Project

Canon Medical/Royal Academy of Engineering Research Chair in Healthcare AI

In the UK alone, currently 7 million people live with cardiovascular disease and this number will increase as the population ages. Under-resourced and under-staffed healthcare systems are struggling with the rising caseload and the large volumes of information being generated. Currently, excitement in Artificial Intelligence (AI) for healthcare is high, because of its potential to help stem this information overload and reduce healthcare costs.

The AI paradigm fuelling this excitement heavily depends on well-curated training data and is largely seen as a black box. In contrast we will:

learn from complex, multimodal, healthcare records with minimal supervision;
focus on problems underpinning learning causal data representations optimised to provide a transparent base for the desired diagnoses and predictions. We will then translate these techniques to automated estimation of cardiac biomarkers, disease diagnosis, and most ambitiously, cardiac episode prediction, thus opening roads to preventive care.

News

Tutorial: Diffusion Models for Medical Imaging, MICCAI Tutorial, October 2023.
Workshop: Co-organizing “Medical Applications with Disentanglements”, Workshop, MICCAI, 2022. link
Workshop: Co-organizing “Domain Adaptation and Representation Transfer”, Workshop, MICCAI, 2022. link
Invited talks: Big AI in healthcare, Upenn (May 2021), DKFZ (March 2022)
Keynote: Algorithms to do more with less?, International Conference on Image and Vision Engineering IMPROVE 2021.
Tutorial: DREAM: Disentangled Representations for Efficient Algorithms for Medical data, MICCAI Tutorial, MICCAI, September 2021. This year, the tutorial has an associated paper.
Keynote: Doing More with Less by Disentangled Data Representations, Keynote DART, MICCAI, September 2021.
Tutorial: DREAM: Disentangled Representations for Efficient Algorithms for Medical data, MICCAI Tutorial, MICCAI, October 2020.
Keynote: Doing More with Less by Disentangled Data Representations, Keynote DART, MICCAI, October 2020.
Event: Sotos invited to present at The Global AI Summit, organized by the Saudi Data and Artificial Intelligence Authority, March 2020. [postponed due to COVID-19]
News: How can AI help? Reading the language of medicine, Canon Medical Blog, October 2019.
Event: Public Event in Japan, British Embassy. AI and Healthcare, October 2019.
Keynote: Disentangled Representation Learning in Medical Imaging, Keynote STACOM, MICCAI, October 2019.
Tutorial: Simulation and Synthesis, Tutorial at International Conference on Acoustics Speech and Signal Processing (ICASSP), May 2019.
News: Interview in Greek Huffington Post, September 2019.
Event: Intervention at Future-proofing society - how digital health and social care can empower and transform lives, Scottish Parliament and IET Event.
News: Prof. Tsaftaris featured in School of Engineering News.
News: Prof. Tsaftaris featured in RAEng News.
Kick-off: We started a new project.

Publications

IEEE Transactions on Medical Imaging

Learning to Segment From Scribbles Using Multi-Scale Adversarial Attention Gates

Gabriele Valvano, Andrea Leo, Sotirios A. Tsaftaris

News

Publications

Learning to Segment From Scribbles Using Multi-Scale Adversarial Attention Gates

Pseudo-healthy synthesis with pathology disentanglement and adversarial learning

Diffusion Models for Causal Discovery via Topological Ordering

Indication as Prior Knowledge for Multimodal Disease Classification in Chest Radiographs with Transformers

CTR: Contrastive Training Recognition Classifier for Few-Shot Open-World Recognition

OMASGAN: Out-of-distribution Minimum Anomaly Score GAN for Anomaly Detection

vMFNet: Compositionality Meets Domain-Generalised Segmentation

What is Healthy? Generative Counterfactual Diffusion for Lesion Localization

HSIC-InfoGAN: Learning Unsupervised Disentangled Representations by Maximising Approximated Mutual Information

Why Patient Data Cannot Be Easily Forgotten?

Diffusion Causal Models for Counterfactual Estimation

Semi-supervised Meta-learning with Disentanglement for Domain-Generalised Medical Image Segmentation

Controllable Cardiac Synthesis via Disentangled Anatomy Arithmetic

Have You Forgotten? A Method to Assess if Machine Learning Models Have Forgotten Data

Measuring the Biases and Effectiveness of Content-Style Disentanglement

INSIDE: Steering Spatial Attention with Non-imaging Information in CNNs

Disentangled Representations for Domain-Generalized Cardiac Segmentation

Max-Fusion U-Net for Multi-modal Pathology Segmentation with Attention and Dynamic Resampling

Semi-supervised Pathology Segmentation with Disentangled Representations

Consistent Brain Ageing Synthesis

Multimodal cardiac segmentation using disentangled representations

Temporal Consistency Objectives Regularize the Learning of Disentangled Representations

Conditioning Convolutional Segmentation Architectures with Non-Imaging Data

People

Funding