AI technology: Vision (image recognition, image generation, text recognition OCR, etc.) and Deep Learning
The goal of this project is to reach a new frontier in precision radiology through novel deep learning techniques applied at unprecedented scale.
By using Self-Supervised Learning (SSL) techniques, such as DINO and CLIP, on a dataset of over 1 billion anonymized radiological images, this initiative seeks to overcome the significant challenge of requiring extensive annotations for medical image analysis, a major barrier in applying deep learning technologies to radiology.
The scientific contribution of this work lies in the adaptation and scaling of SSL methods to an unprecedentedly large multimodal medical dataset, demonstrating the potential to improve diagnostic accuracy, reduce the time burden on radiologists, and facilitate a deeper understanding of various pathologies.
This project has the potential to greatly enhance the quality and accessibility of healthcare by supporting the development of advanced diagnostic tools and personalized medicine. It pushes the boundaries of what is possible in computer vision and AI in healthcare.
This project will lead to the submission of findings to a high-impact journal like Nature Medicine, demonstrating the scalability of SSL in medical imaging and its application to various downstream tasks with limited labeled data. Additionally, the release of some model weights will enable further research and development in the field.
Ronot Maxime, Institut national de la santé et de la recherche médicale (INSERM) - France