Machine learning for medical applications, and in particular supervised deep learning, imposes constraints that are difficult to verify: the results must be interpretable, of good quality, reproducible and auditable. It is also known that performance depends on the basic quality of the learning base and annotations.
At the same time, learning about medical data imposes the need to respect the privacy of patients. In Europe, medical data are in restricted access. The combination of these factors makes the use of deep learning for imaging difficult in practice.
This research project aims to find innovative solutions to these issues, adding an important application context: the use of these techniques in the field of immunotherapy in cancerology.
The objective of this project is to propose an approach based on the very structure of the architectures used, allowing the a priori regularization of the trained networks in order to allow a faster, safer and more efficient convergence of the proposed antagonistic generative networks.
In terms of application, the project proposes to implement these networks in the classification of tumour lesions in ultrasound and CT scans, in order to obtain a rapid and objective assessment of the response of patients to immunotherapy treatment in cancerology.
The project is therefore ambitious because data and annotations are difficult to obtain in this field. A CVN-LR4M collaboration will make it possible to produce a sufficient quantity of these data to initiate training and demonstrate that the approach is feasible, through the collaboration of the two teams within the Gustave-Roussy Institute. At the application level, the objective will be to demonstrate the good properties of the networks obtained, by applying them to the problem of obtaining (1) segmentation of lesions in ultrasound and CT scans; (2) characterization of these lesions in the context of immunotherapy. A conditional extension of the proposed deep learning architecture could allow this type of application.