Introduction and Research Goal: Epilepsy is a common neurological disorder that severely impacts patients’ quality of life. Current diagnostic standards rely on the presence of seizures or interictal epileptiform discharges (IEDs) in the electroencephalogram (EEG). However, some patients who are ultimately diagnosed with epilepsy do not present with seizures or IEDs on their initial EEG, which delays their diagnosis and appropriate medical treatment. Previous studies by Thangavel et al. 2022 and Mirwani 2024 suggest the potential of machine learning methods applied to IED-free EEGs for classifying epilepsy. The aim of this study is to evaluate whether adding clinical characteristics and visual EEG interpretation to existing machine learning models based on quantitative EEG improves the performance of these models for the identification of epilepsy in IED-free EEGs. Additionally, the study explores model interpretability to promote clinical application.
Methods: We focus on subjects who presented at the emergency room following a first clinical seizure and were not diagnosed with epilepsy based on their initial EEG. Ten quantitative EEG feature sets based various mathematical transforms were readily available from previous research by Mirwani 2024. EXtreme Gradient Boosting (XGBoost) models were trained using Leave-One-Subject-Out Cross-Validation (LOSO CV) on these feature sets as benchmark models. Model performance was assessed using the Area Under the Curve (AUC). Clinical and EEG report features were added individually to each quantitative EEG feature set to evaluate their added value to model performance. The best performing clinical and report features for each quantitative EEG feature set were determined using a two-fold grid search, and significance was tested via Welch’s t-test. Ensembles were created from the best performing models of each quantitative EEG feature set including their best performing clinical and report features. SHapley Additive exPlanations (SHAP) and XGBoost feature importance were used for model interpretation, while Bayesian statistics were applied to gain insight into clinical implementation.
Results: The best performing clinical and report features varied across EEG feature sets and did not consistently yield significant performance gains. Only the addition of EEG background to graph metrics of phase lock values showed a significant increase in model performance. SHAP analysis identified residual focal sharp activity as the primary contributor to this improvement. Combining individual models into ensembles substantially improved performance, achieving AUCs up to 0.870. To align model performance interpretation to the guidelines of the ILAE, sensitivity at P(posterior) ≥ 0.6 was proposed as a key evaluation metric. The best XGBoost model with clinical and report features achieved 0.48 sensitivity at P(posterior) = 0.6, while the best ensemble attained 0.81.
Conclusion: Incorporating clinical and report features into XGBoost models based on quantitative EEG data does not consistently improve the detection of epilepsy based on IED-free EEGs. The variability of the best performing clinical and report features across quantitative EEG feature sets suggests that their impact is dependent on the quantitative EEG feature sets. Combining individual models into ensembles significantly enhances performance, achieving a sensitivity of 0.81 at P(posterior) ≥ 0.6. However, external validation is required to confirm these findings. Future models should assess model performance according to the sensitivity at P(posterior) ≥ 0.6 to comply with ILAE guidelines.

Automated diagnosis of epilepsy for differentiating epileptic EEGs without Interictal Epileptic Discharges (IEDs) from normal EEGs remains a critical challenge in clinical settings.
Current state-of-the-art methods use algorithms that can effectively detect epilepsy seizures which improves the current treatment methods for people suffer from epilepsy. Electroencephalograms (EEGs) analyzed by neurologists which are not able to meet the criteria are further looked into to obtain an efficient classification. This thesis presents an automated epilepsy diagnosis approach using a multi-algorithmic feature extraction pipeline. The final models include the development of a robust multi-processing feature extraction pipeline, the application of advanced machine learning / deep learning algorithms, and the validation of the proposed methods on comprehensive EEG datasets. The results, achieved using an XGBoost classifier with leave-one-subject-out (LOSO) cross-validation, demonstrate comparable performance to state-of-the-art epilepsy detectors. The study emphasizes the detection of epilepsy without IEDs, optimizing models through nested cross-validation, and evaluates their performance on the Temple University Hospital (TUH) and Erasmus Medical Center (EMC) Rotterdam datasets.

