Search for a command to run...
Liver cancer remains a leading cause of cancer-related mortality, where early detection is critical for survival. Beyond initial diagnosis, sustained longitudinal monitoring is essential for evaluating treatment response and surveilling for recurrence. Current practice relies on the manual assessment of serial Computed Tomography (CT) scans (called longitudinal follow-up), a time-intensive process complicated by the liver's deformable nature, inter-patient variability, and the heterogeneous temporal evolution of lesions.While deep learning offers potential for automation, progress is hindered by the scarcity of annotated longitudinal datasets and the complexity introduced by anatomical variability and temporal changes. This thesis addresses these challenges through a methodological pipeline for the automated analysis of liver lesions in CT, encompassing the utilization of synthetic data and the integration of temporal context. The approach focuses on three key objectives: reducing dependence on manual annotation, establishing robust anatomical alignment across time-points, and leveraging temporal context to enhance lesion detection.First, to overcome the bottleneck of data scarcity, we introduce a data-efficient training strategy based on the generation of synthetic tumors. By integrating expert domain knowledge into a parameterized generation pipeline, we demonstrate that models pre-trained on synthetic data achieve strong generalization, thereby reducing the reliance on large-scale manual annotation.Second, to ensure observations across scans correspond to the same anatomical context, we propose a structure-preserving registration approach. Unlike standard intensity-based methods that can distort lesions, this approach accommodates liver deformation while preserving the morphology of internal structures, ensuring a reliable foundation for comparative analysis.Building on these components, we introduce LiFE-Net, a novel deep learning architecture designed to exploit temporal context. By effectively aggregating information from prior examinations through feature-level fusion and self-attention mechanisms, the model enhances detection performance, particularly for small or subtle lesions that are often missed in single-timepoint analysis.Finally, the clinical relevance of the framework is evaluated through a multi-reader study, assessing the concordance between automated measurements and expert consensus.Collectively, these contributions advance automated, consistent, and efficient liver cancer monitoring, while offering extensible strategies for broader longitudinal imaging applications.