Verbeterde analyse van hersen-PET-beelden met machine learning

This thesis investigates how machine learning can be integrated with established PET kinetic modelling, non-invasive quantification, supervised classification, and network analysis to advance brain PET image analysis. Across these four topics, the work progresses from local voxel-wise estimates to global network-level characterizations, while maintaining a consistent emphasis on tracer kinetics, quantification, classification, and practical clinical applicability. Chapter 2 introduces a PET kinetics-informed neural network (PKiNN) to generate voxel-wise one-tissue compartment model (1TCM) parametric mapsfrom dynamic PET data using an arterial input function. Using a compact LSTM architecture, PKiNN enables fast inference and reduced computational load compared with conventional voxel-wise fits. Embedding 1TCM constraints into the loss function improves parameter estimation, yielding close agreement with conventional 1TCM fits in simulated data and strong concordance with volume of interest (VOI)-based kinetic parameter estimates in clinical 11C-UCB-J data, while mitigating the underestimation observed for Logan graphical analysis. Chapter 3 explores SVM-derived weight maps as a data-driven strategy for pseudo-reference region selection to support non-invasive quantification using SUVR or reference-tissue modelling and to facilitate voxel-wise group comparisons without arterial sampling. Applied to 11C-PiB (AD vs controls), 11C-UCB-J (baseline vs blocking), and 18F-DPA-714 (chemo-treated breast cancer patients vs controls), this approach produces tracer-specific pseudo-reference region candidates, with cerebellum and brainstem emerging as consistent options across cohorts. The stability of candidate region rankings supports feasibility in small cohorts while highlighting tracer- and dataset-specific requirements. Chapter 4 assesses the potential and limitations of voxel-wise 18F-FDG PET for ALS phenotyping using SVM classification. High performance is achieved for ALS versus healthy volunteers (>80%), with weight maps reflecting established ALS-related metabolic patterns. In contrast, classification remains limited for more challenging tasks, including ALS vs mimics and bulbar vs spinal onset, with accuracy <60%, indicating that FDG PET alone is insufficient for reliable subtype discrimination in these settings. The chapter also demonstrates pronounced scanner and reconstruction effects, with high accuracy in distinguishing PET/CT systems within the same clinical subgroup, underscoring the importance of harmonized acquisition and reconstruction for multi-centre applications across different PET/CT systems. Finally, Chapter 5 shifts to network-level analysis by constructing subject-specific metabolic connectivity networks from static 18F-FDG PET using Wasserstein distance (WD) and compares this framework with KLSE-based connectivity, demonstrating very similar metabolic connectivity networks. Despite substantial dimensionality reduction from three-dimensional volumes to adjacency matrices, connectivity-based representations achieve ALS-versus-healthy classification performance comparable to voxel-wise models and improve discrimination for ALS versus ALS mimics. Graph metrics derived from these networks are interpretable in relation to regional uptake patterns, with WD-based networks showing slightly greater sensitivity to distributed disease-related changes in uptake patterns. Overall, this thesis demonstrates that applying machine learning-based methods to PET kinetic and quantitative analysis can enable fast and accurate voxel-wise parametric imaging, support data-driven selection of pseudo-reference regions, clarify the potential and limitations of FDG PET for ALS phenotyping, and provide compact yet informative network representations of metabolic connectivity.