Abstract:

Brain tumour segmentation is a key application of AI in neuroimaging. Recently, federated learning (FL) has emerged as a strategic and increasingly relevant paradigm in neural computing due to its ability to address key challenges in large-scale neural network training, such as data access, privacy, collaborative learning, and model robustness. However, its adoption is currently hindered by high communication costs and the heterogeneity of client data. In this study, we investigated an efficient FL framework for brain tumour segmentation based on communication-aware optimization. We evaluated FedWSOComp, which integrates sparsification, quantization, and entropy-based encoding, in combination with a 3D U-Net architecture under both homogeneous and heterogeneous data distributions. The multi-institutional FeTS 2024 dataset was employed and partitioned into independent and identically distributed (IID) and non-IID settings, with an independent test set of 67 patients. An overall of 18 configurations combined sparsification rates and quantization levels. Performance was measured using Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff Distance (HD95). Experimental results demonstrated that aggressive compression caused severe degradation in segmentation quality, with HD95 exceeding 60 mm. In contrast, higher retention with finer quantization achieved the best balance between efficiency and accuracy, reaching a DSC and HD95 mm on the test set under non-IID conditions. The findings demonstrated that, when configured with moderate-to-fine quantization and high sparsification retention, FedWSOComp enabled accurate and communication-efficient federated brain tumour segmentation. This study provides quantitative evidence and practical guidance for the deployment of FL-based segmentation models in privacy-sensitive and bandwidth-constrained clinical settings.

Abstract:

Coronary artery disease remains a leading cause of mortality worldwide. Coronary computed tomography angiography (CCTA) provides a non-invasive basis for diagnosis; however, accurate and connectivity-preserving segmentation of the coronary artery tree is essential for robust automated and quantitative analysis. CNN-based architectures, in particular nnU-Net, achieve strong volumetric accuracy but often produce fragmented vessel trees when segmenting thin tubular structures such as coronary arteries. Skeleton Recall (SR) loss has recently been proposed to address these limitations by explicitly penalizing missing centerline segments, yet its effectiveness for coronary artery tree segmentation remains to be established. This work compares a generic baseline loss with an SR-augmented variant within a standard nnU-Net 3D segmentation pipeline for coronary artery segmentation from CCTA. Evaluation is performed using five-fold cross-validation on 98 CCTA volumes from the public ASOCA dataset and an additional in-house cohort. Performance is evaluated using complementary volumetric and connectivity metrics, with quantitative findings contextualized through targeted qualitative visualizations. Results show that augmenting the generic loss with SR significantly improves coronary artery tree connectivity in a practically relevant manner, while maintaining comparable volumetric overlap. Overall, SR constitutes an effective and efficient loss that is simple to integrate, making it a practically viable choice for connectivity-preserving coronary artery tree segmentation.

Abstract:

Stereoelectroencephalography is a minimally invasive neurosurgical technique used to lo-calize the epileptogenic zone (region responsible for seizure generation) in patients withdrug-resistant epilepsy. The procedure requires the placement of multiple intracranial elec-trodes along carefully planned trajectories that reach predefined brain targets while avoidingcritical structures such as blood vessels. Planning these trajectories is complex and timeconsuming, relying heavily on manual inspection of medical images and surgical expertise.Consequently, there is increasing interest in computer-assisted approaches that can supportclinicians during the planning process.This thesis investigates the feasibility of using deep reinforcement learning to automatethe trajectory generation step of the SEEG planning workflow. The proposed frameworkintegrates medical imaging, vascular modeling, and a reinforcement learning agent into amodular pipeline. The trajectory planning task is then formulated as a sequential decision-making problem in which an agent iteratively adjusts electrode entry and target points. Inorder to approximate the action-value function and guide the search through the anatomicalspace, a Deep Q-Network is used, together with a reward formulation that incorporates safetyand geometric constraints, particularly vessel avoidance.Experimental evaluation using patient imaging data demonstrates that the RL agent cangenerate candidate electrode trajectories that reach predefined targets while maintaining safedistances from vascular structures. In this work, 2000 candidate trajectories were generatedand evaluated, with feasibility rates ranging from about 26% to 52% depending on anatomicalcomplexity. Generation of trajectories requires only a few seconds per electrode, indicatingthe potential for substantial reductions in planning time.Overall, the results demonstrate that reinforcement learning can effectively automate thegeneration of candidate SEEG trajectories and provide a foundation for future computer-assisted surgical planning systems.

