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.