Abstract:
Atrial fibrillation is responsible for a significant and steadily rising burden. Simultaneously, the treatment options for atrial fibrillation are far from optimal. Personalized simulations of cardiac electrophysiology could assist clinicians in the risk stratification and therapy planning for atrial fibrillation. However, the use of personalized simulations in clinics is currently not possible due to either too high computational costs or non-sufficient accuracy. Eikonal simulations come with low computational costs but cannot replicate the influence of cardiac tissue geometry on the conduction velocity of the wave propagation. Consequently, they currently lack the required accuracy to be applied in clinics. Biophysically detailed simulations on the other hand are accurate but associated with too high computational costs. To tackle this issue, a regression model is created based on biophysically detailed bidomain simulation data. This regression formula calculates the conduction velocity dependent on the thickness and curvature of the heart wall. Afterwards the formula was implemented into the eikonal model with the goal to increase the accuracy of the eikonal model without losing its advantage of computational efficiency. The results of the modified eikonal simulations demonstrate that (i) the local activation times become significantly closer to those of the biophysically detailed bidomain simulations, (ii) the advantage of the eikonal model of a low sensitivity to the resolution of the mesh was reduced further, and (iii) the unrealistic occurrence of endo-epicardial dissociation in simulations was remedied. The results suggest that the accuracy of the eikonal model was significantly increased. At the same time, the additional computational costs caused by the implementation of the regression formula are neglectable. In conclusion, a successful step towards a more accurate and fast computational model of cardiac electrophysiology was achieved.
Abstract:
The generation of Synthetic Computed Tomography (sCT) images has become a pivotal methodology in modern clinical practice, particularly in the context of Radiotherapy (RT) treatment planning. The use of sCT enables the calculation of doses, pushing towards Magnetic Resonance Imaging (MRI) guided radiotherapy treatments. Moreover, with the introduction of MRI-Positron Emission Tomography (PET) hybrid scanners, the derivation of sCT from MRI can improve the attenuation correction of PET images.Deep learning methods for MRI-to-sCT have shown promising results, but their reliance on single-centre training dataset limits generalisation capabilities to diverse clinical settings. Moreover, creating centralised multi-centre datasets may pose privacy concerns. To address the aforementioned issues, we introduced FedSynthCT-Brain, an approach based on the Federated Learning (FL) paradigm for MRI-to-sCT in brain imaging. This is among the first applications of FL for MRI-to-sCT, employing a cross-silo horizontal FL approach that allows multiple centres to collaboratively train a U-Net-based deep learning model. We validated our method using real multicentre data from four European and American centres, simulating heterogeneous scanner types and acquisition modalities, and tested its performance on an independent dataset from a centre outside the federation.In the case of the unseen centre, the federated model achieved a median Mean Absolute Error (MAE) of 102.0 HU across 23 patients, with an interquartile range of 96.7–110.5 HU. The median (interquartile range) for the Structural Similarity Index (SSIM) and the Peak Signal to Noise Ratio (PNSR) were 0.89 (0.86–0.89) and 26.58 (25.52–27.42), respectively.The analysis of the results showed acceptable performances of the federated approach, thus highlighting the potential of FL to enhance MRI-to-sCT to improve generalisability and advancing safe and equitable clinical applications while fostering collaboration and preserving data privacy.