This report presents the design process of a project aimed at the automatic recognition of 3D structures formed by magnetic spheres in a turbulent water-filled cylinder. This field of research holds promise for future technologies, as macroscopic self-assembly might be the key to three-dimensional storage on chips. With the macroscopic setup, the microscopic self-assembly process is imitated. To automatically recognise a 3D structure, this report is divided into three subgroups that research the optimal test setup, develop an image processing program and create a deep learning model. A labelled result of the 3D formation will be outputted, all while limiting the data size, computation time and inaccuracies. The subgroup responsible for the setup and the underlying physics produces images of the magnetic spheres forming a structure in the test setup. The Image Processing subgroup extracts the properties of the spheres from the image. Finally, the subteam for deep learning, in combination with data management, gives the extracted properties as input to a neural network model, which determines the structure of the spheres. Each submodule has demonstrated successful functionality on its own. However, due to time constraints, a fully integrated system with high accuracy has not been achieved yet. Future work will involve expanding the dataset to enhance the robustness of the recognition algorithms.

Intense precipitation can have extensive economic outcomes, from disrupting outdoor activities to causing severe infrastructural damage, such as landslides, and endangering public safety. The urgency to mitigate these impacts underscores the need for improved early warning systems. Enhanced short-term weather prediction, or nowcasting, is critical for addressing these severe weather events effectively. Traditional meteorological forecasting methods, while foundational, are often constrained by simplistic physical assumptions and fail to capture the complex, nonlinear patterns of intense weather events. These methods also struggle with high computational demands and lack the resolution needed to detect crucial microscale atmospheric phenomena for accurate short-term forecasts.
To address these challenges, this research introduces a novel deep learning approach utilizing a streamlined architecture that combines a Vector Quantized Variational Autoencoder (VQVAE) and an Autoregressive (AR) Transformer. This model aims to predict weather conditions up to 180 minutes ahead, using data analyzed at 30-minute intervals. The proposed model displays comparable performance with the state-of-theart conventional methods and other deep learning nowcasting models in predicting precipitations and sometimes extreme events. This study seeks to enhance forecasting accuracy and efficiency, providing valuable contributions to the field of meteorological nowcasting.

This report serves to finalize the bachelor graduation project on the topic of self-supervised federated learning, specifically the implementation of the algorithms in Python. The goal of the project is to implement a self-supervised learning setup in a decentralized approach using Field-Programmable Gate Arrays (FPGAs) for the processing of data. This serves as a proof of concept that decentralized machine learning on unlabeled data using FPGAs is possible. Multiple algorithms based on the literature were considered to allow for a low-profile learning setup, with simplifications done to be able to reduce the compute required. The results are promising: scaled-down models that can run on an FPGA show that self-supervised learning functions as expected from the theory. By decentralizing the computations increases in performance are possible in favorable conditions. The authors hope that the concept of self-supervised federated learning can be employed to FPGAs on a larger scale to help in the processing of the abundant yet underutilized unlabeled data present at the edges of information networks.

This thesis serves to finalise the bachelor graduation project on the topic of self-supervised federated learning, specifically the on-chip implementation of the algorithms. The goal of the project is to implement a self-supervised learning setup in a decentralised approach using Field-Programmable Gate Arrays (FPGAs) for the processing of data. In this thesis, we endeavour to illustrate the possibility of employing FPGAs to move the fairly compute-intensive self-supervised learning algorithms to the edge. We have developed a number of modules that can accelerate key algorithmic blocks that underlie the major bottlenecks of the classical application of the algorithms and showcase prospective results, which are extensively discussed afterwards to pave a clear path towards truly autonomous and efficient edge-intelligence.

