Advancing Medical Computer Vision Through Teacher-Student Learning and Privacy-Preserved Deep Learning Models
Dr Fani Deligianni (University of Glasgow)
Thursday 23rd January 14:00-15:00 Maths 116
Abstract
This talk introduces novel technical approaches for enhancing AI systems in medical imaging and human motion analysis. At the core of our methodology is an innovative teacher-student architecture that enables effective semi-supervised learning through cross-supervision and multi-scale cross contrastive learning. We demonstrate how this framework optimally leverages both labeled and unlabeled data in medical image segmentation, while introducing certainty-guided mechanisms to improve feature representation. For efficient model deployment, we present a novel knowledge distillation approach using global filters, which significantly reduces computational complexity while maintaining high accuracy in human pose estimation. Finally, our advanced privacy preservation framework that integrates differential privacy with integrated decision gradients, provides robust protection against membership inference attacks while preserving model utility. Through extensive empirical evaluation, we demonstrate how these technical innovations enable the development of lightweight, privacy-aware AI models suitable for real-world clinical deployment.
Add to your calendar
Download event information as iCalendar file (only this event)