Skip to Main Content

An official website of the United States government

Self-Supervised Multimodal Medical Image Analysis

Principal Investigator
Name
Cristian Simionescu
Degrees
Ph.D. Student
Institution
Alexandru Ioan Cuza University
Position Title
Ph.D. Student
Email
About this CDAS Project
Study
NLST (Learn more about this study)
Project ID
NLST-796
Initial CDAS Request Approval
May 21, 2021
Title
Self-Supervised Multimodal Medical Image Analysis
Summary
Use state-of-the-art self-supervised deep learning methods such as DINO (DINO: Emerging Properties in Self-Supervised Vision Transformers), Image-GPT, and BYOL (Bootstrap your own latent: A new approach to self-supervised Learning) to investigate the performance of these algorithms on medical data. The project aims to use multiple modalities of medical images, from MRI's to CT's and MG's of various parts of the human body: brain, chest, knee, breast, etc. We want to develop training techniques and models that, after being trained on a large unlabeled collection of datasets, can outperform models trained from scratch on downstream tasks.
Aims

- Identifying the best performing self-supervision methods for medical images and adapt them to the domain, for example applying them on 3D volumes
- Develop a domain adapted self-supervised training procedure and models that make use of large unlabeled datasets in order to outperform models trained from scratch on downstream tasks

Collaborators

Alexandra Hanganu - "Alexandru Ioan Cuza" University, Faculty of Computer Science, Iasi, Romania
Ingrid Stoleru - "Alexandru Ioan Cuza" University, Faculty of Computer Science, Iasi, Romania
Robert Herscovici - "Alexandru Ioan Cuza" University, Faculty of Computer Science, Iasi, Romania
Georgiana Coca - "Alexandru Ioan Cuza" University, Faculty of Computer Science, Iasi, Romania
Dr. Prof. Adrian Iftene - "Alexandru Ioan Cuza" University, Faculty of Computer Science, Iasi, Romania