Les commentaires sont clos.

Motion estimation in 2D echocardiography based on deep learning

1 Septembre 2022

Catégorie : Post-doctorant

18-months post-doc 2022 in Lyon, France

Motion estimation in 2D echocardiography based on deep learning

This work is funded by the ANR project named ORCHID (French national grant). Part of this project is also in collaboration with the VITALab, University of Sherbrooke (Canada).


Medical and overall project context

Echocardiography plays a paramount role in day to day clinical practice. It is the modality of choice to highlight structural or functional dysfunction of the heart. The observed abnormalities combined with clinical data of the patient lead to diagnoses, many of which requiring the performance of complementary examinations to obtain a specific etiology (cause) of the cardiac pathology which will modify the prognosis of patient and the subsequent therapy. Some of these examinations are useless and do not find any specific abnormality, resulting in constraints for the patients and a cost for the society. The ultimate goal of the ORCHID project is to develop rigorous and explainable models for the prediction of the origin of cardiac diseases from echocardiography based on artificial intelligence (AI) paradigm. To this aim, high-dimensional descriptors of the cardiac function needs to be extracted from echocardiographic sequence to efficiently characterize targeted pathologies.

Scientific context of the postdoctoral project

Several studies have shown that myocardial deformations are relevant descriptors of the cardiac function [DUC-20]. This high-dimensional information is usually estimated by conventional motion estimation techniques (i.e. block matching-based or optical flow-based methods) that suffer from difficulties inherent in ultrasound images, such as artifacts (shadow, reverberation), lack of information or speckle decorrelation. This results in a lack of accuracy and reproducibility in current embedded solutions, which limits their use in clinical practices.

In this context, two recent studies have proposed to solve this issue thanks to deep learning (DL) methods [OST-21] [EVA-22]. In particular, based on pre-trained models on natural synthetic images, we have recently shown that it is possible to perform efficient transfer learning from realistic ultrasound simulations [EVA-22]. Results from 60 real patients acquired with two different ultrasound scanners show that dedicated DL models can obtain global longitudinal deformations calculated from the estimated motion fields with an average error of less than 3%, which is four times better than the results obtained with state-of-the-art methods. These preliminary results confirm that, as with segmentation, DL methods have the strong potential to revolutionize motion/deformation estimation in echocardiographic imaging.

Methodological objectives

We propose here to design a framework able to compute robust and accurate myocardial deformation maps based on a DL technique for motion estimation in ultrasound imaging. This will be achieved through three main steps.

  1. We will first reinforce an existing simulation pipeline developed in our team (matlab/C code) to generate the largest and richest synthetic dataset in terms of diversity of cases, pathologies and acquisition view. A virtual cohort of 3000 B-mode sequences is targeted. Generating such synthetic dataset will require a large computational power, but will have to be performed only once and could be parallelized on clusters. We have access to several clusters (CNRS IN2P3, CNRS Jean-Zay, CREATIS lab cluster) and are experienced in this domain.
  2. This synthetic dataset will then be used to feed DL models whose architectures are adapted to optimize the motion estimation task. We already have a pipeline in place in echocardiographic images [EVA-22], and the goal of this project is to investigate new solutions such as recurrent networks, whose architectures currently produce among the best results on natural scene images [TEE-20].
  3. Finally, we will exploit our already existing DL method on myocardial region segmentation [PAI-22] to compute efficient deformation maps over the full cardiac cycle. Starting from an initial mesh automatically generated from the first segmentation mask, we will propagate it on the whole sequence from an optimization algorithm that will involve three terms related to the estimated motion from our DL solution, the boundary conditions from the segmentation masks and a smoothing regularization term.


The recruited person will work in a multidisciplinary team composed of medical cardiologist experts, researchers and computer scientists of CREATIS laboratory (Lyon, France).

Expected skills and other information

  • Expected skills: computer sciences, artificial intelligence
  • Technical skills: Python (Pytorch/Tensorflow) and C programming are required
  • Knowledge of ultrasound imaging will be appreciated
  • An interest in the field of health is essential
  • English and/or French
  • Expected start: October 2022
  • Duration: 18 months
  • Salary: ~ 2255€/month (net)
  • Location: CREATIS Laboratory, Lyon, France
  • Send CV to: and