AI for healthcare: Building more accurate lung digital twins with less data
Ali Keshavarzi, a PhD candidate at Télécom Paris, is using AI to tackle one of the most complex challenges in medical imaging: modeling the human airway from CT scans. His work focuses on enhancing how we segment and analyze the branching structures of the lungs, key to understanding and diagnosing respiratory diseases like asthma, COPD, and pulmonary fibrosis.
Why this matters
Respiratory diseases affect hundreds of millions of people worldwide, yet clinicians still face major limitations in analyzing the complex structure of the lungs, especially the small and fragile airway branches, on CT scans. Ali Keshavarzi’s research tackles this challenge with AI models that are both more accurate and more efficient. By designing deep-learning tools that work well even with limited annotated data, his work makes high-precision lung modeling more accessible and scalable.
This has the potential to improve diagnosis, support the development of digital twins for personalized medicine, and reduce the workload on radiologists, especially in healthcare systems with constrained resources.
Ali's approach
Ali’s research combines deep learning, sparse encoding, and domain adaptation to overcome these limitations. His innovations include:
Topology-aware loss functions that help preserve the shape and continuity of the airways during segmentation, especially for smaller branches.
Curriculum learning strategies that prioritize training on CT scans based on their complexity, reducing the need for large amounts of annotated data.
BifDet, a new dataset with over 7,500 annotated airway bifurcations, enabling better 3D object detection for tracking disease progression.
Sparse priors and boundary-focused learning to improve the robustness and consistency of airway segmentation models, even with few examples.
“My goal is to create AI tools that work even when we don’t have perfect data. In healthcare, this is often the reality? so we need solutions that are both efficient and reliable.”
— Ali Keshavarzi, PhD Candidate at Télécom Paris
Figure: Lung CT scan highlighting the airway region (left), 3D rendering of the airway tree ground-truth with bifurcation points annotated by us (middle), and zooms on annotated bifurcation bounding boxes for the trachea, right bronchus, and left bronchus (right).
The big picture
This work directly supports Hi! PARIS’ mission to apply cutting-edge AI to real-world health challenges. It shows how data-efficient AI can lead to better models, even in domains where annotated data is scarce.
The goal: scalable tools that can support radiologists, improve diagnoses, and eventually enable better treatment outcomes.
“Beyond performance, we’re also thinking about accessibility,” Ali adds. “I want this technology to benefit patients everywhere, even in settings where computing power or expert-labeled data is limited.”
— Ali Keshavarzi, PhD Candidate at Télécom Paris
Looking ahead...
Ali’s techniques could be extended to other areas of medical imaging, contributing to the development of frugal AI solutions that work across diverse healthcare environments—including those with limited access to expert annotations or computing resources.
This research is supervised by Elsa Angelini (Hi! PARIS affiliate and researcher at Télécom Paris) and contributes to the broader Hi! PARIS effort to accelerate AI-driven innovation in healthcare.
A vision for living with AI
This work is featured in the first chapter of our latest Visions of Research white book, Living with AI.
In this section, we explore how humans are learning to coexist with artificial intelligence and how it can meaningfully improve our lives.
