

Anatomical landmark correspondences in medical images can provide additional guidance information for the alignment of two images, which, in turn, is crucial for many medical applications. However, manual landmark annotation is labor-intensive. Therefore, we propose an end-to-end deep learning approach to automatically detect landmark correspondences in pairs of two-dimensional (2D) images. Our approach consists of a Siamese neural network, which is trained to identify salient locations in images as landmarks and predict matching probabilities for landmark pairs from two different images. We trained our approach on 2D transverse slices from 168 lower abdominal Computed Tomography (CT) scans. We tested the approach on 22,206 pairs of 2D slices with varying levels of intensity, affine, and elastic transformations. The proposed approach finds an average of 639, 466, and 370 landmark matches per image pair for intensity, affine, and elastic transformations, respectively, with spatial matching errors of at most 1 mm. Further, more than 99% of the landmark pairs are within a spatial matching error of 2 mm, 4 mm, and 8 mm for image pairs with intensity, affine, and elastic transformations, respectively. To investigate the utility of our developed approach in a clinical setting, we also tested our approach on pairs of transverse slices selected from follow-up CT scans of three patients. Visual inspection of the results revealed landmark matches in both bony anatomical regions as well as in soft tissues lacking prominent intensity gradients.

Marco Virgolin, Ziyuan Wang, Tanja Alderliesten et al.

Machine Learning (ML) is proving extremely beneficial in many healthcare applications. In pediatric oncology, retrospective studies that investigate the relationship between treatment and late adverse effects still rely on simple heuristics. To assess the effects of radiation therapy, treatment plans are typically simulated on phantoms, i.e., virtual surrogates of patient anatomy. Currently, phantoms are built according to reasonable, yet simple, human-designed criteria. This often results in a lack of individualization. We present a novel approach that combines imaging and ML to build individualized phantoms automatically. Given the features of a patient treated historically (only 2D radiographs available), and a database of 3D Computed Tomography (CT) imaging with organ segmentations and relative patient features, our approach uses ML to predict how to assemble a patient-specific phantom automatically. Experiments on 60 abdominal CTs of pediatric patients show that our approach constructs significantly more representative phantoms than using current phantom building criteria, in terms of location and shape of the abdomen and of two considered organs, the liver and the spleen. Among several ML algorithms considered, the Gene-pool Optimal Mixing Evolutionary Algorithm for Genetic Programming (GP-GOMEA) is found to deliver the best performing models, which are, moreover, transparent and interpretable mathematical expressions.

S. C. Maree, T. Alderliesten, P. A. N. Bosman

This report presents benchmarking results of the Hill-Valley Evolutionary Algorithm version 2019 (HillVallEA19) on the CEC2013 niching benchmark suite under the restrictions of the GECCO 2019 niching competition on multimodal optimization. Performance is compared to algorithms that participated in previous editions of the niching competition.

Marco Virgolin, Tanja Alderliesten, Peter A. N. Bosman

Feature construction can substantially improve the accuracy of Machine Learning (ML) algorithms. Genetic Programming (GP) has been proven to be effective at this task by evolving non-linear combinations of input features. GP additionally has the potential to improve ML explainability since explicit expressions are evolved. Yet, in most GP works the complexity of evolved features is not explicitly bound or minimized though this is arguably key for explainability. In this article, we assess to what extent GP still performs favorably at feature construction when constructing features that are (1) Of small-enough number, to enable visualization of the behavior of the ML model; (2) Of small-enough size, to enable interpretability of the features themselves; (3) Of sufficient informative power, to retain or even improve the performance of the ML algorithm. We consider a simple feature construction scheme using three different GP algorithms, as well as random search, to evolve features for five ML algorithms, including support vector machines and random forest. Our results on 21 datasets pertaining to classification and regression problems show that constructing only two compact features can be sufficient to rival the use of the entire original feature set. We further find that a modern GP algorithm, GP-GOMEA, performs best overall. These results, combined with examples that we provide of readable constructed features and of 2D visualizations of ML behavior, lead us to positively conclude that GP-based feature construction still works well when explicitly searching for compact features, making it extremely helpful to explain ML models.

S. C. Maree, T. Alderliesten, D. Thierens et al.

Model-based evolutionary algorithms (EAs) adapt an underlying search model to features of the problem at hand, such as the linkage between problem variables. The performance of EAs often deteriorates as multiple modes in the fitness landscape are modelled with a unimodal search model. The number of modes is however often unknown a priori, especially in a black-box setting, which complicates adaptation of the search model. In this work, we focus on models that can adapt to the multi-modality of the fitness landscape. Specifically, we introduce Hill-Valley Clustering, a remarkably simple approach to adaptively cluster the search space in niches, such that a single mode resides in each niche. In each of the located niches, a core search algorithm is initialized to optimize that niche. Combined with an EA and a restart scheme, the resulting Hill-Valley EA (HillVallEA) is compared to current state-of-the-art niching methods on a standard benchmark suite for multi-modal optimization. Numerical results in terms of the detected number of global optima show that, in spite of its simplicity, HillVallEA is competitive within the limited budget of the benchmark suite, and shows superior performance in the long run.

S. C. Maree, T. Alderliesten, D. Thierens et al.

This report presents benchmarking results of the latest version of the Hill-Valley Evolutionary Algorithm (HillVallEA) on the CEC2013 niching benchmark suite. The benchmarking follows restrictions required by the GECCO 2018 competition on Niching methods for Multimodal Optimization. In particular, no problem dependent parameter tuning is performed. A number of adjustments have been made to original publication of HillVallEA that are discussed in this report.