Projects as PI
- PI of MEDISeg: ANR funded project (2021-25) with LITIS, LMI (INSA de Rouen) and ImVIA (Université de Bourgogne)
The automatic segmentation of medical images plays an important role in diagnosis and therapy. Deep convolutional neural networks (CNN) represent the state of the art, but have limitations, particularly on the plausibility of the generated segmentations. Our hypothesis is that the improvement of segmentations will come from the addition of external information, via medical knowledge for example, and auxiliary tasks, such as registration, which will guide and constrain the segmentation. On the other hand, the uninterpretable nature of CNN hinders their use in the medical field. If there are explicability methods for classification, everything remains to be done for segmentation. We will aim to develop such methods, in order to understand the mechanisms underlying the addition of knowledge and tasks. Although our developments will be generic, we will target use cases to demonstrate the impact of the results on clinical practice.
- PI of the DAWmal project, a STIC AmSud project 2021-22 with Violeta Chang from the Universidad de Santiago de Chile, Mauricio Cerda from Universidad de Chile
Cell segmentation and classification share a central common problem: the lack of labeled images for training automatic methods. In this regard, adequately addressing the shortcomings of current computational approaches and enabling the clinical use of decision support tools requires training and validation of models on large-scale datasets representative of the wide variability of cases encountered every day in the clinic. At that scale, it is impossible to rely on costly and time-consuming manual annotations. Therefore, to strengthen the application of state-of-the-art segmentation and classification methods in biomedical applications, two strategies are proposed: domain adaptation and weakly supervised machine learning.
Motivated by the promising prospects of deep learning-based segmentation and classification and fueled by the lack of fully annotated training data, in this project, we are interested in studying domain adaptation and weakly supervised methods to train neural networks for cell segmentation and classification. In this sense, we will investigate how to leverage self-training and co-training to train high-quality cell segmentation algorithms using weak labels taking advantage of domain adaptation techniques. Therefore, we will explore weak supervision (extremely one-point annotation per cell), aiming to obtain segmentation performance close to full supervision (mask annotation for each cell) but with the lowest human annotation effort. On the other hand, we will investigate strategies for domain adaptation for cell classification using weak labels. Our focus will be on applications in human sperm cells, cancer biopsies, blood parasites, and neuronal morphologies.
- PI of the WeSmile project, a PHC Van Gogh project 2019-20 with Veronika Cheplygina from TU Eindhoven, NL
- Leader of the Alveo project, funding by Ligue contre le cancer 60K€, 2008 and ADIR (patients association) 30k€, 2009
Projects as participant:
- ANR Icub, APi
- Medical image segmentation WP leader in the M2NUM project, funded by the Normandy region (66k€) and FEDER funding (103400k€), 2014-18
- Involved in the STIC-AmSud Project “Dynamic Selection of Classifiers with Applications in Real Environments” with UFPR and PUCPR (Brazil) and Universidad de Chile, PUC Chile and Orand (Chile), 2014-2015