Radiotherapy (RT) is a widespread and effective technique to treat cancers by killing cancerous cells with rays of radiation. Building upon advances in image guidance and dose delivery technology like Proton Therapy, Adaptive RT promises more effective tumor decimation and a reduction of the incidence and severity of side effects. Unfortunately, the clinical implementation of adaptive workflows is challenging due to their resource-intensive nature. Therefore, their successful adoption lingers on overcoming several bottlenecks in the treatment planning process.

In this dissertation, we focus on methods used for the image segmentation or contouring step, which allows the localization of the anatomical structures required for dose optimization and evaluation. Until recently, clinicians had to manually delineate dozens of organs-at-risk and target volumes across hundreds of slices of the patient’s three-dimensional images. A process that is extremely time-consuming. The advent of deep learning-based artificial intelligence (AI) has changed the landscape: a modern auto-segmentation AI can produce segmentations for most of a patient’s anatomy in minutes.

Despite increasing automation in the segmentation process, it remains time and resource-intensive. Due to the segmentations’ criticality for the patient’s outcome and the errors the AI will commit, clinicians must perform a quality assessment of the AI’s outputs. Depending on the case’s complexity, the duration of the quality assessment process can negate the time gains auto-segmentation tools bring.

Deep ensemble AIs represent an advancement in medical image segmentation. Instead of providing a deterministic output, deep ensemble AIs produce a set of plausible candidates that aim to model inter-clinician annotation variability. Consensus segmentations obtained from ensembles tend to be more accurate and robust than the single-prediction deterministic counterpart. Nevertheless, by only using the consensus, a lot of potentially useful information is being discarded.

In this dissertation, we contribute to different phases of the segmentation quality assessment process. We characterize this process and introduce methods that leverage the raw outputs of deep ensemble AIs to support and speed up quality assessment tasks. The methods presented show new ways of analyzing and using ensembles in RT. Nevertheless, since these are relevant outside RT, we keep the presentation of the methods general and evaluate them in other application scenarios, such as the analysis of simulation ensembles or meteorological data.

Before fixing segmentation failures, clinicians must find them. This process can be time-consuming and fatiguing when failures are sparse and spread through the patient’s three-dimensional images. We present and evaluate a delineation error detection system, which guides clinicians to slices of three-dimensional images that contain potentially clinically relevant segmentation failures. We co-designed the DEDS with clinicians and refined it based on an observational study, which allowed us to characterize clinicians’ navigation patterns and the use of information sources like AI uncertainty and patients’ dose distributions. We evaluated the DEDS’ potential to speed up the QA process through a simulation study with a retrospective cohort of patients. Results indicate that speed-ups are the most significant when equipping the DEDS with information sources indicative of clinical priority, which prevents unnecessary edits.

Visual inspection of the segmentation ensemble permits understanding the main trends and detecting anomalies that might indicate segmentation failures. Using a spaghetti plot to visualize all ensemble members is straightforward but prone to clutter. Contour boxplots prevent clutter and extra complexity by distilling essential ensemble information, which permits more efficient ensemble inspection. Nevertheless, they are time-consuming to compute, reducing their practical value. We present Inclusion Depth for contour ensembles. Inclusion Depth yields per ensemble member centrality scores that allow characterizing the distribution of segmentation ensembles in terms of properties like the median, trimmed mean, confidence bands, and outliers. Compared to previous contour depth notions, Inclusion Depth is significantly faster, making it more applicable in practice for time-critical contexts like QA in adaptive RT. We show how Inclusion Depth permits creating contour boxplots for ensembles with hundreds of segmentations in seconds.

It is not uncommon for distinct representative shapes to co-occur within a contour ensemble. With ensembles created by clinicians, for instance, different institutions, training sessions, or experience levels can lead to distinct shapes (i.e., modes of variation) for the same structure. When trained on these data, deep ensemble AIs would yield similarly multimodal ensembles. In quality assessment, being able to extract these representatives would pave the way for new ensemble-based interactive segmentation workflows. Applying traditional contour depth notions to these multi-modal ensembles collapses the existing variation modes and can lead to uninformative centrality scores. To address this issue, we present the first framework for multi-modal contour depth, which also includes notable runtime improvements for depth computation. When used with Inclusion Depth, multi-modal contour depth permits clustering the different modes of variation and determining cluster-dependent scores that appropriately characterize the data. Variation modes can be then independently analyzed using uni-modal depth machinery like contour boxplots. xiii

The global perspective of contour depth methods, which consider the entire volume, may be insufficient when parts of the contours are noisy or when the resolution of the ensemble is too large to process within a reasonable time. Correlation clustering methods provide a solution by partitioning the spatial domain of the ensemble into highly correlated regions that can be used to localize analyses. Existing correlation clustering algorithms do not scale well as the resolution of the ensemble increases. We introduce the Local-to-Global Correlation Clustering (LoGCC) method, which partitions the ensemble’s spatial domain into coarser primitives, representing areas of consistent ensemble member behavior. Unlike previous correlation clustering methods, the proposed LoGCC achieves significantly faster runtimes by leveraging the ensemble’s spatial structure and decoupling computations into local and global steps. Like with Inclusion Depth, these speed gains enable LoGCC to analyze large datasets in time-critical fields such as adaptive radiotherapy (RT).