Introduction: Heart failure (HF) poses a significant burden on public health. This can be largely attributed to recurrent hospitalizations in consequence of HF decompensation. Detection of early signs of impending fluid retention may facilitate timely medical intervention and thereby prevent hospitalizations. Monitoring of Cardiac Implantable Electronic Devices (CIEDs)-derived parameters has been proposed as promising solution, as the sensor inherent in CIEDs provide the ability to continuously monitor physiological signals. The aim of this study was to develop personalized machine learning (ML) models that can identify upcoming HF decompensation based on CIED-derived parameters. Methods: Two ML models, a support vector classifier (SVC) and an extreme gradient boosting (XGBoost) model, were developed for all patients. Features known to be associated to HF decompensation were extracted from daily CIED data. The output of the models is the daily classification of the patient’s HF status, either ‘stable’ or ‘unstable’. Model performance was evaluated through area under the precision-recall curve (AUPRC). First, the models were tested on a development dataset with leave-one-out cross-validation, and subsequently on an independent test set. Results: In total, for 62 patients two models were developed. The average AUPRC on the independent test set of the XGBoost models was 0.63 ± 0.28 and of the SVC models was 0.57 ± 0.26. Finally, for each patient, the model that resulted in the highest AUPRC was selected. The final models achieved an AUPRC on the independent test set of 0.61 ± 0.28. Conclusion: The findings of this study show promising results for the use of personalized CIED-derived models. However, significant variability in model performance across patients highlight the need for further research.

Introduction
Opioids are vital for pain management but are highly addictive and may lead to opioid-induced respiratory depression (OIRD), which is the primary cause of death related to both prescription and illicit opioid use. This study employed unsupervised machine learning (ML) to examine potential changes in cluster patterns post-opioid administration and their relationship with respiratory depression levels. Additionally, comprehensive data analysis was conducted based on questions that resulted from observations of cluster behaviour.

Methods
Following preprocessing, a preliminary study evaluated three different models, each trained on the baseline and post-opioid epochs, aiming to identify the most efficient distance metrics and feature space combinations. These models included principal component analysis with Euclidean distance, computed feature space with Euclidean distance, and time series with dynamic time warping. Next, the preferred approach, determined by testing different hypotheses regarding the desired cluster behaviour, underwent further refinement, and predictions were generated for post-antagonist epochs. Finally, feature influence was determined and questions were identified for the data-analysis.

Results
In total, 34 subjects were included, resulting in a total of 6630 epochs for model development. Based on the preliminary study, it was decided to opt for a fuzzy clustering approach using calculated features as input, resulting in membership values indicating the probability of an epoch belonging to a certain cluster. A change in membership values was observed post-opioid as well as a recovery to baseline values post-antagonist in the majority of subjects when clustering was obtained for each subject individually. However, it was expressed as a sudden switch followed by a prolonged plateau phase rather than a gradual transition which was expected. Conversely, when trained collectively over all subjects, the majority of subjects showed no difference, probably due to the presence of inter-subject variability. Nevertheless, SHAP value analysis identified the same feature behaviour among subjects, despite variations in orientation and positioning, hinting at the potential for adjustment using static features. Yet, no significant correlations were found between static features and feature behaviour within this study.

Conclusion
In this study, a fuzzy clustering model was implemented, incorporating SHAP value analysis to enhance the interpretability of the clustering results. Although the model successfully identified changes in respiratory flow patterns associated with OIRD and subsequent recovery after naloxone administration, it requires further refinement.

The thesis explores an innovative technique for enhancing the precision of short-term weather forecasts, particularly in predicting extreme weather phenomena, which present a notable challenge for existing models such as PySTEPS due to their volatile behavior. Leveraging precipitation and meteorological data sourced from the Royal Netherlands Meteorological Institute (KNMI), the research innovates through the development of a physics-informed neural network. Central to this approach is the implementation of a Physics-Informed Discriminator GAN (PID-GAN), a method that embeds physical principles directly into the adversarial training regime. The architecture is marked by the integration of a Vector Quantization Generative Adversarial Network (VQ-GAN) and a Transformer as the generator, complemented by a temporal discriminator as the discriminator component. Results from this study indicate a notable advancement over traditional numerical weather prediction and cutting-edge deep learning models, underscoring the PID-GAN model's superiority in delivering accurate precipitation nowcasting metrics.

