Stroke remains a leading cause of morbidity and mortality worldwide, despite advances in treatment modalities. Endovascular thrombectomy (EVT), a revolutionary intervention for ischemic stroke, is limited by its reliance on 2D fluoroscopic imaging, which lacks depth and comprehensive vascular detail. We propose a novel AI-driven pipeline for 3D CTA to 2D DSA cross-modality registration, termed DeepIterReg. The proposed pipeline integrates neural network-based initialization with iterative optimization to align pre-intervention and peri-intervention data. Our approach addresses the challenges of cross-modality alignment, particularly in scenarios involving limited shared vascular structures, by leveraging synthetic data, vein-centric anchoring, and differentiable rendering techniques. We assess the efficacy of DeepIterReg through quantitative analysis of capture ranges and registration accuracy. Results indicate that our method is able to accurately register 70% of a testset of 20 patients, and is able to improve capture ranges when performing an initial pose estimation using a convolutional neural network.

Introduction: Trauma-induced rib fractures are a common injury, affecting millions of individuals globally each year. The number and characteristics of these fractures influence whether a patient is treated conservatively or surgically. Rib fractures are typically diagnosed using CT scans, yet 19.2% to 26.8% of fractures are still missed during assessment. Another challenge in managing rib fractures is the interobserver variability in their classification. In 2023, a deep learning-based algorithm for the automatic detection and classification of rib fractures (DCRibFrac v1.0) was developed based on the Chest Wall Injury Society (CWIS) taxonomy. Although DCRibFrac v1.0 demonstrated promising results, there remained room for improvement, particularly in the accuracy of rib number labelling. This project aims to develop and assess DCRibFrac v2.0, an enhanced version of the original algorithm.
Methods: Two novel approaches for automatic rib number labelling were proposed and evaluated: a pre-trained deep learning model named TotalSegmentator and a custom-developed nnUNet. Additionally, three nnDetection models were developed based on multi-centre data for the automatic detection of fractures and the classification of their type, displacement, and location. The performance of DCRibFrac v2.0 was evaluated. Finally, an external validation was conducted to assess the generalizability and robustness of DCRibFrac v2.0.
Results: For the development and evaluation of DCRibFrac v2.0 a total of 170 patients were included. The custom-developed nnUNet was the best-performing method for rib number labelling, correctly labelling 95.5% of all ribs and 98.4% of fractured ribs in 30 patients. On the internal test set, DCRibFrac v2.0 achieved a detection sensitivity of 80%, a precision of 87%, and an F1 score of 83%, with a mean FPPS (false positives per scan) rate of 1.11. Classification sensitivity varied across fracture types, with the lowest being 25% for complex fractures and the highest being 97% for posterior fractures. The correct rib number was assigned to 94% of the detected fractures. The detection and classification performance on the external validation dataset was slightly better, with a fracture detection sensitivity of 84%, precision of 85%, F1 score of 84%, FPPS of 0.96 and 95% of the fractures assigned the correct rib number.
Conclusion: The developments resulting in DCRibFrac v2.0 have improved the automatic detection and classification of rib fractures, with particular advancements in rib number labelling performance and detection precision. These improvements are important steps towards establishing a more accurate and standardised method for rib fracture assessment, which could enhance clinical decision-making and improve patient outcomes in trauma care.