Abstract:

To ensure adequate cerebral oxygenation, bypass surgery may be required to reroute blood around an occlusion or stenosis. Therefore, intraoperative evaluation of blood flow is crucial. Current methods are ultrasound-based and require direct vessel contact.Therefore, non-contact optical methods based on Optical Coherence Tomography (OCT) are investigated. OCT provides depth-resolved structural imaging. By applying the so-called Speckle Variance (SV) algorithm, which calculates the variance of the OCT signal over time, moving scatterers such as red blood cells within vessels can be distinguished from the surrounding static tissue, thereby enabling vessel visualization, which is referred to as OCT angiography (OCTA). While the SV algorithm enables visualization of vessels, the broader objective is the quantification of red blood cell flow velocity within the vessels.The SV signal depends on flow velocity and exhibits a sigmoidal shape at low flow velocities and decreases again at high velocities after reaching saturation. The influence of several factors on this curve was investigated. First, the A-scan rate was varied, where exposure time and interscan time are identical. In a second step, the effects of exposure time and interscan time were examined separately in order to assess their individual impact.To evaluate the influence of these parameters, two velocities were defined for objective assessment: the slowest detectable velocity (SDV) and the fastest distinguishable velocity (FDV). Experiments were performed using a milk–water mixture as a blood analogue, pumped through a microchannel with a syringe pump, and scans were acquired at seven A-scan rates (5, 10, 25, 50, 100, 200, and 248 kHz).Higher A-scan rates increased both SDV and FDV, and also expanded the measurable velocity range, while low A-scan rates improved sensitivity for slow flows but reduced the overall range. Post-processing showed that higher interscan times increase sensitivity for slow flows but lowers the FDV, whereas increasing exposure time decreases the SDV without a clear effect on FDV.Furthermore, it was investigated that the drop in SV signal after saturation is mainly caused by the washout effect, which reduces the intensity of the OCT signal and consequently the variance. This was demonstrated by normalizing the velocity response curve by the square of the intensity. On the normalized curve, no drop was observed.In conclusion, this thesis provides a quantified basis for selecting OCTA measurement parameters for clinical applications such as intraoperative bypass surgery.

Abstract:

Lung collapse is an important physiological process that occurs when the pleural cavity is opened during thoracic surgery[1, 2], and it is also intentionally induced through artificial pneumothorax to improve surgical exposure and operating space[3, 4]. Reconstructing this phenomenon in a controlled and observable manner is therefore essential for surgical simulation and collapsed lung modelling[5], enabling safer training conditions and more reliable evaluation of operative strategies. However, most existing lung phantom models focus primarily on anatomical form or material elasticity and do not reproduce the dynamic pressure behavior that governs real collapse[6, 7], which limits their usefulness for physiology oriented studies. This study presents an experimental platform developed to simulate pressure variations associated with lung inflation and collapse. The system integrates medical image processing, mechanical and electronic design, embedded system development, product level prototyping, and experimental data analysis into a unified workflow. Lung geometry was first segmented from CT data using 3D Slicer[8] and was then further simplified in Autodesk Fusion 360, AutoCAD, and Blender to achieve a physiologically appropriate scale while ccommodating manufacturability and surgical access constraints such as incision and opening placement. The physical model was fabricated through 3D printing with iterative efinement in material selection, SLA and FDM evaluation[9, 10], and optimization of printing accuracy and reliability. A pneumatic system driven by air pumps and relay control was mplemented to generate controlled pressure environments representing different collapse conditions. A pressure sensor was integrated to monitor internal pressure changes over time, enabling the capture of characteristic pressure trends. Electronic schematics were designed in EasyEDA, while control logic was programmed using Arduino IDE in C++. Python cripts developed in Visual Studio Code provided automated serial data acquisition, CSV storage, and line-plot visualization[11], forming the basis for interpreting and comparing xperimental results. To evaluate physiological consistency, three categories of collapse simulation experiments were conducted: controllable collapse produced by airway suction[2], uncontrollable col-lapse driven by pleural-side inflow[4], and a more surgery oriented pneumothorax pattern achieved through alternating pleural side inflow. Although the platform rovides trend-level pressure information rather than absolute physiological values, the resulting pressure–time curves clearly reflect the transition from a supported state toward collapse as the pressure environment moves toward equilibrium[12]. Based on these findings, the model is capable of reproducing a range of collapse patterns, including normal inflation, full collapse, and behavior similar to artificial pneumothorax[3], demonstrating its suitability for studying both controlled and uncontrolled collapse mechanisms.Overall, the platform offers a practical and low cost tool for investigating collapse dynamics[5], supports thoracic surgical education and hands on training[2], and provides a solid foundation for the future development of higher fidelity and physiology based thoracic simulation systems.

Abstract:

Diagnostic bronchoscopy is an uncomfortable procedure that requires careful navigation through narrow, highly branching airways. High-fidelity simulation can improve clinician training and enable the development of vision-based navigation methods, but realistic bronchoscopic video data is dicult to obtain in practice: clinical recordings are often inaccessible due to privacy constraints and strict hospital governance. Existing common unpaired baselines typically require substantial target data and often degrade over long sequences due to temporal drift and flicker, making them poorly suited to the practical realities of bronchoscopy constraints where target data is scarce and stability over long, clinically realistic durations is essential.To address these limitations, we introduce a pipeline that learns from fewer than 500 target frames and performs inference using a single target reference frame. The proposed method produces stable long-horizon translations from CT-driven virtual bronchoscopic videos, preserves airway geometry—especially small bifurcations critical for navigation—and reduces temporal artifacts over extended sequences through world consistency, while still enabling controllable transfer of discrete appearance modes without requiring separate training for each style.This thesis studies unpaired video-to-video translation for bronchoscopic simulation in an extreme low-data setting, with an explicit focus on clinically relevant constraints. Rather than aiming for stylization alone, the work prioritizes: (1) airway structure and geometry preservation, keeping lumen boundaries and fine bifurcations stable to avoid misleading navigation cues; (2) temporal consistency over long sequences, mitigating flicker and drift; and (3) appearance transfer as a secondary objective, aligning texture, color, and wet-surface highlights to the chosen reference.