Extreme precipitation, like floods and landslides, poses major risks to safety and the economy, underscoring the need for sophisticated weather forecasting to predict these events accurately, enhancing readiness and resilience. Nowcasting, which uses real-time atmospheric data to predict short-term weather, is key in addressing this challenge. Traditional nowcasting systems, reliant on extrapolation from rainfall radar observations and constrained by simplistic physical assumptions, often struggle to detect complex, nonlinear weather patterns. This gap has opened the door for deep learning models, which have shown significant promise in improving the accuracy and reliability of short-term weather predictions, making them a focal point of recent research and the basis of this thesis's approach.

This thesis introduces a deep generative model designed for the nowcasting of extreme precipitation events up to 3 hours ahead, utilizing a Vector-Quantized Variational Autoencoder (VQ-VAE) to compress radar data into a low-dimensional latent representation, and an Autoregressive Transformer for predicting future radar images. Additionally, a binary classifier works in conjunction with the Autoregressive Transformer to identify extreme versus non-extreme weather events, using these classifications to inform an Extreme Value Loss (EVL) function. This loss function aims to improve the accuracy of predicting extreme weather events by addressing the data imbalance between normal and extreme precipitation occurrences. The proposed model displays comparable performance with the state-of-the-art conventional methods and other deep learning nowcasting models in predicting extreme events.

To address the challenges of manual egg sorting on poultry farms, such as labor intensity and
inconsistent quality standards, Moba is developing a machine for automated egg candling
using computer vision. As part of the prototype phase, this project evaluates various methods
to process the egg images for this sorting task.
This project presents an algorithm that segments the individual eggs within the camera’s
field of view, classifies them into five categories, and tracks them across multiple images
to perform the outlet control prediction. To enhance computational efficiency, an anomaly
detector is introduced, directing classification resources exclusively to eggs with a high like-
lihood of defects.
For segmentation, the number of channels in the YOLOv8n model was significantly re-
duced, resulting in YOLOv8xn. With a 97.2% AP95, this model outperformed a classical
algorithm using domain-specific information, as well as more complex deep learning meth-
ods, such as YOLOv8n and the prompt-based FastSAM and MobileSAM models.
Various low-complexity deep learning models were evaluated and compared to classical
feature extraction combined with ensemble machine learning methods for the classification
task. Balancing accuracy and runtime, MobileNetV3 proved to be the optimal choice, achiev-
ing 95.6% accuracy on the final test set.
For anomaly detection, both classical methods and the autoencoder reconstruction loss
were evaluated. Modeling the edge count of non-defective eggs with an exponential distribu-
tion to set a threshold reduced the computational load for classification by 83%. Furthermore,
using the anomaly detector the batch size can be reduced from 32 to 8 images.
To send the eggs to the correct outlet, the predictions for each egg were combined using
thresholds on count-per-class and normalized cumulative class predictions, both without and
with anomaly detection enabled. This resulted in overall accuracies of 97.3%, 98.4%, and
98.0%, respectively.

Deep-learning-based object detectors, while offering exceptional performance, are data-dependent and can suffer from generalization issues. In this thesis, we investigated deep neural networks for detecting people and medical instruments in the vision-based workflow analysis system inside Catheterization Laboratories (Cath Labs). The central problem explored in this thesis is the fact that the performance of the detector can degrade drastically if it is trained and tested on data from different Cath Labs.

Our research aimed to investigate the underlying causes of this specific performance degradation and find solutions to mitigate this issue. We employed the YOLOv8 object detector and created datasets from clinical procedures recorded at Reinier de Graaf Hospital (RdGG) and Philips Best Campus, supplemented with publicly accessible images. An aggregated version of object detection metrics was created for multi-camera system evaluation. Through a series of experiments complemented by data visualization, we discovered that the performance degradation primarily stems from data distribution shifts in the feature space. Notably, the object detector trained on non-sensitive online images can generalize to unseen Cath Labs, outperforming the model trained on a procedure recording from a different Cath Lab. The detector trained on the online images achieved an mAP@0.5 of 0.517 on the RdGG dataset. Furthermore, by switching to the most suitable camera for each object, the multi-camera system can further improve detection performance significantly. An aggregated 1-camera mAP@0.5 of 0.679 is achieved for single-object classes on the RdGG dataset.

