Xiao Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Xiao Liu has a total of 24 co-authors. Here are the top co-authors by number of shared publications:
- Sotirios A. Tsaftaris
The University of Edinburgh
10 shared publications Spyridon Thermos
Moverse, Thessaloniki, Greece
7 shared publications- Alison Q. O'Neil
University of Edinburgh
6 shared publications Sotirios A. Tsaftaris
University of Edinburgh
5 shared publications- Pedro Sanchez
School of Engineering, University of Edinburgh, Edinburgh, U.K.
4 shared publications Alison Q. O’nEil
University of Edinburgh
3 shared publications
Latest Publications
- Open Access
- 19 March 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (15)
Model-informed double image prior for thermochemical process monitoring using chemical species tomography
Accurate monitoring on the thermochemical processes of combustion-based power generation systems is critical for optimizing combustion efficiency and minimizing harmful emissions. Chemical species tomography (CST), enhanced by in situ sensing, stands out for reconstructing thermochemical parameters in industrial applications. However, the inverse problem inherent in CST is ill-posed, resulting in poor-reconstructed reactive-flow images, even with regularization methods. While most data driven-based algorithms offer potential improvements, their heavy reliance on simulated dataset leads to degraded generalizability in harsh industrial environments. To address this, we propose a model-informed double image prior (MI-DIP), an untrained neural network (UNN) for dataset-free, dual-path reconstruction of two-dimensional thermochemical parameters-temperature and species concentration. This network incorporates the physical model of CST, regularizing the inverse problem with its physical model and stabilizing image reconstruction. This approach is validated through the lab-scale experiments on combustion devices with various burner profiles and further assessed using numerical simulations of real reactive flows, obtained via large eddy simulation (LES) in the fire dynamics simulator (FDS). Results demonstrate that the developed algorithm outperforms either the traditional regularization method or the key UNN-based algorithm, deep image prior (DIP), in terms of higher reconstruction accuracy. In addition, the robustness study and an ablation study are carried out to demonstrate the reliability of MI-DIP in industrial thermochemical processes.
- 19 March 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (15)
Open AccessModel-informed double image prior for thermochemical process monitoring using chemical species tomography
Accurate monitoring on the thermochemical processes of combustion-based power generation systems is critical for optimizing combustion efficiency and minimizing harmful emissions. Chemical species tomography (CST), enhanced by in situ sensing, stands out for reconstructing thermochemical parameters in industrial applications. However, the inverse problem inherent in CST is ill-posed, resulting in poor-reconstructed reactive-flow images, even with regularization methods. While most data driven-based algorithms offer potential improvements, their heavy reliance on simulated dataset leads to degraded generalizability in harsh industrial environments. To address this, we propose a model-informed double image prior (MI-DIP), an untrained neural network (UNN) for dataset-free, dual-path reconstruction of two-dimensional thermochemical parameters-temperature and species concentration. This network incorporates the physical model of CST, regularizing the inverse problem with its physical model and stabilizing image reconstruction. This approach is validated through the lab-scale experiments on combustion devices with various burner profiles and further assessed using numerical simulations of real reactive flows, obtained via large eddy simulation (LES) in the fire dynamics simulator (FDS). Results demonstrate that the developed algorithm outperforms either the traditional regularization method or the key UNN-based algorithm, deep image prior (DIP), in terms of higher reconstruction accuracy. In addition, the robustness study and an ablation study are carried out to demonstrate the reliability of MI-DIP in industrial thermochemical processes.
