Post-Doc. 3D MRI image segmentation using machine learning techniques

3 Novembre 2023

---

Catégorie : Post-doctorant

---

Project :

Spina bifida, a malformation visible from birth and characterised by poor development of the nervous system and spinal column, can have a number of consequences for different organs, including the bladder, leading to multiple disabilities. Depending on the severity of the condition and its symptoms, a wide range of medical and surgical specialities are involved in its management. Indeed, a case of spina bifida considered severe by the specialists requires surgical intervention, hence the interest in studying the numerical characterisation of the bladder wall in order to identify the degree of severity of this disease, useful information for the decision of the medical team. In order to characterise the bladder wall and differentiate between the different cases of spina bifida, a segmentation stage (trimming) and then characterisation of the bladder wall are necessary.

The aim of this research project is to develop digital methods to help characterise the bladder wall on 3D MRI images, based on image processing techniques and machine learning approaches, in order to provide healthcare professionals with a computer-assisted diagnostic tool.

The main scientific hurdles that will need to be overcome are, on the one hand, the 3D segmentation of MRI images using semi-supervised learning techniques based on deep neural network architectures and, on the other hand, the choice of relevant radiomic attributes for characterising the bladder wall. One of the major problems of neural networks is the annotation of a non-negligible quantity of medical images by specialists in order to satisfy the learning phase of the automatic segmentation process. As part of this project, we propose to use minimal 2D annotation (qq images out of the hundred or so that make up a 3D MRI) and to propagate the process via a deep convolutional network-type architecture combining image registration and 2D segmentation. Secondly, in order to better understand the correlation between disease severity and the numerical attributes extracted from the regions of medical interest segmented in the previous phase, we plan to extract morphological, statistical and 3D texture characteristics and use an auto-encoder neural network to select and rank the most relevant attributes.

Beyond the case of the bladder, the aim is to develop a general methodology for 3D segmentation by propagation of 2D segmentation and image registration, using minimalist semi-supervised learning, without the need for a medical team to annotate a considerable mass of data.