Optical Character Recognition (OCR) is a pivotal technology used to extract text information from images, finding wide-ranging applications in document digitization and medical records management. The integration of machine learning has ushered in an era of swift and precise OCR models. Broadly, OCR comprises two key components: detecting the bounding boxes around text instances and recognizing the characters within them. Presently, prevailing OCR models are primarily intricate two-stage systems necessitating real-time operation on remote servers. Nevertheless, end-to-end models exhibit superior performance from a data utilization perspective. There exist scenarios where offline models prove indispensable, such as in environments with restricted internet access or locales with stringent data privacy and security requirements.

This project delves into various end-to-end models, leveraging the PaddleOCR end-to-end model as a foundational reference to devise a compact OCR model tailored for edge devices. Through meticulous optimization of the backbone architecture and the introduction of diverse Feature Pyramid Network (FPN) structures within the stem network, we achieved a remarkable reduction in model size, down to 19MB. This represents a substantial advancement, constituting merely one-tenth of the original PaddleOCR end-to-end model's footprint.

By leveraging an extensive database and conducting a series of fine-tuning experiments specifically tailored for end-to-end OCR tasks involving curved text images, the model exhibits an impressive precision rate of 47.3% and an f-score of 45.3%. This achievement highlights the effectiveness of the customized loss function relative to the original model, despite its reduced size. Notably, this performance is comparable to certain end-to-end models with larger backbones. Furthermore, an Android demo has been carefully developed to demonstrate the model's capabilities on mobile devices, achieving an average processing time of 433 milliseconds per image.

The rise of streaming and video technologies has underscored the significance of efficient access and navigation of digital content, particularly for scholars in fields like history and art. Scholars actively seek streamlined approaches to index, retrieve, and explore digital content, with a focus on locating specific instances. The process of searching for specific instances in video search is complex that requires the analysis of video sequences and the identification of relevant video segments. Advanced techniques and algorithms are necessary to ensure effective content-based retrieval of the required information.

In response to the escalating demand for accurate and swift access to relevant visual data within the vast spectrum of video resources, our research has been dedicated to the development of novel, efficient content-based image retrieval methods tailored for videos by integrating deep learning methodologies. Our comprehensive system contains two crucial components: keyframe extraction and content-based image retrieval. Keyframe extraction involves identifying significant frames within videos, while content-based image retrieval enables the retrieval of similar frames to a query image through feature extraction and ranking.

A unique aspect of our research lies in the exploration and analysis of a diverse range of feature extraction techniques derived from compact deep learning networks. We have compared our proposed method with state-of-the-art retrieval systems, evaluating performance metrics in terms of both accuracy and speed. Our method harnesses the power of compact deep learning network features in the initial ranking stage, effectively sublisting frames, and subsequently introduces re-ranking using a larger network. This innovative approach promises to deliver the best of both worlds: exceptional efficiency without compromising retrieval accuracy.

Anesthesia-related hypotension is a significant concern during surgery, occurring shortly after induction and potentially leading to severe complications. Since the anesthetic drug is believed to have an important role in the occurrence of post-induction hypotension (PIH), anesthesiologists now advocate for the appropriate selection of anesthetics dosage to avoid PIH.To facilitate such decision-making, an accurate prediction of PIH associated with a certain dosage of anesthetics is necessary. This thesis presents a high-accuracy prediction model for PIH that supports anesthesia decision-making. The model is trained on data from the VitalDB database of 320 patients undergoing general anesthesia. The target output of this classification model is the occurrence of PIH, as defined through comprehensive analysis that incorporates clinical operations. Besides demographic data and vital signs, our model incorporates the dosage of propofol administered during the induction period as an input variable, mimicking real-world anesthetic plans. By employing the model in the target control infusion system of anesthesia, the anesthetics dosage can be varied as input, providing outcome predictions as security suggestions. An ensemble algorithm is employed to balance the prediction performance and the ability to elucidate the positive relationship between propofol and PIH risk, forming an anesthetics advice model. Compared to previous PIH prediction studies, our prediction model is validated in more reliable nested cross-validation approach and achieves a higher performance (precision of 0.83 and recall of 0.84). We believe utilizing demographic and dynamic vital signs to predict HIP can be useful in determining the appropriate anesthetic dosage plan, offering potential improvements in patient care and safety.

