Description
Date depot: 1 mars 2023
Titre: Anomaly detection in pulmonary vessels using few shot learning
Directeur de thèse:
Thierry GERAUD (LRE)
Encadrante :
Elodie PUYBAREAU (LRE)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Images et vision
Resumé: This research project is a CIFRE proposal presented by General Electric Healthcare (GEHC) in
collaboration with Epita, the industrial and academical partners of this project respectively. The GEHC
scientific supervision will be ensured by Laura Dumont, currently Senior Software Engineer and Jorge
Hernandez Londono, Staff Software Engineer. The administrative supervision will be ensured by
Baptiste Perrin, manager of AI-DReAM team in the AV department. The Epita scientific direction will
be ensured by Elodie Puybareau, AI and Computer Vision assistant professor, and Thierry Géraud, AI
and Computer Vision professor. This project will be conducted with the help of the Kremlin-Bicêtre hospital. The Kremlin-Bicêtre
hospital from AP-HP houses the French Pulmonary Hypertension Reference Centre, a world leader in
the field of PH education, research, and care. It benefits from an exceptional positioning to offer a global
approach for chronic thromboembolic pulmonary hypertension (CTEPH) detection, assessment,
treatment, and translational research.
This research project falls into the domain of image processing and machine learning for healthcare.
More precisely in developing automated solution to analyze medical images (CT scan) to support
thoracic clinical workflow. This thesis is part of a work package included in RHU Destination 2024,
which aims to better detect, assess, and treat CTEPH.
Venous thromboembolism, clinically presenting as deep vein thrombosis or pulmonary embolism (PE),
is globally the third most frequent acute cardiovascular syndrome behind myocardial infarction and
stroke. Annual incidence rates for acute PE range from 39 to 115 per 100,000 (Konstantinides SV,
2014). In France, at least 35,000 episodes of acute PE are diagnosed every year. In 1 to 3% of cases,
abnormal persistent obstruction of proximal or distal pulmonary arteries by residual organized thrombi,
combined with a variable degree of small pulmonary vessels disease, may lead to CTEPH, causing
exercise limitation, right heart failure and premature death in more than 50% of untreated patients within
5 years of diagnosis (Kim NH, 2019) (Delcroix M, 2018) (Galiè N, 2016).
This severe cardiovascular condition is a major clinical challenge: the insufficient healthcare
professionals’ awareness leads to delayed diagnosis and management, thus resulting in significant
morbidity and mortality. Recent international data reported survival rates of 92, 75, and 60% at 1, 3,
and 5 years, respectively, highlighting the need for better treatment strategies in inoperable CTEPH
(Delcroix M, 2018) (Kim NH, 2019). Patients followed through the national PH Registry (12,210 PH
patients, including 2,610 CTEPH cases) of the French PH Reference Centre suffer from severe disease.
Most of them display marked exercise limitation, poor quality of life, frequent hospitalizations and
reduced life expectancy (Kim NH, 2019) (Delcroix M, 2018) (Hoeper MM, 2014), underscoring the
need for better treatment strategies.
A first clinical presentation of CTEPH may actually mimic acute PE. Careful analysis of initial
Computed Tomography Pulmonary Angiography (CTPA) at the time of PE diagnosis may identify
signs of coexistent CTEPH. However, because of insufficient awareness of CTEPH, these cases are
usually missed, and patients then develop progressive and life-threatening right-heart failure. These
index acute PE episodes are opportunities of early diagnosis of CTEPH.
This project aims to help to detect, assess and treat CTEPH, a severe complication of PE. From a
technical point of view, the aim will be to investigate Machine Learning and Deep Learning methodscombined with anatomical and/or medical priors, and classical image processing techniques such as mathematical morphology. The objective is to provide an easily reproducible and explainable method
to detect and predict this pathology. A special focus on few shot learning and anomaly detection
techniques will be address in this thesis. A dataset publicly available for research will be used for this work, containing 96 540 images that were annotated as Positive for PE. The database was
presented in the 2020 RSNA Pulmonary Embolism Detection Challenge invited researchers to develop
machine-learning algorithms to detect and characterize instances of PE on chest CT studies.
Delcroix M, e. a. (2018). Risk assessment in medically treated chronic thromboembolic pulmonary
hypertension patients. Eur Respir, 52:1800248.
Galiè N, H. M. (2016). 2015 ESC/ERS Guidelines for the diagnosis and treatment of pulmonary
hypertension. Eur Heart J, 37:67-119.
Hoeper MM, e. a. (2014). Chronic thromboembolic pulmonary hypertension. Lancet Respir Med,
2:573-82.
Kim NH, D. M. (2019). Chronic thromboembolic pulmonary hypertension. Eur Respir J, 53:1801915.
Konstaninides SV, H. M. (2014). Guidelines on the diagnosis and management of acute pulmonary
embolism. Eur Heart J, 35:3033-73.
Doctorant.e: Descarpentries Adam