Doctoral Thesis: Towards Label-Efficient Visual Learning for Biological Image Analysis

Presenter: Xinyi Gu
Presenter’s Affiliation (CSAIL, RLE, LIDS, MTL, etc.): PILM (The Picower Institute for Learning and Memory)
Thesis Supervisor(s): Prof. Kwanghun Chung

Details

• Date: Tuesday, October 13
• Time: 11:00 am - 1:00 pm
• Location: 46-6199

Additional Location Details:

Abstract: Modern AI systems have advanced rapidly by scaling models and training data, yet in many high-value but specialized biological imaging domains, the availability of expert-annotated data remains a major bottleneck, limiting the development of generalizable AI models. This thesis addresses this challenge by developing label-efficient visual learning approaches that leverage unlabeled images as an alternative source of supervision. Using 3D brain images acquired by light microscopy as a model system, we develop a framework that learns transferable representations through masked self-supervised pretraining and adapts them to downstream tasks, including cell detection and neurite segmentation, with limited task-specific annotations. We demonstrated that these learned representations substantially reduce annotation requirements, improve robustness across imaging conditions, and generalize across imaging platforms. Together, these results establish self-supervised representation learning as a scalable strategy for developing label-efficient AI systems for biological image analysis.

Host

• Xinyi Gu
• Email: gxy@mit.edu