The advent of streaming and video has sparked a revolutionary shift in the presentation of materials across various fields, such as history, art, and media copyright protection. In this context, scholars and rights holders are seeking efficient solutions to index, retrieve, and browse through digital content searching for a specific instance. Unlike searching a specific instance in an image, searching in a video requires more than analyzing the visual features of an image and then comparing these features to a database, for it includes processing video sequences and retrieving video segments.

Motivated by the urgent need and promising applications across diverse disciplines, we present a novel deep-learning-empowered content-based video image retrieval (CBVIR) system with a strong emphasis on real-world applications. This system offers high efficiency and considerable accuracy, addressing the challenges associated with accessing and utilizing video materials effectively. Our initial approach revolves around the extraction of informative keyframes that effectively capture essential objects within the video. This process, known as Key Frame Extraction (KFE), enables us to distill the most crucial visual representations for further analysis. After the extraction of keyframes, the relatively smaller dataset allows for content-based image retrieval (CBIR) to be conducted, retrieving similar images from a database solely based on the content of the query image. In this project, a wide range of methods are investigated and analyzed, including traditional representations, handcrafted feature extraction methods, and up-todate machine learning-based image representations. Our contribution is striking a balance between high-level and low-level image representations for this task. Targeting efficiency improvement, enhanced color-based features together with dynamic clustering KFE module is proposed and implemented, achieving high efficiency ratio and satisfactory accuracy. While targeting accuracy, a traditional and deep learning-based hybrid feature is proposed, achieving valid efficiency ratio and highest accuracy. Overall, an automatic retrieving system requiring much less user engagement is provided, together with a system GUI prototype。

Determining ground level elevation is essential in assessing flood risk. This report presents a com-prehensive framework using state-of-the-art object detection model YOLOv8 trained on a relatively small dataset containing houses, doors and steps. The object detection model is followed up by a post-processing block designed specifically for ground-level elevation assessment. The ground-level elevation is calculated by means of counting the number of steps leading up to the front door of a house. The final system has an precision of 0.908 for the detection of steps and an RMSE value of 1.19 steps per image, based on a test dataset containing 111 images. The report starts with a program of requirements, followed by an overview of the framework. Then a short evaluation of the state-of-the-art YOLOv8 object detection model and a describtion the met-
rics used are outlined. An extensive description of the post-processing block, with design choices substantiated by visualized results follows. The final results are presented, discussed and compared to the program of requirements. The report concludes by highlighting potential areas for further research and improvement, such as exploring larger and more diverse datasets, investigating alter- native object detection models, and refining the post-processing techniques. Overall, this study presents a robust framework for ground level elevation assessment, leverag-ing the capabilities of YOLOv8 and most importantly incorporating tailored post-processing tech-niques. The findings provide valuable insights for flood risk assessment and contribute to the broader field of geospatial analysis.