Abstract:

Atrial fibrillation (AFib) is the most common clinically relevant cardiac arrhythmia. A promising treatment modality is pulsed field ablation (PFA), a non-thermal catheter ablation technique that offers increased tissue selectivity and a reduced risk of collateral damage compared to thermal ablation methods. PFA induces irreversible electroporation, whereby electric pulses create pores in the cell membrane, ultimately leading to cell death. While electroporation depends on pulse parameters, field strength, and cell orientation, previous studies have focused on single cells. This work addresses the question: How do orientation and pulse parameters affect electroporation in coupled cardiomyocytes? This thesis investigates the influence of anisotropy on electroporation through development and extension of a unicellular model to two-cell and five-cell models of electrically coupled cardiomyocytes. Electrical coupling between the cells was modelled by incorporating gap junction conductivity and its reduction due to electroporation. Finite element simulations were performed for six different pulse durations (100 ns, 1 μs, 10 μs, 100 μs, 1 ms, 10 ms) and electric field strengths in the range 10 to 10^5 V/cm for both parallel and perpendicular cell orientations relative to the electric field. The contribution of gap junctions to the induced TMV and pore density was assessed by comparing simulations with functional and inhibited gap junctions. The two-cardiomyocyte model exhibited electroporation onset and extent consistent with unicellular studies, with pore formation occurring at lower field strengths in the parallel orientation for pulse durations of 100 μs and longer. In contrast, the five-cardiomyocyte model demonstrated electroporation only for 1 ms and 10 ms pulses at field strengths exceeding 10^4 V/cm, with shorter pulses failing to induce electroporation. In this five-cell configuration, electroporation onset was unaffected by cell orientation, though orientation-dependent differences in electroporation extent persisted. Gap junction mediated intercellular coupling influenced membrane voltage approximately twice as strongly when the intercalated disc was aligned parallel to the electric field. Key limitations of the model include the use of a fixed, real-valued membrane capacitance, and the simplified representation of the intercalated disc as a single-capacitor element. These simplifications led to altered TMV dynamics for nanosecond pulses and exaggerated immediate effects of gap junction decoupling. Overall, this work provides insight into anisotropic electroporation in coupled cardiomyocytes and demonstrates the importance of incorporating multicellular effects when designing pulse protocols. The findings suggest that short pulses reduce orientation dependence and may improve lesion uniformity, while gap junction coupling modulates electroporation thresholds and extent. By linking unicellular phenomena to multicellular behaviour, this thesis offers a foundation for optimising PFA protocols to achieve more predictable and effective ablation outcomes.

Abstract:

The extracellular–membrane–intracellular (EMI) model enables cardiac electrophysiologysimulations with explicit resolution of individual cardiomyocytes and their microstructuralinteractions. By directly representing intracellular, extracellular, and membrane domains, theEMI framework allows the investigation of cell-level conduction mechanisms that cannot becaptured by homogenized mono- or bidomain models. However, this increased physiologicalfidelity comes at the cost of substantially higher computational demands due to the largenumber of degrees of freedom required for spatial discretization.The main objective of this bachelor’s thesis is to systematically investigate the influenceof spatial mesh resolution on simulation accuracy and computational performance in EMIbased cardiac electrophysiology simulations using the openCARP framework. For thispurpose, a convergence analysis was performed. Starting from a high-resolution referencemesh, a large set of coarsened meshes was generated by varying edge length constraintsand geometric deviation parameters. To assess numerical convergence, three clinically andphysiologically relevant metrics were analyzed: unipolar electrograms, conduction velocities(CVs), and upstroke velocities of the action potentials. The finest mesh configuration servedas a reference solution against which all other configurations were compared.The results demonstrated that substantial reductions in computational cost can be achievedthrough spatial coarsening while maintaining acceptable accuracy, defined by individuallimits depending on the chosen metric. For the optimal mesh configuration with a meaneffective edge length of hedge = 10.3µm (σedge = 2.98µm), computational runtime wasreduced by approximately 80% compared to the highest-resolution reference mesh whilepreserving electrophysiological accuracy for all considered metrics. Based on the meshconfiguration identified as the optimal trade-off for CV evaluation, a parameter tuningstudy was conducted to obtain physiologically realistic longitudinal CV values within theEMI framework in the openCARP environment. Specifically, physiologically realisticlongitudinal conduction velocities of approximately 1m/s were obtained by selecting eithera gap junction resistance of Rg = 0.00613kΩcm2 at an intracellular conductivity of σi =0.4S/m, or an intracellular conductivity of σi = 0.1575S/m at a gap junction resistance ofRg =0.003kΩcm2.Overall, this work provides practical guidelines for selecting mesh resolution and parametervalues in EMI-based simulations and contributes to the validation and efficient application ofmicrostructural cardiac electrophysiology models.