Experience

Working with Prof. Jianxing He's thoracic surgery team on a visual foundation model programme for lung-cancer diagnosis from histopathology and computed tomography.

• Investigated region-of-interest activation failure in weakly supervised segmentation of lung-cancer histopathology under cross-domain transfer with varying staining protocols, and designed a diffusion-based source-free domain adaptation pipeline that restored high-accuracy segmentation without source data or additional annotation.
• Developed training-free multimodal LLM frameworks for CT and whole-slide-image analysis, adapting long-context models to high-dimensional medical imaging without task-specific fine-tuning.
• Output: one oral paper at ICIC 2025 and one oral paper at IEEE ICME 2026.

• Investigated step-level reward modelling to improve the reliability of LLM reasoning chains in biomedical question answering, supervising each reasoning step rather than only the final answer.
• Designed and evaluated a process reward model that transfers across tasks, with emphasis on robustness under distribution shift and in out-of-distribution clinical scenarios.
• Output: one paper accepted at ACM Multimedia 2026.

• Represented high-dimensional biomedical data, including functional MRI and whole-slide images, as graphs so that language models can reason over inputs carrying far more variables than a context window admits.
• Developed a training-free graph in-context learning method for medical multimodal prediction and benchmarked it against fine-tuned baselines.

• Designed a convolutional diffusion-denoising pipeline for MRI and X-ray segmentation and spinal disease classification, improving performance on the BUU and lumbar-spine MRI datasets.
• Applied representation learning and NLP to hospital chief-complaint records, evaluating clustering and vectorisation methods including TF-IDF, and developed lightweight algorithms for edge computing.

• Improved the visual quality of AI-generated video by developing and applying noise-detection algorithms.
• Built label preprocessing for the performance evaluation benchmark used to assess generated video quality.

• Contributed to development and testing of .NET, MVC and ERP modules across the full stack.

• Carried out the data-processing component of two anomaly-detection studies, covering cleaning, preprocessing and preparation of large-scale log and time-series datasets for model training and evaluation.
• Produced the results visualisation for both studies, including the comparative and ablation figures reported in the published analyses.
• Output: co-authored two journal papers (Applied Soft Computing, IJMLC).