The aim of this study is to develop and evaluate software for detecting eye features based on standardized video recordings, with the ultimate goal of integrating it into the existing Philips Vital Signs Camera. The software consists of three modules: eye and iris detection, blink detection and pupil detection. It was developed using open-source libraries, with MediaPipe FaceMesh as the primary component. To validate the eye/iris and blink detection modules, the VSC-MEDlib dataset was used, which had previously been employed to evaluate the Philips Vital Signs Camera. For pupil detection, video recordings of a subject under various lighting conditions were analyzed.
For eye detection, 6 out of 7 test IDs achieved a 100% recognition rate. For iris detection, 5 out of 7
test IDs achieved a 100% recognition rate. Test ID 7 reached a 99% recognition rate. For both eye
detection and iris detection, the face mask condition showed a significant deviation (p < 0.01), partic-
ularly for individuals with darker skin colors (p < 0.1). After parameter optimization, the Bland–ltman
analysis showed that 84 out of 90 data points (93.3%) fell within the 95% limits of agreement, indicating good overall agreement between the software and the true blink counts, with a small number of outliers beyond the expected limits. Pupil detection accurately identified pupil position but consistently underestimated pupil size, especially when the pupil was small. Nevertheless, the software was able to detect appropriate pupil size changes in response to lighting conditions and demonstrated sensitivity to the pupillary light reflex. These findings indicate that reliable measurements of eye features can be obtained from standard video recordings. This opens up possibilities for non-invasive and remote patient monitoring, with potential integration into existing systems such as the Philips Vital Signs Camera. Future research could explore the combination of eye-tracking and neuromonitoring to expand the range of accessible physiological metrics.

This study investigated the feasibility of deep learning-based segmentation in intra-abdominal kidney ultrasound registration, to enable image-guided robotic-assisted partial nephrectomy (RAPN). Two state-of-the-art models, DeepLabV3+ and SAMUS, were trained and evaluated using a novel intra-abdominal kidney ultrasound (IAKUS) dataset of 2,265 images from 15 RAPN patients. Moreover, a transfer-learning approach was adopted using the publicly available open kidney ultrasound (OKUS) dataset for pre-training. Results showed that SAMUS consistently outperformed DeepLabV3+ across all metrics, achieving an average Dice score of 88.0±2.0% and Hausdorff distance of 13.7 ± 3.8 mm, consistent with literature. SAMUS was pre-trained on ∼ 30k ultrasound images which enabled a zero-shot test, outperforming trained DeepLabV3+ configurations. Furthermore, no measurable difference was seen between OKUS and IAKUS training. Both findings suggest ultrasound specific features may be more important than organ specific features for training, and data diversity may be more important than strict anatomical similarity. The SAMUS model obtained a registration accuracy of 4.3 ± 2.8 mm and inference time of 4.35 fps, in line with literature reported clinical feasibility. The target registration was even improved by an average of 2.6 ± 4.2 mm in 11/13 patients compared with manual-based registration. This proves that deep learning-based registration is not only feasible in a clinical setting but exceeds manual-based registration.

Objective: This work focuses on the feasibility of using Electrical Impedance Spectroscopy (EIS) to differentiate between cancerous and healthy tissues in real-time alongside the GI tract. This is to see whether it is possible to support surgeons when taking biopsies, or to see if it is possible to make biopsies obsolete with EIS. Methods: The study is limited to three tissue types representing the GI tract. These tissues are the esophagus, ileum, and colon tissue. Impedance measurements are taken with a 4-electrode probe on a Hewlett Packard 4192A LF HP Impedance Analyzer. All measurements are taken over a frequency range of 1 kHz to 7 MHz in 300 steps. Seven patients are included: three esophagus, two
leum, and two colon. Forty-eight ex vivo measurements are taken; 32 are on healthy tissue and 16 on cancerous tissue. Almost all measurements are verified by histological assessment (golden standard). In this work, a combination of parameterization with a classification method is made to create a classification strategy. These can be listed as the Cole impedance model, the two-pole Cole impedance model, the two-pole Cole impedance model in combination with a
Constant Phase Element (CPE), and PCA. The Cole impedance models are combined with thresholding, and the PCA is combined with a SVM. Results: The thresholding algorithm is created in combination with the α2 parameter (P < 0.05) from the two-pole Cole impedance model in combination with a CPE. The thresholding value is determined in a LOOCV approach via ROC curves created to search for the threshold that gives the most significant summation of sensitivity and specificity. In the combination of PCA and SVM, the PCA uses three principal components (containing 96.37% of the total variance), and the SVM applies an Radial Basis Function (RBF) kernel. For the 2-pole Cole impedance model with CPE
in combination with thresholding, we find an accuracy 0.5208, sensitivity 0.4375, specificity 0.5625, Positive Predictive Value (PPV) 0.3333, Negative Predictive Value (NPV) 0.6666, and Mathews Correlation Coefficient (MCC) 0.0000. For the PCA in combination with the SVM, we find an accuracy of 0.4167, a sensitivity of 0.0000, specificity of 0.6250, PPV of 0.0000, NPV of 0.5556, and MCC -0.4082. Conclusion: It is concluded that we cannot create an algorithm that either supports surgeons or replaces a biopsy in the GI tract with the current setup. This is highly likely due to the extremely low amount of data that is included to train
the algorithm.