- 26 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 43 (6) , 2169-2179
Compositionally Equivariant Representation Learning
Deep learning models often need sufficient supervision (i.e. labelled data) in order to be trained effectively. By contrast, humans can swiftly learn to identify important anatomy in medical images like MRI and CT scans, with minimal guidance. This recognition capability easily generalises to new images from different medical facilities and to new tasks in different settings. This rapid and generalisable learning ability is largely due to the compositional structure of image patterns in the human brain, which are not well represented in current medical models. In this paper, we study the utilisation of compositionality in learning more interpretable and generalisable representations for medical image segmentation. Overall, we propose that the underlying generative factors that are used to generate the medical images satisfy compositional equivariance property, where each factor is compositional (e.g. corresponds to human anatomy) and also equivariant to the task. Hence, a good representation that approximates well the ground truth factor has to be compositionally equivariant. By modelling the compositional representations with learnable von-Mises-Fisher (vMF) kernels, we explore how different design and learning biases can be used to enforce the representations to be more compositionally equivariant under un-, weakly-, and semi-supervised settings. Extensive results show that our methods achieve the best performance over several strong baselines on the task of semi-supervised domain-generalised medical image segmentation. Code will be made publicly available upon acceptance at https://github.com/vios-s.
- 26 January 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 43 (6) , 2169-2179
Compositionally Equivariant Representation Learning
Deep learning models often need sufficient supervision (i.e. labelled data) in order to be trained effectively. By contrast, humans can swiftly learn to identify important anatomy in medical images like MRI and CT scans, with minimal guidance. This recognition capability easily generalises to new images from different medical facilities and to new tasks in different settings. This rapid and generalisable learning ability is largely due to the compositional structure of image patterns in the human brain, which are not well represented in current medical models. In this paper, we study the utilisation of compositionality in learning more interpretable and generalisable representations for medical image segmentation. Overall, we propose that the underlying generative factors that are used to generate the medical images satisfy compositional equivariance property, where each factor is compositional (e.g. corresponds to human anatomy) and also equivariant to the task. Hence, a good representation that approximates well the ground truth factor has to be compositionally equivariant. By modelling the compositional representations with learnable von-Mises-Fisher (vMF) kernels, we explore how different design and learning biases can be used to enforce the representations to be more compositionally equivariant under un-, weakly-, and semi-supervised settings. Extensive results show that our methods achieve the best performance over several strong baselines on the task of semi-supervised domain-generalised medical image segmentation. Code will be made publicly available upon acceptance at https://github.com/vios-s.
- Open Access
Group Distributionally Robust Knowledge Distillation
Knowledge distillation enables fast and effective transfer of features learned from a bigger model to a smaller one. However, distillation objectives are susceptible to sub-population shifts, a common scenario in medical imaging analysis which refers to groups/domains of data that are underrepresented in the training set. For instance, training models on health data acquired from multiple scanners or hospitals can yield subpar performance for minority groups. In this paper, inspired by distributionally robust optimization (DRO) techniques, we address this shortcoming by proposing a group-aware distillation loss. During optimization, a set of weights is updated based on the per-group losses at a given iteration. This way, our method can dynamically focus on groups that have low performance during training. We empirically validate our method, GroupDistil on two benchmark datasets (natural images and cardiac MRIs) and show consistent improvement in terms of worst-group accuracy.
Open AccessGroup Distributionally Robust Knowledge Distillation
Knowledge distillation enables fast and effective transfer of features learned from a bigger model to a smaller one. However, distillation objectives are susceptible to sub-population shifts, a common scenario in medical imaging analysis which refers to groups/domains of data that are underrepresented in the training set. For instance, training models on health data acquired from multiple scanners or hospitals can yield subpar performance for minority groups. In this paper, inspired by distributionally robust optimization (DRO) techniques, we address this shortcoming by proposing a group-aware distillation loss. During optimization, a set of weights is updated based on the per-group losses at a given iteration. This way, our method can dynamically focus on groups that have low performance during training. We empirically validate our method, GroupDistil on two benchmark datasets (natural images and cardiac MRIs) and show consistent improvement in terms of worst-group accuracy.