Introduction
Calcaneus fractures represent 60% of tarsal bone fractures and are particularly challenging to assess when they involve the posterior talocalcaneal (PTC) facet. Accurate evaluation of intra-articular fractures, including metrics such as gap area and step-off, is crucial for treatment planning but remains difficult due to limitations of using 2D cross-sections from 3D CT imaging. Deciding between surgical or conservative management from these images often leads to significant inter- and intra-observer variability, complicating clinical decisions and affecting outcomes. This study aims to develop and validate an AI-based method for automatic segmentation and 3D quantitative analysis of PTC fractures, improving assessment and providing more objective data for decision-making.
Methods
A retrospective study was conducted on CT scans from 44 patients for training using 5-fold cross-validation and 33 patients for external validation. The nnU-Net framework was trained to segment the PTC fragments and posterior talar facet (PTF). Automatic 3D measurements, including gap area, inter-articular distances, maximal step-off, and maximal gap, were computed from the segmented models. Results were validated against manual 2D measurements performed by two observers, as well as through comparison with the external validation set.
Results
The nnU-Net achieved a Dice score of 0.78 for PTC segmentation in the training set and 0.75 in the external validation set. Moderate positive correlations were observed between the 3D automatic measurements and manual 2D measurements. Specifically, the correlation between 3D gap area and 2D maximal gap measurement was Spearman’s rho = 0.62, while the correlation between 3D and 2D maximal step-off measurements was rho = 0.52. Additional correlations were found between the 3D fracture area and the 2D maximal gap and step-off measurements, with rho values of 0.65 and 0.64, respectively, indicating that the 3D analysis is consistent with corresponding 2D measurements.
Conclusion
This study introduces an AI-based method for automatic 3D analysis of calcaneus fractures, offering faster and more detailed fracture metrics, which may improve treatment planning over traditional 2D slice evaluations.

Ischemic stroke, a leading cause of death and disability worldwide, occurs when a blood vessel is occluded by a thrombus. Current therapies for ischemic stroke, include Intravenous thrombolysis (IVT) and Endovascular thrombectomy (EVT). EVT relies on a Thrombolysis in Cerebral Infarction (TICI) score for assessing treatment effectiveness. This score, based on the visual evaluation of medical images by physicians, suffers from inter- and intra-observer variability, as it is influenced by the individual rater’s judgment. Digital Subtraction Angiography (DSA) imaging is commonly utilized both before therapy to identify the occluded vessel and after therapy to evaluate the treatment outcome. Accurate vessel correspondence before and after treatment is crucial for a reliable assessment. To enhance current evaluation methods and address this challenging task, we propose two automated approaches for determining vessel correspondence in pre- and post-EVT DSA imaging. The proposed methods
utilize graphical representations of the cerebral vascular network and distinct matching procedures. We refer to the methods as registration-based vessel matching (RB-VM) and graph based vessel matching (GB-VM). The methods were evaluated using manually annotated data with the RB-VM and GB-VM methods achieving a recall of 82.7% (78.2; 85.7) and 51.3% (47.4; 54.4) respectively. This work marks a significant step towards automatic stroke therapy assessment and showcases the potential benefits of graph based algorithms for this task, paving the way for more reliable and objective treatment assessments.

Ischemic stroke, defined as the sudden onset of a focal neurological deficit resulting from hypo-perfusion due to an occluded cerebral artery, is a leading cause of mortality and long-term disability worldwide. The standard treatment for acute ischemic strokes (AIS) is an endovascular thrombectomy (EVT), a minimally invasive procedure for locally removing the occlusion. The effectiveness of an EVT is assessed by estimating the reperfusion status on digital subtraction angiography (DSA) images using a six-category visual grading system termed the extended Thrombolysis In Cerebral Infarction (eTICI) scale. However, eTICI scoring suffers from inter-observer variability, complicating EVT outcome comparisons. To address this, ‘autoTICI’, a fully automated and objective deep-learning-based TICI scoring method, was developed by our research group. This thesis aimed to improve and evaluate autoTICI in clinical practice.

Chapter 1 provides a general introduction, highlighting the clinical background and relevance of automated TICI scoring.

Chapter 2 presents a pilot study that explored the technical feasibility and clinical implementability of autoTICI within the interventional radiology workflow. The pilot study showed that the performance of autoTICI is currently insufficient for implementation in clinical practice. Nevertheless, with an average System Usability Score (SUS) of 81.3, autoTICI still demonstrated excellent usability within the clinical workflow, indicating its potential for future clinical use. Besides, its potential as a decision-support tool was recognized by multiple interventional radiologists, particularly in complex or high-stress scenarios.

Chapter 3 presents a core-laboratory study assessing the inter-observer agreement and efficiency of eTICI scoring with and without autoTICI feedback, showing a substantial agreement both with and without autoTICI, with weighted kappa values of 0.65 and 0.67, respectively. Additionally, autoTICI had no significant effect on scoring efficiency, suggesting that the supplementary information did not increase the time required for eTICI scoring.