Throughout history, different intraocular illumination methods for vitreoretinal surgery have been used and the most common limitations to the state-of-the-art intraocular illumination methods are the glare and reflections created by the illumination device and the relatively small illuminated area. A way to tackle these limitations could be by using a ring-light configuration, which produces diffuse scattered illumination for a full 360° around the object.
In this study, we developed and experimentally evaluated two ring-light configurations for intraocular illumination during vitreoretinal surgery: one for insertion through the pars plana (20 mm), and one for insertion through the cornea and placement in the posterior chamber (11 mm). Both prototypes are made from a 27 gauge Eckardt TwinLight (DORC, the Netherlands), placed inside a capillary tube cut in half through the length. Cuts were made in the fiber every 2 mm, perpendicular to the length of the fiber, to allow light to exit the fiber through the cladding.
By measuring the light intensity and illuminated area, we found that the ring-light configurations produce a lower light intensity but also a slightly larger illuminated area and a more diffuse and uniformly distributed illumination than the Eckardt TwinLight.
The placement of the ring-light in the posterior chamber proved to be unsafe and complex, since the installation procedure in ex-vivo porcine cadaver eyes could not be executed without damaging the lens bag, lens zonules, posterior iris and cornea, due to the lack of space in the posterior chamber, and the rigidity and sharp tip of the ring.
The ring-light configuration that is inserted through the pars plana also caused damage to the retina and lens bag, due to the ring’s rigidity and its sharp tip. However, this configuration shows potential when the ring would be made from a more flexible material and has an incorporated pigtail curve a the distal end, which can be easily achieved in a future version of the prototype.

Introduction: Although being the most advanced spinal stabilization method, pedicle screw fixation has an increased number of complications compared to other surgical procedures, due to the complex spinal anatomy, large inter-patient anatomical variability, and the highly vulnerable surrounding tissue of the spine. Screw fixation strength can be increased by anchoring a flexible screw in a hole following the curved trajectory of the more dense cortical shell of the vertebral body, instead of the porous core. The curved trajectory can be determined by local diffuse reflectance spectroscopy (DRS) guiding. Currently, there is no commercial phantom or phantom recipe known that provides the correct optical and structural material characteristics for testing DRS guiding and curved trajectory drilling on, comparing it to similar devices, and perform surgical training sessions on. The research goal is to design and validate an integrated phantom that closes this gap. Preliminary test results: Preliminary optical and structural tests were performed separately on hybrid optical filler samples and structural matrix samples, respectively. The hybrid optical filler samples made out of water, fat (30%, 40% and 50% fat fraction) and silicone (95%, 80%, and 50%) were inspected visually, and spectra were obtained by means of a DRS device in week 0 as well as in week 10. The most promising optical filler sample consisted of 50% of silicone and 50% of water+fat, with a fat fraction of 40%. The measured DRS signals showed a clear correlation between the fitted and actual water+fat content, which, can thus be used as a tissue differentiation parameter. The stability and preservability of phantoms that were enhanced with silicone and resulted in homogeneous mixtures, was increased compared to samples without any silicone. For the structural matrix, different 3D printers, materials, and infill densities were evaluated on drill characteristics and compression characteristics. The best bone mimicking structural phantom consisted of an FDM printed part made out of ABS with a gyroid infill structure with an infill density of 50%. Integrated phantom validation: The combination of the optical filler and the structural matrix was validated by infusing the optical filler in the structural matrix with a vacuum chamber and measuring the drill characteristics and DRS spectra at various depths inside the phantom. A gradual increase in axial drilling force and a gradual change in spectral shape was noticed as a function of depth from the cancellous zone, to the pre-cortical zone and cortical zone. Discussion: The ABS material and the actual optical filler constituents (such as peanut oil and specific silicone types) should be implemented inviii the fitting algorithm. Nevertheless, the results of this research show not only great potential phantom design for a wide variety of bone types because of the structure and shape design flexibility of 3D printing, but also for soft tissues by using only the optical filler with a much better stability and preservability than traditional soft tissue phantoms.

