Accurate segmentation of organs at risk in head-and-neck radiotherapy is important for limiting radiation-induced damage to healthy tissue, yet fully automatic segmentation methods still require manual correction in difficult cases. Interactive refinement offers a practical alternative, but many refinement systems are trained with simulated point annotations while being used in practice with human scribbles. This creates a train-test mismatch in the user-guidance channels, which we study in this thesis as annotation-shift. To address this problem, we propose a training pipeline for interactive segmentation refinement based on simulated human-like scribbles, combined with annotation-aware supervision, distance-map encodings, and explicit manipulation of the initial segmentation. The method is evaluated on head-and-neck CT data for parotid gland segmentation and compared with both a standard point-based baseline and a balanced point-based baseline. The results show that outperforms the original point-based refinement strategy, with the clearest gains appearing in the local evaluation near the annotated region. At the same time, the smaller gap between scribbles and balanced points indicates that refinement quality depends not only on annotation geometry, but also on the amount and spatial distribution of corrective information. These findings provide evidence that annotation-shift affects interactive refinement in interactive medical image segmentation and that simulated scribbles provide a useful step toward refinement systems that are better aligned with real clinical annotation practice. ... Read more

In the hospitality sector, food waste remains a persistent challenge, despite the growing availability of strategies to reduce it. So, rather than being solely a matter of available solutions, the adoption of these strategies is also shaped by the organisational and systemic context in which hospitality businesses operate. This research investigates to what extent independent hospitality businesses in Amsterdam have adopted food waste reduction strategies and what factors influence their integration into everyday practices. The Multi-Level Perspective (MLP) and absorptive capacity theory are combined to examine how food waste reduction strategies, approached as niches, interact with wider system dynamics and organisational capacity to implement change. A qualitative approach was used, which consisted of a document analysis and 13 semi-structured interviews with representatives of independent hospitality businesses. Four categories of niche strategies were identified: managerial practices, behavioural changes, technological services, and social networks. The results further show that adoption remains uneven, with internally driven strategies more widely embedded in everyday operations than externally oriented ones. Furthermore, financial pressures encouraged more cost-conscious behaviour, which also reduced waste, while existing routines, supplier relationships, food safety requirements, and limited infrastructure constrained further change. Although operators possessed a general awareness about certain strategies, practical barriers such as effort and cost involved, as well as established routines and attitudes towards food waste, often made it difficult to implement these strategies. In addition, a wasteaware organisational culture and active individuals within businesses generally supported further food waste reduction. The findings indicate that reducing food waste is not only a matter of developing effective strategies, but also of understanding how these strategies are absorbed into existing organisational and sectoral structures. For Amsterdam, this research provides insight into where sustainability initiatives are successfully integrated into hospitality practices and where further support may be needed to enable broader adoption. ... Read more

**Introduction:** Blunt thoracic trauma frequently results in pneumothorax (PTX), haemothorax (HTX), and pulmonary contusion (PC), injuries that may contribute to post-traumatic pulmonary complications such as pneumonia. Although computed tomography (CT) enables detailed assessment of these injuries, their extent is generally not quantified routinely. Automated image analysis may enable objective quantification of thoracic injury burden and provide quantitative imaging features for prediction of post-traumatic pneumonia. This study investigated whether PTX, HTX, and PC can be automatically and reliably quantified from admission CT and whether these quantitative imaging features can be integrated into pneumonia prediction.

**Methods:** PTX and HTX were segmented using a multi-class nnU-Net model with lung parenchyma as an auxiliary anatomical class. Model configurations and post-processing thresholds were optimised during development, and the final pipeline was evaluated on an independent test set. For PC, six clinical observers assessed contusion extent within twelve predefined lung regions. Inter- and intraobserver agreement were evaluated, and the mean assessment of the six observers was used as a consensus reference for development of a 3D convolutional neural network (CNN) for automated PC quantification. Finally, the automatically derived PTX, HTX, and PC measurements were evaluated as radiological features within an existing clinical–radiological model for post-traumatic pneumonia prediction.

**Results:** On independent testing, PTX achieved a sensitivity of 1.00, specificity of 0.80, and mean absolute volume error of 5.2 mL. For HTX, sensitivity was 0.81, specificity 0.82, and mean absolute volume error 35.4 mL. Including lung parenchyma as an auxiliary class significantly reduced HTX volume error and false-positive predictions. PC assessment showed moderate interobserver agreement and substantial variation in intraobserver agreement. Automated PC quantification achieved a patient-level mean absolute error of 5.70 percentage points and a calibration slope of 0.68 on independent testing, with better performance at patient than regional level. Integration of the resulting automated pulmonary injury measurements into the pneumonia prediction framework yielded a test area under the receiver operating characteristic curve (AUC) of 0.81, compared with 0.83 for the previously developed combined clinical–radiological model.

