����Vlog

Led the development of LyssaScan, the first AI-driven workflow for automated Rabies lyssavirus infection detection and antibody titre quantification from fluorescence microscopy images at APHA. Built and curated a large-scale cell culture imaging dataset capturing biological and technical variability across virus dilution series, staining conditions, and assay artefacts. Designed state-of-the-art deep learning architectures, achieving >99% well-level accuracy and robust generalisation under real laboratory conditions.

Extended the framework to the WOAH-recommended Fluorescent Antibody Virus Neutralisation (FAVN) assay, enabling automated antibody titre prediction directly from raw images. The best-performing model achieved complete concordance with expert operators across critical regulatory thresholds (≥0.5 IU/mL), demonstrating reliability for vaccine assessment, serological surveillance, and international pet travel compliance. The system processes thousands of images within minutes, supporting scalable and objective diagnostic workflows.

Building on this foundation, I am leading the development of multimodal foundation models integrating fluorescence microscopy with whole genome sequencing (WGS) data to enable genotype-informed virus identification and strain-level characterisation. By aligning cellular phenotypic signatures with viral genomic embeddings, the framework enables cross-modal reasoning between morphology and genotype, advancing rapid pathogen identification, enhanced epidemiological intelligence, and next-generation AI-driven biosecurity infrastructure.

Led the development of KnifeHunter, the world’s first large-scale AI system for forensic knife image retrieval, designed and deployed in collaboration with the Metropolitan Police. Framed knife catalogue matching as a fine-grained instance-level retrieval problem under real-world forensic constraints, including clutter, occlusion, uncontrolled illumination, and large distractor galleries.

Built the first dedicated KnifeHunter dataset comprising 25,843 images across 543 knife classes collected from police evidence repositories, retail catalogues, and UK Border Force seizures, supported by structured forensic metadata. Designed KHNet, a novel retrieval architecture integrating Weibull-based activation shaping, saliency-guided prototype aggregation, and bi-directional reciprocal fusion to enhance discrimination under cluttered operational conditions. The system achieves 88.1% mAP on the Medium protocol and maintains strong performance under million-scale distractor retrieval, without query expansion or re-ranking.

KnifeHunter was operationally validated during Operation Sceptre, achieving >99% rank-1 accuracy in field deployment. The system is hosted on secure infrastructure and is used by police forces across 22 UK regions, enabling scalable catalogue matching, intelligence aggregation, source attribution, and data-driven enforcement against illicit knife sales. The platform has also contributed to national policy discussions on knife crime prevention and online retail regulation.

Led the development of novel foundation modelling frameworks for cellular biology, advancing scalable representation learning for high-content microscopy, morphological profiling, and drug discovery.

Developed the Hybrid subCellular Protein Localiser (HCPL), a large-scale single-cell AI system for protein localisation within the Human Protein Atlas, integrating diverse deep architectures, wavelet-based hybrid modelling, and parametric activation pooling to robustly handle extreme cellular variability and weak labelling. The system achieved state-of-the-art performance in single-cell subcellular classification, enabling biologically meaningful protein localisation analysis at scale.

Subsequently developed TRex (Task-guided Representation Exaptation), a modular adaptation framework bridging self-supervised learning and mechanism-of-action (MoA) discovery in Cell Painting datasets. TRex refines generic foundation embeddings into biologically task-aware representations, doubling MoA classification performance and significantly improving compound recognition.

This work establishes a new paradigm for biologically grounded foundation models that move beyond generic embeddings toward task-aware, interpretable cellular representations. Applications include mechanism-of-action prediction, protein localisation modelling, Cell Painting analysis, phenotypic screening, and AI-driven drug discovery. The framework enables scalable cross-task adaptation without retraining large-scale models, supporting rapid deployment across new assays, cell lines, and therapeutic discovery pipelines.

In this project, I developed a novel system, HCPL, to localise proteins at a subcellular level. Our system processes high-resolution confocal microscopy images to segment individual cells, assesses their visual integrity, and robustly predicts protein localisation patterns. HCPL makes large-scale single-cell data annotation feasible by helping to avoid the need to hand-label individual cells. HCPL currently defines state-of-the-art in the single-cell classification of protein localisation patterns.
 

I worked on the InnovateUK project “RetinaScan: AI-enabled automated image assessment system for diabetic retinopathy screening”. Automated grading of DR has important benefits such as increasing efficiency, reproducibility, and improving patient outcomes. Our deep learning system ‘HydraNet’ achieved high sensitivity and specificity for identifying DR using retinal images from multi-ethnic populations, significantly outperforming the current best system. As part of this project our team also too part in Human Protein Cell Classification challenge 2021 organized by the Human Protein Atlas team in Sweden and Kaggle. Here we designed a novel DNN architecture capable of segmenting and classifying human protein cells, which is expected to help accelerate our understanding of how cells function and how diseases develop. Our design achieved top 2% position from 800 world-leading universities and companies. My current focus is to develop novel algorithms to detect eye diseases in ultra-widefield (UWF) fundus imaging and to further improve the performance of the protein cell classification system.

The project aim was to develop a smartphone-based intelligent "virtual journey assistant", providing end-to-end routing with proactive contextual information to the traveler, including real-time visual recognition through the smartphone's camera. During the project, our team took part in the Google Landmark Retrieval Challenge 2018. The challenge was to develop the world’s most accurate technology to automatically identify landmarks and retrieve relevant photographs from a database. Our deep learning-based system “REMAP: Multi-layer entropy-guided pooling of dense CNN features for image retrieval” won the prestigious competition, significantly outperforming well known international corporations and global university groups including Google, Layer6 AI Canada, Facebook, Naver Labs Europe, Stanford USA and MIT USA. Our winning solution was published in IEEE Transactions on Image Processing (impact factor 6.8). Furthermore, my recent image recognition system “ACTNET: end-to-end learning of feature activations and aggregation for effective instance image retrieval” won Gold Medal in Google Landmark Retrieval Challenge 2019 beating competition from 300 teams.

The project aim was to develop an advanced video classification system. As part of the project, our team participated in the Google YouTube 8M Video Understanding Challenge 2018. The task set by Google was to develop algorithms which accurately and automatically assign labels to videos using a dataset created from over seven million YouTube videos. We developed a deep-learning based AI system which can learn to understand the story behind any video and give a short verbal summary of what it is about. Our system won the Google Gold Medal from the 650 entries worldwide. The deep learning system we developed is being used by the BBC and the Metropolitan Police. The work on video classification was published in IEEE Transactions on Circuits and Systems for Video Technology (impact factor 4.06). 
 

The project was to develop large-scale visual search algorithms for the broadcast industry. My research led to the development of a novel method for deriving a compact and distinctive representation of image content called Robust Visual Descriptor (RVD). It significantly advanced the state-of-the-art and delivered world-class performance. The ����Vlog filed a US patent based on RVD representation for visual search. My work was also published in IEEE Transactions on Pattern Analysis and Machine Intelligence with an impact factor of 17.