Glioblastoma multiforme tumors (GBM) are the most frequent and aggressive brain tumor type and have a very low survival rate. In order to improve the patient’s outcome for a patient suffering from GBM the production of brain phantoms can be of great help. A brain phantom can be used for multiple purposes, such as neurological planning, training and for testing new medical devices. Nevertheless, no brain phantom has been created with realistic stiffness yet. Investigating the stiffness of the brain is challenging, due to the complexity of the brain and for not being able to touch it in person. However, through a relatively new technology called magnetic resonance elastography (MRE), it is possible to examine the stiffness of the brain. Hence, the
objective for this thesis was to develop a brain phantom with a GBM tumor with similar stiffness compared to the biological human brain and tumor tissue.

In the literatureMRE data can be found on the complex shear modulus of the brain and tumors, which can be converted to an indicative Young’s modulus (E*). The indicative Young’s modulus will be used as a guideline for the desired stiffness needed for the production of a brain phantom. First various polyvinyl alcohol (PVA) samples were made and a compression test was carried out to derive the Young’s modulus (E). By comparing E with E* , it can be seen which PVA sample most resembles the brain and tumor tissue. Thereafter, the head phantom is created, which consists of 3 different production processes; production of the mould, production of the skull and production of the brain phantom. Firstly, the moulds for the brain phantom and tumor are 3D printed using polylactic acid (PLA). The skull phantom is derived from a patient specific CT scan, and is 3D printed with PLA. The brain phantom itself is made of 4,8 wt% PVA and 2 freeze-thaw (FT) cycles, and the tumor phantom is made of 5,2 wt% PVA and 1 FT cycle. The mass fractions and FT cycles have the closest stiffness compared to real brain and tumor tissue and were therefore selected. Within the brain phantom a hollow space is created where the tumor can be placed in. For a parallel project of the TU Delft and Utrecht University a brain phantom was required. For this specific case a brain phantom was produced with barium sulfate (BaSO4), which functions as a contrast additive for CT and MRI. Therefore, the second brain phantom
was made of 4,8 wt% PVA, 1 wt% BaSO4 and 2 FT cycles. The tumor phantom with BaSO4 failed and therefore the first tumor without BaSO4 was used for this use case. An experiment was carried out where a HoMS substance was injected into the tumor phantom. During the experiment 9 CT scans and 2 MRI scans were carried out. Due to the MRI and CT scans it could be investigated whether the tumor was in place and
whether the brain phantom was visible in the CT and MRI. After the use case one more mechanical test was carried out consisting of PVA samples with BaSO4 to examine whether BaSO4 has an effect on the stiffness of the phantom.

The results of the mechanical test led to a desired mass fraction and FT cycle that can represent the brain and tumor phantom. These were immediately used for the production of the brain and tumor phantom. The Young’s modulus of the brain phantom (E = 8,9 kPa) is a little stiffer compared to the actual brain tissue (E* =
9,57 kPa). The Young’s modulus of the tumor phantom (E = 5,8 kPa) is softer compared to the GBM tissue (E* = 4,93 kPa). However, adding BaSO4 to the PVA solution changed the stiffness of the PVA samples, has had an effect on the compactness of the samples, and the BaSO4 did not dissolve properly. In the results of the CT and MRI scans it was found that the BaSO4 was only visible in the inferior part of the brain. In addition, the brain and tumor phantom were clearly visible in theMRI and CT scans. It is therefore not recommended to use BaSO4 for future research. Furthermore, the CT scan showed that the tumor phantom was well surrounded
by the brain phantom and stayed in place even when the tumor was inserted by a cannula. The results show that a brain phantom with a tumor phantom can be produced. However, some improvements to the design of the mould are needed, because the mould was leaking, which subsequently led to an anatomically incorrect brain phantom.

