Research
My work has moved through four areas, in this order. Each one handed its questions to the next, and the projects I am running now are at the end.
Signal, image and energy
2006-2015 · MUET Jamshoro, then IST Islamabad
Hyperspectral document imaging, digital signal processing and power-grid communications. Engineering fundamentals, and the first time an image was something to be measured rather than looked at.
Earlier work
- Hyperspectral document image processing, published in Pattern Recognition
- Communication technologies for smart grids
- Deep learning for sensor modelling in precision irrigation
- Imaging spectroscopy and its state-of-the-art applications
Computational oncology
2018-2021 · PhD, City University of Hong Kong
Three and a half years simulating how a tumour binds, and stops binding, the drug meant to kill it. A question with a shape, a dataset and a deadline.
Drug resistance in lung cancer
Why do lung cancers stop responding to the drugs that should kill them?
- EGFR L858R and T790M mutation dynamics
- Correlated motions and community network analysis in receptor domains
- Personalised drug response prediction from tumour data
- Protein-drug interaction visualisation
From molecules to patients
2021-2024 · Qatar Foundation, MD Anderson, FAST-NUCES, UCF CRCV
Clinical cohorts, imaging physics and the first foundation-model work. Where it became clear that the useful questions rarely arrive with labels attached.
Cancer data science and precision oncology
Which patients will benefit from a given treatment?
- Predicting benefit from immune checkpoint inhibitors in non-small-cell lung cancer
- Identification of prognostic and predictive cancer biomarkers
- Recurrence-free survival modelling in head and neck tumours
- Long non-coding RNA and RNA-binding protein interactions in cancer
Medical imaging and computer vision
How do we make imaging models accurate enough to be trusted clinically?
- Weighted fusion transformers for dual PET/CT tumour segmentation
- Semi-supervised knee cartilage segmentation from MRI
- Medical object detection, including a systematic review of the YOLO family
- Explainable diagnosis, including SHAP-based skin cancer classification
- Image restoration and vessel segmentation
Responsible and generative AI
Who is accountable when a generative system causes harm?
- Multimodal disinformation evaluation with regulatory alignment
- Responsible generative AI across data, models, users, and regulation
- Generation and detection of synthetic media, and evaluating it against regulation
- Explainability and auditability in clinical machine learning
Pretraining over per-task training
2024-now · Pépin Laboratory, MGH and Harvard Medical School
One pretrained model adapted to many tasks, rather than a model per task. Biology and medicine are the domains where that shift is least finished.
Foundation models for biology and medicine
What happens when we pretrain on biological data instead of training one model per task?
- Whole-embryo foundation models for understanding development
- Biomedical imaging foundation models and their evaluation
- Multimodal learning across imaging, genomics, and clinical text
- Women's health and whole-body health applications
Large language model applications
Where do large language models help, and where do they fail?
- Surveying LLM applications, challenges, limitations, and practical usage
- Efficient retrieval-augmented generation with intent-aware retrieval
- Multimodal LLMs applied to agriculture and the life sciences
- Brain-inspired and cognitive architectures for general intelligence
Current
Current projects
Modelling how development unfolds at cellular scale, rather than segmenting one structure at a time.
In developmentCompanion work to a Nature Biomedical Engineering Perspective on foundation models in imaging.
In preparation
A companion repository for a Perspective on foundation models in biomedical imaging is listed as in preparation. It is given no title, because none is public.