PhD defense Ali Keshavarzi: Deep Learning for Anatomically-Consistent Airway Tree Modeling

Wednesday 17 December, 2025, at 14:00 (Paris time) at Télécom Paris

Télécom Paris, 19 place Marguerite Perey F-91120 Palaiseau [getting there], amphi 2 and in videoconferencing

Full title: Deep Learning for Anatomically-Consistent Airway Tree Modeling: Visual Features, Loss Functions, Learning Strategies, and a Novel Detection Task

Jury

Hervé Delingette, Research Director (HDR), Inria Sophia Antipolis – Méditerranée (Reviewer)
Petr Dokladal, Senior Researcher (HDR), MINES Paris – PSL (Reviewer)
Catalin Fetita, Professor (HDR), Télécom SudParis, Institut Polytechnique de Paris (Examiner)
Maria A. Zuluaga, Professor (HDR), EURECOM (Examiner)
Elsa Angelini, Professor (HDR), Télécom Paris, Institut Polytechnique de Paris (Thesis supervisor)

Abstract

The human airway tree is a complex branching network whose geometry and connectivity are tightly linked to pulmonary function. Accurately modeling this tree from chest CT is essential for quantitative analysis, disease monitoring, bronchoscopic planning, and digital lung twins, yet remains difficult due to fine distal bronchi, anatomical variability, and heterogeneous CT acquisition protocols. Standard deep learning methods often fail to reconstruct a continuous airway tree and to generalize across image cohorts, while annotated data is scarce.

Learn more

This PhD thesis work proposes complementary contributions to improve robustness, anatomical fidelity, and generalization in airway-tree modeling, combining deep-learned segmentation and detection tasks. First, convolutional sparse priors act as data-driven structural biases, learning recurring airway patterns and producing compact, geometry-aware visual representations particularly beneficial for few-shot and cross-domain settings. Second, the Boundary-Emphasized Loss (BEL) is proposed to reinforce morphological constraints derived from boundaries and reduce distal branch breakages through dynamic weighting. BEL offers a more robust alternative to classic centerline-based loss constraints. Third, a complexity-based curriculum domain adaptation framework is proposed, using a novel scan-level complexity score to guide supervised learning from simple to challenging cases and improve transfer learning from healthy to pathological cohorts.

Overall experiments on healthy and diseased cohorts show improved distal reconstruction, topological continuity, and cross-domain robustness, providing a general foundation for topology-aware, structure-preserving modeling of tubular structures in medical images with limited annotated data.
Finally, the « BifDet » cohort is introduced as the first large-scale 3D airway annotated dataset for bifurcation detection, which we benchmarked with two deep-learned methods.

All the PhD defenses