This thesis project leads to the following conclusion; it is possible to produce a brain phantom with a GBM tumor phantom with comparable stiffness to the actual brain and GBM tissue. However, it can not be concluded that the various stiffness’s are exactly the same as the actual brain and GBM tissue, but it is close to the values found in the literature.

One of the main elements for creating an optimal operating room schedule is an accurate surgery duration estimation. Currently, this estimation is only done preoperatively. However, multiple factors during the surgery itself could influence the duration, for example bleeding. Literature showed the possibility of intraoperatively estimating a surgery duration based on surgical progress. However, one of the main concerns was the usability of such a system in the operating room workflow. Therefore, this research was focused on two parts: (1) the creation of an automatic intraoperative remaining surgery duration estimation system and (2) the evaluation of such a system for the operating room workflow. Two types of surgeries were used for the estimation, the Laparoscopic Cholecystectomy and the Total Laparoscopic Hysterectomy. The estimation was created using multiple statistical regressor methods, such as linear regression and random forest, and progress-based methods based on the nearest-neighbors algorithm and Dynamic Time Warping method. The evaluation was done on two levels: the system level based on the error of the estimation, and the operating room workflow based on surgical data from 2016 to 2019 and interviews with the operating room program coordinators. Results showed that an intraoperative remaining surgery duration estimation system based on surgical phases was able to re-estimate the duration with an error of about 10 minutes, an acceptable error for the operating room workflow. Moreover, the third quarter of the surgery showed to be the essential part where an accurate estimation is needed. Furthermore, an automatic system showed additional benefits such as being unbiased, continuous, and reducing unnecessary disturbance in the operating rooms. Overall, this research showed that an intraoperative remaining surgery duration system based on surgical phases is promising for the operating room workflow. Future research is needed to understand how to implement such a system in the operating room workflow.

In hospitals, the duration of surgical procedures is susceptible to a number a factors resulting in varying procedure durations of the same procedure. When these procedure times deviate from the predicted surgical duration, scheduling conflicts arise, leading to patient discomfort and inefficient use of hospital resources. Considering the aging population and the subsequent growing demand for healthcare, hospitals are under constant pressure to provide cost-effective care. Indeed, this is also the case for the interventional cardiology unit, where the expected increase of the prevalence of cardiovascular disease is expected to put a strain on hospital human resources. Improving the efficiency has been an area of focus in many hospitals. In fact, advanced methods such as surgical workflow monitoring have been employed to identify points of improvement especially in procedure scheduling. Up until now, this has not been employed in cardiac catheterization laboratories in the Netherland.
The purpose of this study, therefore is identify primarily the factors influencing variability of the procedure time of procedures of the cardiac catheterization laboratory of the Reinier de Graaf Hospital in Delft and secondarily the possible sources compatible with video-based workflow monitoring for phase segmentation of the procedures.
Retrospective data obtained from DoseWise was used to analyze the procedure times and associating factors of coronary angiograms and pacemaker implantation procedures. In addition, process models of the procedures were created and activity recognition identifiers were studied through observer-based data acquisition.
Analysis showed that in coronary angiograms variations of total procedure times are mostly due to variable preparation times, whereas for pacemaker implantation the lead positioning times were the main source for variability. In addition, more experienced operators were shown to complete the procedure in a shorter time (-4.35%|p=0.007) in comparison to less experienced operator.
Instrument and equipment use were identified as possible sources for activity recognition along with the hand kinematics, however further analysis of reliability is required.
The findings in this study provide a basis for future monitoring of the workflow of these procedures using real-time data acquisition methods to improve the efficiency in cardiac catheterization laboratories.

Atrial Fibrillation (AF) is a common type of arrhythmia, characterized by rapid and irregular contraction of the atria. Irregular contraction of the atria is caused by erroneous electrical activation, originating from the tissue around the Pulmonary Veins (PV) in the Left Atrium (LA). Catheter-based treatment of AF includes ablation of the tissue around the PV.

Imaging of ablation catheters and LA-anatomy is of great importance for accurate and successful treatment of AF. In current state of the art LA-ablation procedures, static, pre-procedural acquired 3D models of the LA are used for catheter navigation and monitoring. However, due to the limited accuracy of these static 3D models there is a continuous demand for real-time 3D imaging techniques during catheter based treatment of AF.