Abstract:
This thesis investigates federated learning (FL) techniques for generating synthetic computed tomography (sCT) images from magnetic resonance imaging (MRI) data, aiming to support MRI-only workflows in radiotherapy planning. Traditionally, CT scans are necessary for calculating radiation doses, but these scans expose patients to additional ionizing radiation.An MRI-only approach offers superior soft tissue contrast, reduced radiation exposure, and streamlined clinical workflows. To achieve this, sCT images must be generated from MRI to provide CT-equivalent information safely and robustly.Centralized deep learning (DL) has shown promise for sCT generation, but clinical adoption is limited due to the scarcity of large, diverse datasets, as data sharing across institutionsraises privacy concerns. FL mitigates this by enabling collaborative model training across institutions without centralizing data, thus utilizing diverse datasets while preserving patient privacy.The primary research question is: How can federated model aggregation be optimized for performance and computational efficiency in MRI-to-sCT translation? The first hypothesis suggests that due to the complexity of medical image translation and inter-institutional data heterogeneity, more advanced aggregation strategies may be necessary for robust generalization.This study leverages MRI and CT datasets from multiple institutions with variations in imaging protocols, scanners, and patient demographics to simulate realistic clinical diversity.A robust pre-processing pipeline standardizes data, aligning image dimensions, intensity ranges, and anatomical landmarks to reduce inter-institutional variability and support model convergence. Multiple FL aggregation strategies—FedAvg, FedMedian, FedTrimmedAvg,FedAvgM, optimization-based methods, FedBN, and FedProx—were benchmarked for their ability to manage non-uniform data distributions and improve model generalization.Importantly, data remains on each client, and the global model’s generalization is tested onunseen data from an external institution. Model quality was evaluated using key performance metrics: masked mean absolute error, peak signal-to-noise ratio, and structural similarity index.The findings show that (i) FedAvg exceeded performance expectations by outperforming more complex tsrategies, (ii) FedMedian’s simplistic outlier filtering led to information loss, (iii) FedTrimmedAvg ranked between FedAvg and FedMedian, (iv) FedAvgM provided enhanced stability but has slower convergence, and (v) optimization-based strategies were instable and outperformed by simpler methods. The combination of FedAvg with FedProx and FedBN produced the best results, achieving a median masked mean absolute error of 96 HU on 23 unseen test patients.Contrary to the initial hypothesis, simpler aggregation strategies outperformed more complex methods for MRI-to-sCT translation. This may be attributed to the extensive pre-processing pipeline, which effectively reduced data heterogeneity, allowing FedAvg to perform well.These findings underscore FL’s potential for enabling MRI-only radiotherapy by facilitating sCT generation across decentralized datasets, preserving privacy while maintaining model performance. By demonstrating effective data harmonization and adaptable FL strategies in a multi-institutional setting, this work contributes to developing secure, generalizable DL applications in medical imaging, paving the way for broader clinical implementation.While this study demonstrates the feasibility of FL for privacy-preserving sCT generation, the main goal was to provide a first comprehensive benchmark analysis of various aggregation strategies for this task. Future work should explore additional pre-processing techniques and further refine FL approaches, such as combining FedAvg with emerging strategies like FedDG and FedCE. Moreover, practical challenges, including privacy-preserving aggregation, communication costs, and device variability in real-world federated learning settings, must be addressed to optimize federated learning’s effectiveness and scalability in clinical applications.
Abstract:
Atrial fibrillation (AFib) is responsible for complications and excess death and its prevalence is expected to further increase in the future, due to the increasing life expectancy of humans and age being a major risk factor [1] [2]. At the same time the treatment options available for AFib are far from optimal. Radio-frequency ablation has the potential to be highly effective with comparably small side effects. The major limitation for the success of radio-frequency ablation is the difficulty to correctly locate the areas responsible for abnormal conduction. Personalized computational models to test strategies for the ablation procedure could be of great benefit for the success rate of this treatment. [3] Bidomain models have the required accuracy to theoretically assist the clinicians, but their high computational cost makes them inapplicable in clinical environments [4]. Eikonal models on the other hand could potentially be used in clinics due to their low computational costs, but lack the required accuracy, because of their inability to capture certain electrophys- iological effects. The goal of this project is to refine the eikonal model by implementing the influence of geometrical factors on conduction velocity (CV) into the eikonal model. This was achieved by using previously published data about the influence of geometrical factors on CV obtained with bidomain simulations to create a regression model. These geo- metrical factors were muscle curvature, muscle thickness and bath loading. The regression model was then implemented into the eikonal model’s local speed function (LSF) aiming to increase the accuracy of the eikonal model. The implementation of the regression formula into the eikonal model was followed by the evaluation of the improvements made by the regression model. This was achieved by creat- ing two dimensional simulation setups to compare the eikonal model before and after the implementation of the regression formula to bidomain simulations. By applying the same stimulus on the same mesh, the eikonal and bidomain simulations became comparable. The muscle tissue model used to create the mesh and run the eikonal and bidomain simulations was a two dimensional rectangle with a variable muscle thickness and the possibility to bend it to different uniform curvatures. The use of this mesh allowed to study the influence of muscle thickness and curvature on the simulations. ....