Throughout this dissertation, our approach focused on designing modular, flexible analysis methods applicable across different tasks and domains. We demonstrate how the delineation error detection system, multi-modal Inclusion Depth, and Local-to-Global Correlation Clustering support quality assessment in RT and extend to fields like meteorology. We also speculate on their potential as foundational elements for more complex workflows. For example, extracted modes of variation, which indicate representative shapes in the ensemble, could be repurposed as an interactive segmentation tool. Alternatively, consistent regions detected by correlation clustering could be used as building blocks to enable localized contour analysis and editing.

We hope the proposed contour ensemble visual analysis methods inspire the development of more efficient analysis workflows that harness ensembles’ power in RT and beyond.

Artificial Intelligence (AI)-based auto-delineation technologies rapidly delineate multiple structures of interest like organs-at-risk and tumors in 3D medical images, reducing personnel load and facilitating time-critical therapies. Despite its accuracy, the AI may produce flawed delineations, requiring clinician attention. Quality assessment (QA) of these delineations is laborious and demanding. Delineation error detection systems (DEDS) aim to aid QA, yet questions linger about potential challenges to their adoption and time-saving potential. To address these queries, we first conducted a user study with two clinicians from Holland Proton Therapy Center, a Dutch cancer treatment center. Based on the study’s findings about the clinicians’ error detection workflows with and without DEDS assistance, we developed a simulation model of the QA process, which we used to assess different error detection workflows on a retrospective cohort of 42 head and neck cancer patients. Results suggest possible time savings, provided the per-slice analysis time stays close to the current baseline and trading-off delineation quality is acceptable. Our findings encourage the development of user-centric delineation error detection systems and provide a new way to model and evaluate these systems’ potential clinical value.

The need to understand the structure of hierarchical or high-dimensional data is present in a variety of fields. Hyperbolic spaces have proven to be an important tool for embedding computations and analysis tasks as their non-linear nature lends itself well to tree or graph data. Subsequently, they have also been used in the visualization of high-dimensional data, where they exhibit increased embedding performance. However, none of the existing dimensionality reduction methods for embedding into hyperbolic spaces scale well with the size of the input data. That is because the embeddings are computed via iterative optimization schemes and the computation cost of every iteration is quadratic in the size of the input. Furthermore, due to the non-linear nature of hyperbolic spaces, euclidean acceleration structures cannot directly be translated to the hyperbolic setting. This article introduces the first acceleration structure for hyperbolic embeddings, building upon a polar quadtree. We compare our approach with existing methods and demonstrate that it computes embeddings of similar quality in significantly less time. Implementation and scripts for the experiments can be found at https://graphics.tudelft.nl/accelerating-hyperbolic-tsne .

Background and purpose: Retrospective dose evaluation for organ-at-risk auto-contours has previously used small cohorts due to additional manual effort required for treatment planning on auto-contours. We aimed to do this at large scale, by a) proposing and assessing an automated plan optimization workflow that used existing clinical plan parameters and b) using it for head-and-neck auto-contour dose evaluation. Materials and methods: Our automated workflow emulated our clinic's treatment planning protocol and reused existing clinical plan optimization parameters. This workflow recreated the original clinical plan (P_OG) with manual contours (P_MC) and evaluated the dose effect (P_OG-P_MC) on 70 photon and 30 proton plans of head-and-neck patients. As a use-case, the same workflow (and parameters) created a plan using auto-contours (P_AC) of eight head-and-neck organs-at-risk from a commercial tool and evaluated their dose effect (P_MC-P_AC). Results: For plan recreation (P_OG-P_MC), our workflow had a median impact of 1.0% and 1.5% across dose metrics of auto-contours, for photon and proton respectively. Computer time of automated planning was 25% (photon) and 42% (proton) of manual planning time. For auto-contour evaluation (P_MC-P_AC), we noticed an impact of 2.0% and 2.6% for photon and proton radiotherapy. All evaluations had a median ΔNTCP (Normal Tissue Complication Probability) less than 0.3%. Conclusions: The plan replication capability of our automated program provides a blueprint for other clinics to perform auto-contour dose evaluation with large patient cohorts. Finally, despite geometric differences, auto-contours had a minimal median dose impact, hence inspiring confidence in their utility and facilitating their clinical adoption.

Ensembles of contours arise in various applications like simulation, computer-Aided design, and semantic segmentation. Uncovering ensemble patterns and analyzing individual members is a challenging task that suffers from clutter. Ensemble statistical summarization can alleviate this issue by permitting analyzing ensembles' distributional components like the mean and median, confidence intervals, and outliers. Contour boxplots, powered by Contour Band Depth (CBD), are a popular non-parametric ensemble summarization method that benefits from CBD's generality, robustness, and theoretical properties. In this work, we introduce Inclusion Depth (ID), a new notion of contour depth with three defining characteristics. First, ID is a generalization of functional Half-Region Depth, which offers several theoretical guarantees. Second, ID relies on a simple principle: The inside/outside relationships between contours. This facilitates implementing ID and understanding its results. Third, the computational complexity of ID scales quadratically in the number of members of the ensemble, improving CBD's cubic complexity. This also in practice speeds up the computation enabling the use of ID for exploring large contour ensembles or in contexts requiring multiple depth evaluations like clustering. In a series of experiments on synthetic data and case studies with meteorological and segmentation data, we evaluate ID's performance and demonstrate its capabilities for the visual analysis of contour ensembles.

