Our Projects - Neuro-Reasoning Lab

Brain MRI analysis

Mon, 01 Jan 0001 00:00:00 +0000

Brain MRI analysis plays a central role in neuro-oncology, providing multi-sequence (e.g., T1, T2, FLAIR, T1ce) information for detecting and segmenting brain tumours and guiding treatment planning. However, real-world clinical workflows often face missing or incrementally acquired modalities, which can degrade segmentation performance and make retraining impractical. In our recent work (MICCAI 2025), we propose a domain-incremental framework that uses a replay buffer, a cross-patient hypergraph segmentation network, and a Tversky-aware contrastive loss to learn from newly available MRI modalities without forgetting previously learned ones and to better handle inter-modality imbalance.

Class Imbalanced Learning

Mon, 01 Jan 0001 00:00:00 +0000

Class-imbalanced and long-tailed learning focus on scenarios where some classes have many samples while others have very few. Standard deep networks tend to overfit majority classes and under-represent rare but clinically important categories, leading to biased decision boundaries and poor minority-class performance. Our contributions span both medical and natural image domains. In medical imaging, we propose imbalance-aware methods such as a deep feature graph attention network for histopathology (ISBI 2022), which models patch–patch relationships to better capture minority patterns, and adaptive unified contrastive learning with graph-based feature aggregation (ESWA), which reweights contrastive objectives to reduce majority-class bias. In the natural image domain, we develop a decoupled optimisation framework (AAAI 2024) for long-tailed recognition that separates and optimises different parameter groups with tailored objectives, significantly improving performance across head, medium, and tail classes.

Computational Machine Ethics

Mon, 01 Jan 0001 00:00:00 +0000

#Required

Computational Machine Ethics aims to embed ethical decision-making into machines (e.g., robots). This presents many challenges due to the abstract nature of ethics and the endless possibilities of ethical situations one may encounter. Being “ethical” in a situation can be dependant on context and different individuals and groups may define what is “right” and “wrong” differently depending on factors such as their culture and background. Additionally, ethical dilemmas where different rules may come into conflict are, in themselves, difficult problems with no one correct answer.

Computer Vision Applications

Mon, 01 Jan 0001 00:00:00 +0000

These papers represent research contributions across agricultural image analysis, medical and industrial computer vision, and interpretable deep learning. My work includes building large-scale datasets (e.g., GrainSpace, MANTA), developing image quality inspection systems for cereal grains, advancing anomaly detection and multimodal fusion methods, and contributing to interpretable AI research. These publications span top venues such as CVPR, IEEE TII, TGRS, RA-L, Scientific Data, and ECAI.

Generative Models

Mon, 01 Jan 0001 00:00:00 +0000

These papers extend my research into generative models, vision-language models, and anomaly detection. The work includes human image generation with pose priors (AAAI 2025), counterfactual-aware procedural planning (ACM MM 2025), and visual-language models for anomaly detection (ACM MM 2025). Together, they reflect my broader contributions to advanced deep learning methods in computer vision and multimodal AI.

Histopathology Stain Normalisation

Mon, 01 Jan 0001 00:00:00 +0000

Stain normalization aims to reduce colour and intensity variations in H&E-stained histopathology images that arise from differences in staining protocols, reagents, and scanners. These variations can seriously degrade the robustness and generalisability of downstream algorithms for classification, segmentation, and detection, especially in multi-centre settings. To address this challenge, we have proposed several deep learning–based stain normalization methods. Our earlier work (ISBI 2021) showed that using hematoxylin components as input can effectively enhance texture feature generation, leading to higher-quality stain-normalised outputs. Building on this, we developed a semi-supervised adversarial framework for stain normalization in histopathology images (MICCAI 2021) and later introduced the Colour Adaptive Generative Network (CAGAN) (MedIA 2022).

Machine Ethics

Mon, 01 Jan 0001 00:00:00 +0000

#Required

The study of machine ethics concerns machines that behave appropriately. The complexity of autonomous agents makes it difficult to specify every possible condition in real-world deployment. Therefore, machine ethics research—especially in formal methods—attempts to mitigate any unethical behaviour or intervene if such agents are about to do something unethical.

Ethics is used because it reflects how humans are expected to behave in society. In this case, we take a general approach and consider different kinds of ethics proposed by ethicists and philosophers. Without a doubt, there are many rules and guidelines implemented within local groups. In specific environments, these rules and guidelines can then become their own form of “ethics”, which every interacting agent (robot or human) is required to follow.

Medical Image Analysis

Mon, 01 Jan 0001 00:00:00 +0000

#Required My research focuses on pathology whole-slide images (WSIs) for cancer prognosis and related medical image analysis. The goal is to extract reliable and interpretable biomarkers from gigapixel pathology images to better predict cancer progression and patient outcomes.

Model Fairness

Mon, 01 Jan 0001 00:00:00 +0000

Ensuring group fairness in deep learning models is crucial in sensitive domains like medical diagnosis, yet studies show they often exhibit demographic performance gaps that undermine trust. Existing debiasing methods typically require sensitive-attribute labels, introduce costly architectural changes, or improve fairness at the expense of utility.

We introduce RLU, a medical imaging-oriented bias mitigation framework, which mitigates bias without requiring sensitive‑attribute knowledge or intrusive model modifications.

Model Fairness

Mon, 01 Jan 0001 00:00:00 +0000

We introduce RLU, a medical imaging-oriented bias mitigation framework, which mitigates bias without requiring sensitive‑attribute knowledge or intrusive model modifications.

Nuclei Segmentation

Mon, 01 Jan 0001 00:00:00 +0000

Nuclei segmentation aims to accurately identify individual cell nuclei in histopathology images, a critical step for quantitative pathology and disease diagnosis. However, challenges such as overlapping, clustered, or ambiguous nuclei make this task highly complex.

Our group focuses on advancing this field through two key research directions:

Weakly Supervised & Noise-Robust Learning – reducing reliance on costly pixel-wise annotations.
Domain Generalisation – developing models that remain robust across unseen staining styles, scanners, and clinical settings.

Rip Current Data augmentation

Mon, 01 Jan 0001 00:00:00 +0000

Rip currents are a common beach hazard, and their visual characteristics are inconsistent and highly variable. Automated detection approaches are limited by lacking large scale annotated datasets. We use generative adversarial networks to synthesize rip currents based on input object detection labels, enriching the dataset to improve detector performance without requiring re-annotation of the generated data.

Speaker-Adaptive TTS

Mon, 01 Jan 0001 00:00:00 +0000

Speaker-Adaptive Text-to-Speech (TTS) aims to synthesize natural-sounding speech that accurately mimics the identity, timbre, and prosody of a specific target speaker, often requiring only a few seconds of reference audio (Zero-Shot scenario).

My research expands the frontiers of this field by addressing critical challenges in generalization, cross-lingual adaptation, and data scarcity.

Image Enhancement

Mon, 01 Jan 0001 00:00:00 +0000

Research Overview – Image Enhancement

Image enhancement seeks to restore visual clarity under challenging conditions such as low light, noise, blur, or complex real-world degradations. It plays a crucial role in improving downstream perception tasks and visual understanding.

Our group advances this field through two core directions:

Unsupervised Learning Methods – enhancing images without paired ground truth, enabling scalable real-world deployment.
Multi-Degradation Handling – building models that can adapt to diverse degradations (e.g., low-light, noise, haze) within a unified framework.

Domain adaptation

Mon, 01 Jan 0001 00:00:00 +0000

#Required My topics are on domain adaptation on medical image segmentation tasks. Both published methods are based on Mean-teacher model and the paradigm of self-training and data augmentation.