Chapter 4 details the development of an automated deep-learning approach for direct segmentation of vascular territories on cerebral DSA, aimed at improving the robustness of autoTICI by replacing the atlas registration step in its pipeline. A nnUNet model was trained, validated, and benchmarked against the atlas registration method. The proposed model demonstrated excellent performance on a held-out test set, achieving a 100% success rate compared to 52.5% for atlas registration. Additionally, it significantly outperformed the atlas registration method in both segmentation accuracy and computational efficiency.

Chapter 5 provides a general discussion and conclusion, synthesizing the results from the previous chapters and offering recommendations for future research on autoTICI.

In this master’s thesis, autoTICI was successfully integrated into the clinical workflow of the interventional radiology. Through the pilot study and inter-observer analysis, we were able to identify autoTICI’s intended use and added value within the clinical workflow, along with key issues, challenges, pitfalls, and future opportunities. By addressing one of these identified issues — atlas registration — through automated vascular territory segmentation, the performance of autoTICI was significantly improved. Although autoTICI is not yet ready for clinical implementation, with further optimization needed, the contributions of this master thesis have brought it significantly closer to that goal.

Introduction: Trauma-induced rib fractures are a common injury, affecting millions of individuals globally each year. Although anteroposterior thoracic radiographs are part of the standard posttraumatic screening, the most sensitive modality, and therefore golden standard for diagnosing rib fractures, is computed tomography (CT). Still, between 19.2% and 26.8% of rib fractures are missed. Another problem encountered in rib fracture treatment management is the large interobserver variability on their taxonomy. This thesis aims to automate rib fracture detection and improve consistency in their classification by developing a Deep Learning (DL) model, using CT data.

Methods: The rib fractures were classified according to the Chest Wall Injury Society (CWIS) taxonomy, evaluating rib fracture’s type, displacement and location. Furthermore, the ribs were numbered from 1 up to and including 12 from cranio-caudal direction. For the detection and three CWIS labels, three classification models of the nnDetection framework were trained. The rib numbering consisted of a trained nnU-Net segmentation model. The four models were combined to obtain the proposed DCRibFrac model.

Experiments and results: The dataset is composed of retrospectively collected and anonymized CT scans of 100 randomly selected patients (1010 rib fractures) who were admitted to the Erasmus MC following blunt chest trauma. On the internal test set, DCRibFrac achieved a detection sensitivity of 77%, precision of 79%, and F1-score of 78%, with a mean false-positives per scan of 2.26. The type labels had the lowest scores, with sensitivities between 17% and 90%. The displacement labels had sensitivities between 43% and 91%. The location labels had the highest scores, with sensitivities between 88% and 96%. The rib number was correct in 72% of the rib fractures when wrong segmentations were excluded.

Conclusion: The proposed DL model automates acute rib fracture detection and reaches a sensitivity that is on par with clinicians. This model is the first, to the authors’ knowledge, to incorporate the CWIS taxonomy and shows its potential for achieving a consistent classification.

Introduction: Breast reconstruction after a mastectomy is crucial to improve a patient’s quality of life. The Deep Inferior Epigastric artery Perforator (DIEP) flap procedure is considered the golden standard for breast reconstruction. Three-dimensional (3D) visualization methods have shown promise in providing a better understanding of Computed Tomography Angiography (CTA) information and potentially reducing operative time. However, for these methods manual segmentation of the perforators is needed, which is time-consuming and prone to variability. Automated segmentation using Deep learning (DL) can potentially overcome these limitations and provide accurate and efficient segmentation of the DIEP flap perforators. This study aimed to evaluate the application of DL for automated perforator segmentation in DIEP flap breast reconstruction, improving the efficiency and objectivity of DIEP flap perforator segmentation.

Materials and methods: The dataset comprised 25 CTA scans for training and validation, and 5 CTA scans for testing. DL was employed for automated segmentation, and quantitative evaluation included metrics such as Dice coefficient score, recall, precision, surface distance, and centerline overlap. The qualitative evaluation involved grading the clinical acceptability of segmentations by four experienced plastic reconstructive surgeons.