Widely used pipelines for analyzing high-dimensional data utilize two-dimensional visualizations. These are created, for instance, via t-distributed stochastic neighbor embedding (t-SNE). A crucial element of the t-SNE embedding procedure is the perplexity hyperparameter. That is because the embedding structure varies when perplexity is changed. A suitable perplexity choice depends on the data set and the intended usage for the embedding. Therefore, perplexity is often chosen based on heuristics, intuition, and prior experience. This paper uncovers a linear relationship between perplexity and the data set size. Namely, we show that embeddings remain structurally consistent across data set samples when perplexity is adjusted accordingly. Qualitative and quantitative experimental results support these findings. This informs the visualization process, guiding the user when choosing a perplexity value. Finally, we outline several applications for the visualization of high-dimensional data via t-SNE based on this linear relationship.

Deep learning models for organ contouring in radiotherapy are poised for clinical usage, but currently, there exist few tools for automated quality assessment (QA) of the predicted contours. Bayesian models and their associated uncertainty, can potentially automate the process of detecting inaccurate predictions. We investigate two Bayesian models for auto-contouring, DropOut and FlipOut, using a quantitative measure – expected calibration error (ECE) and a qualitative measure – region-based accuracy-vs-uncertainty (R-AvU) graphs. It is well understood that a model should have low ECE to be considered trustworthy. However, in a QA context, a model should also have high uncertainty in inaccurate regions and low uncertainty in accurate regions. Such behaviour could direct visual attention of expert users to potentially inaccurate regions, leading to a speed-up in the QA process. Using R-AvU graphs, we qualitatively compare the behaviour of different models in accurate and inaccurate regions. Experiments are conducted on the MICCAI2015 Head and Neck Segmentation Challenge and on the DeepMindTCIA CT dataset using three models: DropOut-DICE, Dropout-CE (Cross Entropy) and FlipOut-CE. Quantitative results show that DropOut-DICE has the highest ECE, while Dropout-CE and FlipOut-CE have the lowest ECE. To better understand the difference between DropOut-CE and FlipOut-CE, we use the R-AvU graph which shows that FlipOut-CE has better uncertainty coverage in inaccurate regions than DropOut-CE. Such a combination of quantitative and qualitative metrics explores a new approach that helps to select which model can be deployed as a QA tool in clinical settings.

Bayesian Neural Nets (BNN) are increasingly used for robust organ auto-contouring. Uncertainty heatmaps extracted from BNNs have been shown to correspond to inaccurate regions. To help speed up the mandatory quality assessment (QA) of contours in radiotherapy, these heatmaps could be used as stimuli to direct visual attention of clinicians to potential inaccuracies. In practice, this is non-trivial to achieve since many accurate regions also exhibit uncertainty. To influence the output uncertainty of a BNN, we propose a modified accuracy-versus-uncertainty (AvU) metric as an additional objective during model training that penalizes both accurate regions exhibiting uncertainty as well as inaccurate regions exhibiting certainty. For evaluation, we use an uncertainty-ROC curve that can help differentiate between Bayesian models by comparing the probability of uncertainty in inaccurate versus accurate regions. We train and evaluate a FlipOut BNN model on the MICCAI2015 Head and Neck Segmentation challenge dataset and on the DeepMind-TCIA dataset, and observed an increase in the AUC of uncertainty-ROC curves by 5.6% and 5.9%, respectively, when using the AvU objective. The AvU objective primarily reduced false positives regions (uncertain and accurate), drawing less visual attention to these regions, thereby potentially improving the speed of error detection.

Delineation of tumours and organs-at-risk permits detecting and correcting changes in the patients' anatomy throughout the treatment, making it a core step of adaptive external beam radiotherapy. Although auto-contouring technologies have sped up this process, the time needed to perform the quality assessment of the generated contours remains a bottleneck, taking clinicians between several minutes and an hour to complete. The authors of this article conducted several interviews and an observational study at two treatment centres in the Netherlands to identify challenges and opportunities for speeding up the delineation process in adaptive therapies. The study revealed three contextual variables that influence contouring performance: usable additional information, applicable domain-specific knowledge, and available editing capabilities in contouring software. In practice, clinicians leverage these variables to accelerate contouring in two ways. First, they use domain-specific knowledge and relevant clinical features such as the proximity of the organs-at-risk to the tumour to enable targeted inspection of the delineation. Second, clinicians modulate editing precision depending on the effect they anticipate the edit will have on the patient outcome. By implementing these acceleration strategies in guidelines and contouring tools, developers and workflow builders could increase contouring efficiency and consistency without affecting the patient outcome.