- 15 October 2023
- Published by Springer Nature
- p. 234-242
Group Distributionally Robust Knowledge Distillation
- 15 October 2023
- Published by Springer Nature
- p. 234-242
Group Distributionally Robust Knowledge Distillation
- Open Access
Compositionally Equivariant Representation Learning
Deep learning models often need sufficient supervision (i.e. labelled data) in order to be trained effectively. By contrast, humans can swiftly learn to identify important anatomy in medical images like MRI and CT scans, with minimal guidance. This recognition capability easily generalises to new images from different medical facilities and to new tasks in different settings. This rapid and generalisable learning ability is largely due to the compositional structure of image patterns in the human brain, which are not well represented in current medical models. In this paper, we study the utilisation of compositionality in learning more interpretable and generalisable representations for medical image segmentation. Overall, we propose that the underlying generative factors that are used to generate the medical images satisfy compositional equivariance property, where each factor is compositional (e.g. corresponds to the structures in human anatomy) and also equivariant to the task. Hence, a good representation that approximates well the ground truth factor has to be compositionally equivariant. By modelling the compositional representations with learnable von-Mises-Fisher (vMF) kernels, we explore how different design and learning biases can be used to enforce the representations to be more compositionally equivariant under un-, weakly-, and semi-supervised settings. Extensive results show that our methods achieve the best performance over several strong baselines on the task of semi-supervised domain-generalised medical image segmentation. Code will be made publicly available upon acceptance at https://github.com/vios-s.
Open AccessCompositionally Equivariant Representation Learning
Deep learning models often need sufficient supervision (i.e. labelled data) in order to be trained effectively. By contrast, humans can swiftly learn to identify important anatomy in medical images like MRI and CT scans, with minimal guidance. This recognition capability easily generalises to new images from different medical facilities and to new tasks in different settings. This rapid and generalisable learning ability is largely due to the compositional structure of image patterns in the human brain, which are not well represented in current medical models. In this paper, we study the utilisation of compositionality in learning more interpretable and generalisable representations for medical image segmentation. Overall, we propose that the underlying generative factors that are used to generate the medical images satisfy compositional equivariance property, where each factor is compositional (e.g. corresponds to the structures in human anatomy) and also equivariant to the task. Hence, a good representation that approximates well the ground truth factor has to be compositionally equivariant. By modelling the compositional representations with learnable von-Mises-Fisher (vMF) kernels, we explore how different design and learning biases can be used to enforce the representations to be more compositionally equivariant under un-, weakly-, and semi-supervised settings. Extensive results show that our methods achieve the best performance over several strong baselines on the task of semi-supervised domain-generalised medical image segmentation. Code will be made publicly available upon acceptance at https://github.com/vios-s.
- 1 February 2023
- Published by Springer Nature
- p. 3-11
Applying Disentanglement in the Medical Domain: An Introduction for the MAD Workshop
- 1 February 2023
- Published by Springer Nature
- p. 3-11
Applying Disentanglement in the Medical Domain: An Introduction for the MAD Workshop
- 1 February 2023
- Published by Springer Nature
- p. 15-21
HSIC-InfoGAN: Learning Unsupervised Disentangled Representations by Maximising Approximated Mutual Information
- 1 February 2023
- Published by Springer Nature
- p. 15-21
HSIC-InfoGAN: Learning Unsupervised Disentangled Representations by Maximising Approximated Mutual Information
- 17 September 2022
- Published by Springer Nature
- p. 704-714
vMFNet: Compositionality Meets Domain-Generalised Segmentation
- 17 September 2022
- Published by Springer Nature
- p. 704-714
vMFNet: Compositionality Meets Domain-Generalised Segmentation
- Top Cited PapersOpen Access
- 17 June 2022
- Published by Elsevier in Medical Image Analysis
- Vol. 80, 102516
Learning disentangled representations in the imaging domain
- 17 June 2022
- Published by Elsevier in Medical Image Analysis
- Vol. 80, 102516
Top Cited PapersOpen AccessLearning disentangled representations in the imaging domain