Results: On the training set, a Dice score of 0.58 (± 0.08) and a true positive centerline overlap of 0.66 (± 0.10) were achieved for perforator segmentation. The DL model successfully segmented the intramuscular main branch, but some perforators were missed in the subcutaneous fat tissue. Combined grading by all surgeons showed no statistical difference between manual and automated segmentations and both segmentations were evaluated as clinically acceptable.

Conclusion: Automated DL segmentation holds promise for enhancing the efficiency and objectivity of identifying DIEP flap perforators in CTA images, providing an alternative to manual segmentation. Nonetheless, further research is needed to refine the automated segmentation results and to validate the generalizability and clinical applicability of the DL segmentation approach in larger patient cohorts and different clinical settings.

The complexity and interdependency of hospital activities make it challenging to seamlessly integrate new technologies into clinical routines, especially with various stakeholders involved. Considering contextual factors when implementing innovations has been recognized as critical to success. This highlights the significance of an approach that focuses on integrating technology with the existing healthcare culture and motivating the context to change, rather than focusing on technology alone. Therefore, the aim of this thesis is to provide a method to ensure the clinical context is considered when developing a new technology. This approach is put in practice by utilizing it to introduce augmented reality (AR) at the oral and maxillofacial surgery (OMFS) department of the University Medical Center (UMC) Utrecht, to provide new insights on how to better align the technology with clinical needs and to familiarize and motivate the medical team with the technology. This is done in close collaboration with the UMC Groningen. Currently, the UMC Groningen, UMC Utrecht, and the Center for Information Technology (CIT) of the University of Groningen are developing an AR surgical navigation system. The aim of the introduction is to provide valuable information for the further development of the system and to prepare and involve the clinical environment.
The approach used in this thesis is based on human-centered design and design-thinking. Furthermore, key factors for successful adoption found from previous research are considered. The approach consists of four main components. The first is to assign a med-tech innovation coordinator; someone with both a medical and technical background who will be responsible for driving adoption. Furthermore, it is an iterative three-step process: 1) Gain understanding; 2) Share understanding; 3) Actively engage stakeholders.
The application of this approach to introduce AR at the OMFS department of the UMC Utrecht has resulted in multiple positive effects. In clinical context it resulted in an understanding of the technology and its value, motivations to invest time in the system, and the establishment of a shared future perspective on the implementation of AR. Furthermore, valuable input from a clinical perspective is obtained, including confirmation of valuable system elements and new ideas/insights for the further development of the AR navigation system.
In conclusion, this thesis provides guidance for departments seeking to adopt new technologies based on a practical example. The methodology provides activities that facilitate the early introduction of technologies, ensuring the active involvement of the clinical context.

Purpose. Distal radius fractures are common fractures of the wrist. These fractures are often displaced and need reduction, after adequate reduction, the patients will have follow-up X-rays to check if the fracture stays stable. This is important because surgery might be required if the fracture becomes unstable. This can lead to delayed surgery which can worsen the treatment outcome.
Therefore, it would be valuable to predict which distal radius fractures are likely to become unstable. Machine learning could help predict the stability of distal radius fractures based on CT. In addition, because there is only a small dataset available, this research also studies the effect of different machine learning methods.
Method. Two different methods were evaluated for the stability prediction of distal radius fractures: traditional machine learning (radiomics method) and a residual network with and without transfer learning (deep learning method). For the radiomics method, a python package called WORC was used, which automatically extracts radiomic features and optimizes machine learning models. For the deep learning method, the backbone of a residual network called Med3D with added layers for classification was used.
Results. The radiomics method combined with augmentation, gave the best results (AUC: 0.64± 0.01). Also, it was found that using an augmented data set in the radiomics method resulted in improved performance. This gives a slight indication that the radiomic method can learn to predict DRF when there would be more data available.
Conclusion. The radiomic method is the most promising method for predicting the stability of distal radius fractures, despite the small difference in performance compared to random guessing and the deep learning method. However, for further research, it is highly recommended to acquire a larger dataset.