Cardiac output (CO), a vital hemodynamic parameter that reflects the blood volume pumped by the heart per minute, is crucial for determining tissue oxygen delivery and the heart's ability to meet the body's demands. Researchers have developed various methods to measure cardiac output, including thermodilution using pulmonary artery catheters (PAC), also called Swan-Ganz catheters, the gold standard for cardiac output measurements. Such an approach involves an invasive procedure associated with complications, and it requires specialized equipment and expertise, limiting its use to critically ill patients undergoing operations in intensive care units (ICUs). An alternative, less invasive way to estimate CO is by analyzing arterial blood pressure (ABP) waveforms. However, the relationship between cardiac output and blood pressure is unknown. This study uses machine learning and feature engineering techniques to discover the relationship between CO and ABP. We applied the sparse identification non-linear dynamics (SINDy) algorithm to discover features that significantly contribute to the relationship between CO and ABP. Additionally, we investigated the optimum number of cardiac cycles required for feature extraction to achieve the best performance providing insights into the temporal dynamics of CO estimation. The proposed approach achieved clinically acceptable performance regarding radial limits of agreement and bias. Further, the proposed approach was validated on an external dataset and achieved comparable performance. Finally, the learned model was interpreted as a differential equation describing the blood flow where CO acts as an external force to the system. All materials used in this study, including code, model, raw data, processed data, and extracted features, are available on GitHub to facilitate further development.

This thesis presents a comprehensive analysis and implementation of a program designed to detect the number of steps leading up to the front door of a house using a Google Street View Image. The purpose of counting the steps is to have a proxy of the ground floor elevation by multiplying with an average step height. After determining the street level’s elevation from sea level, it becomes possible to assess the vulnerability of houses to flooding. Additionally, the system developed in this thesis provides a method for enhancing the effective resolution of computer vision models to be able to detect details with more accuracy, this is done through Region of Interest detection and enlargement.
The design process consists of four main stages: Acquiring and labelling data consisting of street view images which contain houses, staircases, and steps. Developing an initial model to gauge the ability of the current technology available. Developing algorithms to detect objects of interest in the images. Use a top-level to combine these algorithms and crop out any information that is not of interest.
In the structure of the final product, a top-level, that allowed a model to select a region of interest for step-detection and then cropped the image to this region of interest, was used. This approach allowed for an effective enhancement in resolution as the model is allowed to only focus on the ’useful’ information. The detection in these models was done by YOLOv8x-algorithm that was transfer learned on the custom dataset.
The final product had a precision of 98.7% in detecting steps, an area under the curve (AUC) of 98% in the PR-curve for steps and a deviation no larger than 1 step.

Catheterization Laboratory (Cath Lab) is a hospital examination room equipped with diagnostic imaging equipment to display the heart arteries and to detect and treat any abnormality or stenosis. This thesis addresses the urgent need to improve the efficiency of operating rooms in Cath Labs due to the increasing number of patients with cardiovascular disease. Currently, measuring key performance indicators (KPI) for efficiency in Cath Labs requires manual observation and registration of patient entry and exit times, as well as the establishment and closure of access points. This process is time-consuming, costly, and prone to errors. To overcome these challenges, we conducted a study at the Reinier de Graaf Groep Hospital in Delft, where we recorded activities in the Cath Lab and developed a video-based event detection framework to automate the identification of these crucial events. The proposed framework consist of a fine-tuned YOLO v8 object detector, a popular object tracker ByteTrack, a post processing step and different event detection algorithms for different events of interest. This thesis presents the related video-based event detection framework and a detailed introduction for the development of object detector YOLO and object tracking technologies. The materials used in the project and how they work are clearly described. Events of interest are determined as needed for KPI calculations and are defined by describing the workflow in the target Cath lab. The steps taken to prepare the event dataset are also described in detail. The process of training, the connection between the tracker and the detector and the algorithms to detect different events for reproduction are explained. A series of performance tests on the target detector and tracker concluded that the tracker can handle the occlusion to some extent, but the performance of the object detector determined the performance of the tracker. It was also found that the simpler network structure in YOLOv8 outperformed the more complex network structure in detecting smaller objects on our dataset. In addition, the event detection algorithm is tested on multiple videos to evaluate its robustness and result showed that the algorithm can correctly and accurately detect the target events. The practical applicability of the framework and some future work that could enhance its performance are discussed. In conclusion, this thesis presents a video-based event detection framework that can detect events of interest in the Catheterization Lab at Delft Reinier de Graaf Groep Hospital.