Pulmonary embolism risk stratification based on vascular network modeling

Pulmonary Embolism (PE), the third most frequent acute cardiovascular syndrome, results from the ostruction of pulmonary arteries by blood clots. Assessing its severity requires characterizing the location, extent, and distribution of thrombi within the Pulmonary Arterial Tree (PAT). Current clinical scores, such as Qanadli and Mastora, rely on simplified anatomical templates and time-consuming manual measurements, limiting their routine use. Additionally, key cardiac biomarkers required for risk stratification are sometimes unavailable in emergency settings. This thesis addresses these limitations by developing an automated pipeline for graph-based modeling of the PAT from Computed Tomography Pulmonary Angiography (CTPA), enabling the extraction of patient-specific imaging-derived biomarkers for PE severity assessment. This work was conducted within the PERSEVERE project, using a cohort of 353 CTPA acquisitions. Four main contributions are presented. First, VSMOD, an open-source 3D Slicer plugin, enables semi-automated vascular annotation, reducing annotation time by approximately 80% while preserving topological consistency. Second, a deep learning-assisted iterative annotation strategy compensates for the lack of annotated data, producing robust segmentation models for pulmonary arteries and emboli across the full cohort. Third, a fully automated graph-based pipeline converts vascular and thrombus masks into an anatomically oriented directed graph, from which local and global biomarkers are extracted, including automated adaptations of existing thrombotic burden scores showing strong agreement with manual measurements. Fourth, the framework is applied to neurovascular imaging, demonstrating its broader applicability by quantitatively comparing vessel segmentation between Dual-Energy CT and Spectral Photon Counting CT. Together, these contributions establish a complete, reproducible workflow from raw CTPA data to structured clinical biomarkers, laying the foundation for automated, patient-specific PE severity assessment and future prognostic studies.

Keywords : pulmonary embolism; machine learning; segmentation.

Jury : 

CONZE Pierre-Henri, PU, IMT Atlantique, Rapporteur

DELINGETTE Hervé, Directeur de recherche, Inria Sophia Antipolis, Rapporteur

BERTOLETTI Laurent, PUPH, Université Jean Monnet Saint-Étienne, Examinateur

JACOB Joseph, PU, University College London, Examinateur

ZULUAGA Maria A., PU, EURECOM, Examinatrice

FRINDEL Carole, Maître de conférence, HDR, INSA Lyon, IUF, Directrice de thèse

MERVEILLE Odyssée, Maître de conférence, INSA Lyon, Co-directrice de thèse