Wrong-level spine surgery (WLS) is a medical error with severe consequences for the patient and medical staff. The current golden standard to prevent WLS during spinal surgery is fluoroscopy images acquired by a C-arm. However, this approach uses ionising radiation and interrupts the standard surgical workflow. To overcome these disadvantages, ultrasound-based navigation may provide an alternative to prevent WLS during spinal surgery. Baka et al. developed the Lumbar Localiser for ultrasound-based level localisation; however, this has not yet been applied in clinical practice. This project aims to improve and evaluate this level localisation approach in clinical practice.
Chapter 1 highlights the social relevance of ultrasound-based level localisation. Chapter 2 shows in a literature review that navigated ultrasound-based guidance is not yet implemented in clinical practice. Identified recommendations for further research were optimising ease of use, workflow integration, registration accuracy and computation time. Chapter 3 describes a prospective multi-centre study (n=34) which evaluated the accuracy of the level localisation approach and the added value of intraoperative ultrasound acquisition. The accuracy of the improved level localisation approach was too low (53%) to recommend implementation in clinical practice. We showed that pre- and intraoperative ultrasound acquisitions could be integrated efficiently into the operation room (OR) workflow, with clinically negligible time differences compared to the current approach for level localisation. Intraoperative ultrasound acquisition added valuable information, enabling a re-check of the level localisation without the attenuating subcutaneous fat layer between the SP and ultrasound transducer. Chapter 4 describes the development of an automated contour segmentation approach based on magnetic resonance imaging (MRI) using deep learning. A nnU-Net was trained to segment the lumbar spinous process automatically. Subsequently, the posterior contours were extracted based on these segmentations and imported into the Lumbar Localiser. The automatic contours showed a successful matching in all 16 test cases. Chapter 5 provides an overall conclusion and discussion.
In this master’s thesis, the technical aspects and clinical workflow of the Lumbar Localiser were improved to get closer to the clinical application of ultrasound-based level localisation for lumbar surgery. Implemented improvements include adding intraoperative ultrasound acquisition, extension to MRI as a preoperative modality and automatisation of the contour annotation. In conclusion, this approach cannot be applied yet in clinical practice, as an accuracy of around 100% should be achieved to minimise the risk of WLS and thereby prevent irreversible damage to the patient.

The efficacy of endovascular therapy in large vessel occlusion (LVO) of the anterior circulation is dependent to a high degree on the selection of patients who are likely to benefit from this procedure. To this end, functional outcome prediction based on clinical parameters is an active area of research. In the preoperative screening of LVO patients, CT-Angiography (CTA) imaging is commonly acquired. We compare the functional outcome prediction performance of multiple deep learning based classifiers with multiple conventional methods, including the clinically validated MR PREDICTS decision tool. Using a dataset composed of 1929 preprocedural CTA images combined with clinical data, we compare a clinical baseline model with an imaging based pipeline and a combined pipeline. For the imaging model backbone we train various state-of-the-art architectures (Med3D, Vision Transformer, Voxel Transformer). These models are used to predict dichotomized modified Rankin Scale score 90 days after mechanical thrombectomy. Binary classifier outcomes are quantified using Area-Under the receiver operating characteristic Curve (AUC). The activation maps of the best performing image based model are further investigated using the GradCAM++ posthoc visualization method. Combining clinical features with information extracted from CTA images does not significantly improve the performance of functional outcome prediction methods compared to the baseline model. The information extracted from the images does not seem to
be complementary to the clinical features. This multimodal technique can however replace radiologically derived biomarkers, as its performance is non-inferior.

Predicting functional outcome after intra-arterial treatment (IAT) in acute ischemic stroke (AIS) patients is an important aspect of treatment decision making and prognostics. Standard methods for functional outcome prediction after stroke combine baseline clinical (and radiological) parameters. In this study, we investigated to what extent baseline CTA images can be used for the prediction of functional outcome and how this relates to standard scoring methods. Furthermore, it was investigated whether combining baseline CTA images with clinical parameters improved the predictive accuracy compared to outcome prediction based on clinical parameters. We proposed two network architectures, a convolutional neural network (CNN) for the processing of image data and a multilayer perceptron for the processing of clinical (and radiological) parameters. Various training strategies were applied for the fusion of image and clinical data. The CNN processing CTA images achieved an average cross-validated area under the curve (AUC) score of 0.67, which was lower than for models processing clinical (and radiological) parameters. The best performing model combining CTA images and clinical parameters was trained end-to-end and applied weight initialization of the pre-trained CNN (AUC = 0.78). The DeLong test showed that the combined model performed significantly better than the model processing clinical parameters (AUC = 0.75). However, the difference is small and might not be clinically relevant. Compared to scoring methods processing clinical and radiological parameters the combined model achieved similar performance.

