Where: Ummon HealthTech (80% of the time) and ImViA (Imagerie et Vision Artificielle) laboratory (20% of the time), Dijon, France
Contract: two years contract, starting from September 2022
Salary: 35 k€ gross per year

Project description

Deep learning algorithms have revolutionised many scientific fields, including medical imaging. However, although results in a controlled environment (same acquisition device, same acquisition protocol, monocentric database, etc.) are often very good, generalisation to new data from other acquisition devices and/or other acquisition protocols and/or coming from other centres and/or including other pathologies suffer from data shift, decreasing the performance. Ummon HealthTech (https://www.ummonhealthtech.com/en/ummon-healthtech/) has started to investigate the impact of data shift on the performance of deep learning algorithms dedicated to histological analysis. The algorithm developed in a recent study [1] explores the use of domain adversarial training [2] to reduce the impact of data shift due to sample preparation in different centres (showing the same semantic but different intensity distributions). Moreover, it has been shown that data from other pathologies (showing different semantics) can be used to improve the robustness of convolutional neural networks. The company recently filed a patent on a method estimating the confidence in the prediction of a neural network to measure the consequences of data shift on algorithms used in clinical routine.

The software research engineer will first have to investigate and quantify the consequences of data shift by constructing benchmarks tasks. Their second objective will be to implement and characterise the methods developed by the company in terms of performance and stability, according to the application and to potential data shift. Finally, the third objective will be to improve these approaches by analysing the obtained results and the state-of-the-art.

Tasks description

The engineer will be responsible for several developments and R&D tasks, including:

• the development of deep learning pipelines, focusing on image processing and classification algorithms
• the development of labelling and quality control pipelines, including unit, functional, and performance tests, as well as integrity checking algorithms
• the development of specific data management, verification, and query libraries, as well as security protocols for communications and data transfers

The software research engineer will also be responsible for writing documentation and participating in scientific and technical monitoring.

Person specification

Candidates must hold an engineering degree or a master’s degree in computer science, ideally with a specialisation in machine learning. They must have at least six months of experience (end-of-study internship) in deep learning.

Skills

Be familiar with Python and R programming, especially with machine and deep learning libraries (scikit-learn, pytorch, tensorflow, keras)
Be familiar with data augmentation, transfer learning, and optimisation of neural networks
Be familiar with parallel and distributed computing
Be familiar with project management and have a high degree of autonomy
A good level of English (C1 or more) is mandatory

Application

Applications should include a CV and a covering letter outlining your motivation for the position and be sent to Sarah Leclerc (sarah.leclerc@u-bourgogne.fr) and Benoît Presles (benoit.presles@u-bourgogne.fr) before the 6th of July, 12 p.m. Interviews will take place the week of the 11th of July.

Publications

[1] Dumas et al., “Inter-Semantic Domain Adversarial in Histopathological Images”, arXiv, 2022 – https://arxiv.org/pdf/2201.09041.pdf
[2] Ganin et al., “Domain-Adversarial Training of Neural Networks”, Journal of Machine Learning Research, 17, 2016

Closing date: 06/07/2022