**Conclusion:** Automated quantification of pulmonary injury from trauma CT was feasible, although performance differed between injury types. The resulting PTX, HTX, and PC measurements could be integrated into the existing pneumonia prediction framework while largely preserving discrimination compared with the original combined model. External validation is required to determine whether these measurements and resulting risk estimates can provide clinically meaningful support for patient management.
... Read more

Social science research produces a wealth of valuable data in the form of social networks. Making these empirical data public benefits research on socially relevant causes such as the spread of diseases, social interactions in a geographical location or online, movement and travel of people, flow of information in communication networks, etc. To protect the privacy of people represented in these networks while publicizing the data, anonymization is required. One way to identify an individual in network data is to study the local structure of the network around them. By altering the networks such that no node has a uniquely identifiable neighborhood, we can provide additional anonymity beyond concealing names and private identifiers.

The difficulty in this process of social network anonymization is that the network must remain useful to scientists, retaining properties important for downstream analysis. So far, research in this field has focused on maximizing anonymity and only measuring the utility of the anonymized data after the fact. The literature is also fragmented, with bespoke methods spanning different objectives unique to the dataset under consideration. This makes it costly for the data owner to understand and compare different approaches when their expertise lies in the domain of the data and not complex anonymization algorithms.

The main contribution of this thesis is a constraint programming-based framework in which scientists can easily explore the trade-offs between levels of anonymity for different attack models and the utility of the data for different measures. Within this sandbox, anonymity and utility requirements are expressed as independent, composable constraints. We model two structural anonymity measures of increasing strictness, k-degree and k-(degree, triangle), and introduce a betweenness-centrality preservation constraint that protects utility proactively by making modifications between central nodes expensive. We choose constraint programming for its declarative nature, which makes it easier to interpret the anonymization process. Evaluated on real-world networks, the framework shows that actively incorporating utility into the anonymization model is a promising direction. ... Read more

Unlike modality or structure-specific segmentation models such as U-Net, foundation models like Medical Segment Anything Model (MedSAM) segment across imaging modalities and anatomical structures without retraining, at a fixed cost. MedSAM pairs a Vision Transformer Base Model (ViT-B) image encoder with a prompt encoder and a mask decoder. The encoder dominates, accounting for 95.66% of the parameters and, at 966 giga Floating Point Operations (FLOPs), for almost all of the computation of a single forward pass, incurred for every image regardless of the modality segmented. This work characterises that cost analytically, deriving a block’s FLOPs from its retained attention heads and Multi-Layer Perceptron (MLP) neurons rather than by profiling, and then asks how much of the encoder is redundant for medical image segmentation. Redundancy is studied at two levels. At the level of the individual weight, redundant weights are zeroed while the weight-matrix shapes are left unchanged, so no computation is saved. At the level of the whole computational structure, removing an attention head or an MLP neuron physically reduces the weight count. It is the only level at which inference is accelerated. Sparsity is induced during training by a penalty on the loss: the Lasso ℓ1 penalty drives individual weights towards zero, paired with unstructured pruning, whilst Sparse Group Lasso (SGL) groups the weights of each head and neuron and drives whole groups towards zero, paired with structured pruning. As both penalties are non-smooth at zero, they are imposed via Proximal Gradient Descent (PGD), in which the gradient step optimises the task loss and a closed-form shrinkage sets small weights to exactly zero. Both pruned regimes are compared against a dense task-loss-fine-tuned reference and a variant with the encoder removed entirely, the latter serving as a floor
that tests whether the encoder is needed at all for a given target. Evaluated across seven imaging modalities and eleven evaluation subsets, the encoder is found to hold substantial but unequal redundancy. At the individual-weight level, 72.49% of the prunable encoder weights can be zeroed while maintaining accuracy comparable to the baseline, but the irregular sparsity yields no computational savings. As whole structures, 91.7% of the prunable encoder weight volume can be removed, reducing the model from 93.74 to 15.82 million parameters and the encoder cost from 966.32 to 81.08 giga FLOPs, which does come at a cost of the Dice Similarity Coeffi-
cient (DSC). This cost is not uniform across structures. Some targets retain usable accuracy with the encoder fully removed, such as the ultrasound foetal head and the fundus optic disc, both above 92% DSC, whilst others depend heavily on it, most notably the Chest X-ray
(CXR) lung fields, whose median DSC falls to 54.84%. The per-structure cost ranges from
the optic cup, which improves by 3.08 points under compression, to GlaS (testB), which
falls by 8.60. The MedSAM encoder therefore carries redundancy that can be removed for
real savings in size and speed. However, the extent of removal depends on the structure
and modality being segmented. ... Read more