A craniotomy is a procedure were a neurosurgeon has to open the skull to gain direct access to the brain. When a brain tumor has to be removed from a patient, the craniotomy position is of great importance. This mostly defines the access path from the skull surface to the tumor and thus also what healthy brain tissue will be removed to access to the tumor itself. To minimize the amount of important brain structures that are removed, the craniotomy has to be carefully planned. This is a complex procedure, where a neurosurgeon is required to mentally reconstruct spatial relations of important brain structures to avoid these as much as possible.

We propose a visualization using augmented reality which may assist in the planning of a craniotomy. In this visualization the goal is to show important brain structures aligned with the physical position of the patient. This should allow better perception of the spatial relations of these structures and thus assist the neurosurgeon. Additionally to the visualization of the structures, we created a heat map that is projected on top of the skull. This should give a quick overview of in which areas there are many important structures between the tumor and the skull surface, and should therefore be avoided.

User studies were conducted amongst neurosurgeons and surgeons from other fields to evaluate the proposed visualization. We found that many of the participants indeed thought that the visualization can assist in surgery. For the specific case of craniotomy planning, several improvements have to be made on the heat map before it can be useful. Nevertheless, the visualization of the structures in itself can assist neurosurgeons in the planning of a craniotomy. Although more work has be performed at practical aspects of the visualization to make it ready for clinical experiments.

Percutaneous coronary intervention is a minimally-invasive procedure to treat coronary artery disease. In such procedures, X-ray angiography, a real-time imaging technique, is commonly used for image guidance to identify lesion sites and navigate catheters and guide-wires within coronary arteries. Due to the physical nature of X-ray imaging, photon energy undergoes absorption when penetrating tissues, rendering a 2D projection image of a 3D scene, in which semi-transparent structures overlap with each other. The overlapping structures make robust information processing of X-ray images challenging. To tackle this issue, layer separation techniques for X-ray images were proposed to separate those structures into different image layers based on structure appearance or motion information. These techniques have been proven effective for vessel enhancement in X-ray angiograms. However, layer separation approaches still suffer either from non-robust separation or long processing time, which prevent their application in clinics.
The purposes of this work are to investigate whether vessel layer separation from X-ray angiography images is possible via deep learning methods and further to what extent vessel layer separation can be achieved with deep learning methods.
To this end, several deep learning based methods were developed and evaluated to extract the vessel layer. In particular, all the proposed methods utilize a fully convolutional network (FCN) with two different architectures (Appendix A and Chapter 2), which was trained by two different strategies: conventional losses (Appendix A and L1 method in Chapter 2) and an adversarial loss (AN +L1 method in Chapter 2).
The results of all the methods show good vessel layer separation on 42 clinical sequences. Compared to the previous state-of-the-art, the proposed methods have similar performance but runs much faster, which makes it a potential real-time clinical application. Both the L1 method and AN + L1 method in Chapter 2 achieve better background than the method proposed in Appendix A, which can remove catheter and other tubular structures well. On the other hand, the L1 method results in better contrast and clearer backgrounds than the AN+L1 method.

Ablation is a medical procedure to treat liver cancer where a needle-like catheter has to be inserted into a tumor, which will then be heated or frozen to destroy the tumor tissue. To guide the catheter, Ultrasound(US) imaging is used which shows the catheter position in real time. However, some tumors are not visible on US images. To make these tumors visible, image fusion can be used between the inter-operative US image and a pre-operative contrast enhanced CT(CECT) scan, on which the tumors are visible. Several methods exist for tracking the motions of the US transducer relative to the CECT scan, but they all require a manual initialization or external tracking hardware to align the coordinate systems of both scans. In this thesis we present a technique for finding an initialization using only the image data. To achieve this, deep learning is used to segment liver vessels and the boundary of the liver in 3D US images. To find the rigid transformation parameters, the SaDE evolutionary algorithm was used to optimize the alignment between the blood vessels and the liver boundary between both scans.