Abstract - 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 eld. If there are explicability methods for classication, 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 particial cases to demonstrate the impact of the results on clinical practice.

Resume - La segmentation automatique des images medicales joue un r^ole preponderant pour le diagnostic et la therapeutique. Les reseaux de neurones convolutionnels profonds (CNN) representent l'etat de l'art, mais presentent des limites, notamment sur la plausibilite des segmentations generees. Notre hypothese est que l'amelioration des segmentations viendra de l'ajout d'information externe, via la connaissance medicale par exemple, et de t^aches auxiliaires, comme le recalage, qui permettront de guider et contraindre la segmentation. D'autre part, le c^ote ininterpretable des CNN freine leur usage dans le milieu medical. S'il existe des methodes d'explicabilite pour la classication, tout reste a faire pour la segmentation. Nous viserons a developper de telles methodes, pour comprendre les mecanismes sous-tendant l'ajout de connaissances et de t^aches. M^eme si nos developpements seront generiques, nous viserons des exemples concrets pour demontrer l'impact des resultats sur la pratique clinique.

Référence Projet :  ANR - AAPG 2021

Période : A partir de Mars 2022...

Financeurs : ANR

Porteur : Caroline Petitjean (LITIS, URN)

Page www du projet : ici

Responsable Scientifique du LMI : C. LE GUYADER (LMI)

Autres participants : LITIS (C. Petitjean) et IMVia (F. Meriaudeau)

Budget du projet / Budget total : 402 000 €  (dont 150 000 pour le LMI)

Participants LMI : C. Le Guyader (responsable INSA), Diane Chan Sock Line, N. Forcadel, C. Gout,  Z. Lambert.

Recrutement dans le cadre du projet : Diane Chan Sock Line (Doctorante, 3 ans): Inclusion of Priors in Deep Networks for Medical Image Segmentation

---

Equipement dans le cadre du projet : DELL précision 7760                                                       

---

LMI is mainly involved in WP2 : Joint image registration and segmentation

Leader: Carole Le Guyader (LMI)
Contributors: N. Forcadel, C. Gout and Z. Lambert (LMI), C. Petitjean and G. Gasso (LITIS)
Objectives The purpose of this WP is to design joint segmentation/registration models combining variational and deep-learning based techniques, with relevant and differentiable energy function that allow to assess the mechanical plausibility of the transformation.