This project develops a generalized membrane-level cell segmentation algorithm for multiplexed immunofluorescence imaging.
Existing approaches fall into two groups: nuclear-based methods that extrapolate cell extents from nuclei (simple but shape-agnostic),
and tissue-specific models that capture boundaries yet require retraining for each tissue.
By leveraging membrane-bound marker fluorescence, we aim to delineate highly concave and deformable cells (e.g., amoeboid immune cells, neurons, migrating cells)
and enhance downstream immunofluorescence analysis.
Current cell segmentation algorithms for multiplexed immunofluorescence imaging are either generalized but shape-agnostic (nuclear-based),
or accurate but narrow in scope (tissue-specific). The lack of reliable membrane-level segmentation across tissues hinders the study of
complex morphologies, especially for immune cells, neurons, and migrating cells.
Current Work
Literature review: comprehensive survey of nuclear-based and tissue-specific methods, identifying limitations in membrane-level segmentation.
Algorithm design & implementation: developing a generalized membrane-driven approach that traces cell boundaries from fluorescence-bound markers.
Data pipeline: collection and preprocessing of multiplexed images from varied tissue sources for robust training/testing.
Training & validation: leveraging GPU clusters to train models on diverse datasets and validate with expert annotations.
Performance evaluation: benchmarking on challenging concave/flexible cell types; iterative error analysis and refinement.
Documentation & communication: maintaining detailed records and presenting progress to the research team.
Expected Outcomes
A generalizable membrane-level segmentation framework that transfers across tissues.
Improved analysis for structurally complex cells and enhanced downstream quantification.
Algorithmic contributions to computational pathology and biomedical data science.
✋ Gesture Recognition from Sensors Using Dual-path Encoding and Attention
Final Year Project | University of Birmingham | 09/2024 – 04/2025
We designed DS-CAN, a multimodal framework for MEMS-based gesture recognition.
The system employs dual convolutional encoders to preserve modality-specific dynamics (accelerometer vs. gyroscope),
a multi-head attention module for adaptive cross-modal fusion, and extends contrastive learning (NT-Xent) to the multimodal setting.
The model attains ~94% accuracy with a lightweight footprint (3.9M parameters, 10.3 ms inference), enabling real-time deployment on wearables.
📄 A full paper (PDF) is available in the details section.
Compared to camera-based methods, MEMS sensors offer low power and strong privacy, but robust multimodal fusion, discriminative representation
learning, and lightweight deployment remain challenging.
Methodology
Dual-path encoding: two convolutional encoders independently model accelerometer and gyroscope streams.
Attention-based fusion: multi-head attention reweights cross-modal features to enhance robustness under noise/variability.
Multimodal contrastive learning: NT-Xent adapted to align modalities while maximizing inter-class margins.
Results
~94% accuracy on public benchmarks (e.g., 6DMG, MGD).
Compact model (3.9M params) with 10.3 ms inference latency suitable for real-time wearables.
Stable performance across subjects and sensor noise conditions.
👵 Creating Supportive Environments to Bridge the Digital Health Literacy for Older People
Research Advisor: Prof. CHU, Samuel K.W. | 01/2024 – Present
Health LiteracyElderly EngagementIntergenerational LearningSPSS, R
Background
Older adults often encounter socioeconomic and technological barriers to adopting digital health tools. Targeted literacy programs can
improve access and outcomes, but designing sustainable, community-driven interventions remains an open problem.
Current Work
Surveyed multiple cities (600+ responses) to identify adoption factors and pain points.
Co-designed and delivered an 8-week intergenerational co-learning program with community day care centers.
Mixed-methods evaluation (questionnaires + interviews) to quantify confidence gains and usage behaviors.
Expected Outcomes
Increased digital confidence and e-health tool usage among older adults.
Actionable insights for scaling up community-based interventions.
Policy implications for inclusive digital health ecosystems.
⚙️ Physics-Informed Neural Networks with Broad Learning Systems