A possible solution for real time imaging is to use trans-balloon, miniaturized transesophageal echocardiography (TEE) as imaging modality for AF ablation procedures. Trans-balloon, miniaturized TEE can provide real-time 3D ultrasound imaging of the LA.

Since trans-balloon, miniaturized TEE is not used in clinical practice yet, it is unknown whether image quality of the miniaturized TEE probe will be sufficient for imaging during AF ablation procedures. Therefore, the image quality of trans-balloon, miniaturized TEE has to be compared to image quality current state of the art TEE probes.

In this thesis, image quality of the miniaturized ultrasound probe was compared to current state of the art TEE probes, and the influence of the balloon on image quality was investigated. Image quality was assessed using standardized ultrasound image quality assessment phantoms and software. Image quality was measured using the following image quality parameters: contrast to noise ratio (CNR), spatial resolution and penetration depth of the ultrasound signal.

The results have shown that CNR of the miniaturized probe is equal to that of state of the art probes, but that spatial resolution of the miniaturized probe strongly depends on the rotation angle of the ultrasound imaging plane. Rotating the imaging plane of the miniaturized TEE probe negatively impacts spatial resolution, which was not observed for the state of the art TEE probes. The penetration depth of the ultrasound signal was significantly less compared to state of the art TEE probes.

The rotation-dependent spatial resolution of the miniaturized probe is undesired, since rotating the imaging plane is common practice in TEE. A review study involving clinical experts is recommended to judge whether image quality of the miniaturized probe with rotated imaging plane is sufficient.
Since the anatomical structures of interest during AF are close to the transducer, the limited penetration depth of the miniaturized probe will not be a major limitation. Despite aforementioned limitations, the results suggest that trans-balloon miniaturized TEE can be used for real-time ultrasound imaging during AF ablation procedures.

The research in this thesis provides a theoretical framework to measure and compare image quality of ultrasound probes, which is an important first step in the development of a novel, real-time 3D imaging technique for imaging of AF ablation procedures.

Brain, outside the body, will decay quickly, causing ex vivo studies to be difficult. Having a phantom model has a great beneficial value for numerous reasons. Phantoms are, among others, used in research centres, to validate new equipment, to develop biomechanical models or to test new treatment methods. They are representations of organs or tissues made from tissues that mimicking the desired properties of the organ. This thesis research is performed in collaboration with Philips Healthcare research department. They benefit from having a brain phantom for the development of a new endoscopic tool that incorporates virtual reality images in the view. Existing phantoms are usually expensive and are not meant for destruction by needle interventions or endoscopic interventions, or are a very rough representation of reality. For this study, it was the aim to develop an anthropomorphic brain phantom containing the hollow space of the ventricles of the brain, with the correct mechanical and optical characteristics. From a digital 3D mesh brain, moulds were made to assess different ways to produce a brain phantom with ventricles. From literature, the tissue mimicking material (TMM) PVA was selected to use for the production of the phantom. After evaluation of the different approaches to fabricate a phantom, it was decided to produce the hollow spaces of the ventricles by using 3D printed soluble PVA, to be removed out of the model after casting. This was done by 3D printing a brain mould in which a 3D printed PVA ventricle could be inserted. A solution of 6% PVA as tissue mimicking material was used. After the mould and production principle of the model was finalized, the imaging part was assessed. In order for the phantom to be used in CT imaging, barium sulphate was added to the PVA solution as a contrast enhancing material. Two phantoms were made with 1% and 2% barium sulphate. Finally, to assess the shape and the quality of the ventricles in the phantom, the phantom was scanned with a CT scan at Philips Healthcare in Best, the Netherlands. The CT files were evaluated using RadiAnt and 3D slicer to assess the ventricular shape and position. 3D slicer was also used to segment the ventricle shape out of the designed models to compare with the originally developed ventricle structure. This study functions as a proof of concept for the development of a PVA brain phantom containing ventricles produced with 3D printed soluble PVA. All in all, the overall evaluation of the prototypes have shown to be promising models for the development of a brain phantom using PVA and ‘homemade’ fabrication techniques. The use of 3D printed soluble PVA is a novelty in this field of application. It makes the design easy to develop and easily adjusted to patient-specific cases as personalized models can be made.