We study how to understand individual patients, explain disease differences and help clinicians act on evidence. Colorectal cancer is our flagship setting, connecting multimodal modelling, biological discovery and the evaluation of AI in medical work.
Multimodal and longitudinal patient representation
How can imaging, pathology and molecular measurements describe an individual patient when observations are incomplete or change during treatment? We study reusable representations across modalities and tasks, alongside models that compare a patient's state over time.
Task-specific labels are scarce in clinical imaging. CRCFound learns from 5,137 unlabelled CT scans and adapts the resulting representation to eight tasks, including staging, molecular prediction and prognosis.
Study evidenceExternal N-stage AUC of 0.827, compared with 0.563 for training from scratch.
Patients often have incomplete molecular measurements. Brim learns an interpretable bridge between pathology and molecular profiles to support prognosis modelling when modalities are missing.
Study evidenceStudied across 12 cancer types, with independent validation in 445 patients with colorectal cancer.
Can a representation learned from multiple omics datasets transfer to new tasks? TMO-Net pretrains on four omics modalities from 8,174 patients across 32 cancer types, then adapts to classification, prognosis and drug-response modelling.
Study evidenceSurvival prediction evaluated in 12 cancer types, with external validation in four CPTAC cancer cohorts.
How much do MRI changes during treatment reveal about response? 3D RP-Net and the subsequent DeepRP-RC study jointly model tumour segmentation and response using paired images acquired before and after neoadjuvant treatment.
Study evidence3D RP-Net achieved an external AUC of 0.92; the later DeepRP-RC study evaluated 1,201 patients across four hospitals.
Long-range modelling for 3D medical image segmentation.
Learning image segmentation from sparse annotations and unlabelled data.
A review of representation, missing data and interpretation in biomedical data fusion.
Our next questions concern clinically meaningful patient embeddings, missing observations and updates over time. We aim to connect these representations to general-purpose models through retrieval and specialist prediction tools, testing their contribution to defined clinical tasks.
Molecular subtypes, mechanisms and disease dynamics
Why do patients with the same diagnosis have different tumour biology, immune environments and outcomes? We combine molecular classification with pathway analysis and experimental studies to identify disease states, investigate their mechanisms and develop reproducible risk stratification.
3
immune subtypes
From multi-omics classification to experimental investigation
MotifCC integrates gene expression, mutation and methylation data to identify CRC immune subtypes. It connects a WNT-active, immune-cold subtype to experiments testing WNT inhibition as a candidate approach to improve antitumour immunity.
Study evidenceThree subtypes identified in 1,604 CRC cases, with WNT-related intervention investigated in cell, patient-derived organoid and animal experiments.
Molecular classifications can change with the expression platform or missing genes. DeepCC converts expression into biological pathway profiles to classify established cancer subtypes across platforms and in individual samples.
Study evidenceEvaluated on 14 CRC datasets spanning 3,578 tumours and six measurement platforms.
Clinical risk assessment needs to work for an individual sample across platforms. PIANOS uses within-sample pathway rankings to stratify CRC prognosis and examine biological and treatment-response differences between risk groups.
Study evidenceIndependent validation in 15 cohorts involving 3,666 patients.
Proteogenomic discovery of non-canonical antigens, with functional tests of T-cell activation and tumour suppression.
We aim to organise findings in knowledge graphs that retain their evidence and clinical context. Modelling disease trajectories under different interventions is a longer-term goal, requiring longitudinal data, biological constraints and evaluation of future predictions.
Clinical AI and human–AI collaboration
How can AI help clinicians complete a specific task, and how should that help be measured? Our studies examine diagnostic assistance and surgical visual perception, from reader accuracy and interpretation time to the recognition of anatomical landmarks.
67.9 → 22.5
minutes per reading
Mean clinician reading time without → with AI assistance
Capsule endoscopy requires clinicians to review long video recordings. INTELCAPE combines small-intestine localisation, lesion detection and video-level Crohn's disease diagnosis, and evaluates how AI assistance changes clinicians' accuracy and reading time.
Study evidenceIn the reader evaluation, AI assistance improved accuracy from 76.7% to 94.8% and reduced mean reading time from 67.9 to 22.5 minutes.
1,715
annotated laparoscopic images
LTS dataset from 145 patients with colorectal cancer
Recognising the White Line of Toldt in complex laparoscopic images is a step toward visual assistance in colorectal surgery. FSA-Net uses fractal-guided texture and anatomical features to study this segmentation task.
Study evidenceOn the LTS test set, Dice was 0.557 versus 0.533 for Swin-UMamba, in an offline image-segmentation evaluation.
Clinical harnesses and medical agents are the next research questions in this direction. We aim to study how patient context, specialist models, evidence and human review can be organised within a clinical task, evaluating omissions, task quality, clinician time and handover before studying effects on care and patient outcomes. The same tasks can test whether upgrades to general-purpose models improve the result.
Clinical cohorts, molecular data and shared analytical resources support these studies. Our hospital-based CRC resources include routine imaging, digitised pathology, selected endoscopy and ultrasound, molecular profiling and longitudinal follow-up; individual studies use defined cohorts and modality subsets.
A disease-focused programme connecting genomic characterisation, clinical information and follow-up for translational colorectal cancer research.
A database bringing together 190 datasets and 11,114 samples across 19 cancer types, with expression, survival and regulatory analyses.
One Person Lab and our specialist tools support evidence review, analysis, research coordination and medical communication. Explore their capabilities and releases in AI Systems & Open Software.
