Suggestions
Kit Bransby
AI-based Cardiac Image Computing, PhD Student @ QMUL
Kit Bransby is a Teaching Fellow at Queen Mary University of London (QMUL) in the School of Electronic Engineering and Computer Science.13 He is an academic staff member with a focus on teaching and research in computer science and related fields.
Academic Role
As a Teaching Fellow, Kit Bransby is involved in various aspects of education and research at QMUL:
- Thesis Supervision: He supervises thesis projects for both undergraduate (BSc) and postgraduate (MSc) students, guiding 8 BSc and 10 MSc students in their research endeavors.4
- Teaching Responsibilities: Bransby runs tutorials and delivers lectures as part of his teaching duties.4
- Research: He is actively engaged in research, with a particular interest in computer vision and medical imaging applications.
Research Contributions
Kit Bransby has made several contributions to the field of computer science:
- He has 11 research works with 5 citations to his name, indicating an active research profile.6
- His research interests include echocardiography and machine learning applications in medical imaging.
- Bransby has presented work at international conferences, such as MICCAI 2023, where he shared research on "Joint Dense-Point Representation for Contour-Aware Graph Segmentation".5
Professional Profile
Kit Bransby's professional profile showcases his expertise and skills:
- He is described as someone who "combines creativity with hard work and business sense" in his professional endeavors.2
- His email contact at QMUL is [email protected].uk.3
- Bransby's office is located in room Eng 153 at QMUL.3
Kit Bransby's work at Queen Mary University of London demonstrates his commitment to both education and cutting-edge research in computer science and related technologies.
Highlights
Excited to have a second paper accepted at MICCAI, this time for an oral presentation at the ASMUS workshop.
✨ We address the task of updating a trained echocardiography view classifier with a stream of new data from multiple sites. This is essential to avoid model drift, but challenging as the time and cost of re-training multiple times is considerable, labels vary between sites, and licensing agreements may restrict the transfer of multi-site data to a single server.
✨ We propose a class-incremental learning method which learns an expert network for each dataset, and combines all expert networks with a score fusion model. The influence of “unqualified experts” is minimised by weighting each contribution with a learnt in-distribution score.
✨ Our work is validated on six datasets from over 20 sites, demonstrating significant reductions in training time while improving view classification performance.
Big thanks to my co-authors at Ultromics: Agis Chartsias, Alberto Gomez, Woo Jin Cho Kim, Jorge Miguel Gomes de Oliveira, Arian B., Alex Thorley.
You can read the preprint here:
Very happy that our paper "BackMix: Mitigating Shortcut Learning in Echocardiography with Minimal Supervision" has been accepted at MICCAI 2024.
We show that echocardiography view classifiers learn naive shortcuts from spurious metadata in the image background, and propose an augmentation mechanism to encourage the network to focus on the image content (heart ultrasound).
Pre-print: https://lnkd.in/eA_4vKZ8
Big thanks to my co-authors Alberto Gomez, Agis Chartsias and team at Ultromics
