TMO-Net
An explainable pretrained multi-omics model for transferable oncology prediction and pathway-level interpretation.
- Transferable multi-task learning
- Pathway-level interpretation
- Precision oncology workflows
Research at FengGao Lab
AI for the full continuum of colorectal cancer care
We use colorectal cancer as a flagship clinical setting to build explainable, multimodal, and workflow-aware AI. The program connects disease biology, medical imaging, treatment response, and clinical reasoning while keeping every claim tied to the evidence that supports it.
Research map
A disease-first research program
Rather than treating models as isolated outputs, we organize the lab's work around connected capabilities that move from disease biology to clinically grounded prediction and intervention.
Program 01
Explainable Omics & Disease Mechanisms
We use biological knowledge and interpretable learning to identify reproducible disease states, molecular subtypes, and mechanisms that can support patient-level stratification.
Selected studies
- TMO-NetGenome Biology / 2024
ContributionPretrained transferable multi-omics representations while tracing predictions to biological pathways.
Why it mattersMakes cross-task oncology prediction more reusable without treating interpretation as an afterthought.
- MotifCC immune classificationCancer Letters / 2026
ContributionIdentified reproducible immune motifs for classifying colorectal cancer microenvironments.
Why it mattersConnects interpretable tumour states with clinically relevant immune heterogeneity.
DeepCC
A pathway-informed deep learning framework for robust cancer molecular subtype classification.
- Pathway-informed ANN
- Single-sample prediction
- Robust to missing data
Program 02
Multimodal Foundation Models
We learn transferable representations across radiology, pathology, and molecular data so that clinically useful signals remain available when labels or modalities are limited.
Selected studies
- CRCFoundAdvanced Science / 2025
ContributionPretrained a 3D CT foundation model on 5,137 unlabeled scans and transferred it across eight clinical tasks.
Why it mattersDemonstrates how routine imaging can support diagnosis, staging, molecular prediction, and prognosis when labels are limited.
- BrimAdvanced Science / 2025
ContributionLearned an interpretable bridge between pathology and molecular profiles across 12 cancer types.
Why it mattersKeeps multimodal prognosis useful when real-world patients have incomplete molecular data.
CRCFound
A self-supervised CT foundation model for diagnosis, staging, molecular prediction, and prognosis in colorectal cancer.
- 5137 unlabeled CT scans
- 8 downstream tasks
- Strong transfer under limited labels
Brim
An interpretable bridged multimodal fusion model that links histology and molecular data for pan-cancer prognosis.
- 12 cancer types
- Robust to missing modalities
- Histology-to-genomics bridging
Program 03
Clinical Prediction & Longitudinal Imaging
We design patient-level systems that remain stable across cohorts and platforms, with particular emphasis on treatment response, recurrence, and deployable single-sample prediction.
Selected studies
- PIANOSNature Communications / 2025
ContributionBuilt a normalization-free single-sample classifier validated across 24 cohorts and 5,439 patients.
Why it mattersReduces platform and cohort dependence in transcriptomic risk stratification for individual patients.
- 3D RP-NetNature Communications / 2021
ContributionJointly modeled tumour segmentation and response from longitudinal rectal MRI.
Why it mattersUses change during treatment, rather than a single scan, to support response assessment.
3D RP-Net
A longitudinal multi-task MRI model for treatment response prediction in rectal cancer.
- 2568 MRI scans
- AUC 0.95 / 0.92
- Segmentation plus response prediction
PIANOS
A normalization-free single-sample classifier for robust colorectal cancer risk stratification across platforms.
- 24 cohorts / 5439 patients
- Single-sample and cross-platform
- Outperformed 105 published models
Program 04
Data-efficient Medical Vision & Computational Surgery
We build efficient visual models for volumetric imaging and operative scenes, reducing annotation burden while preserving anatomical and temporal context.
Selected studies
- SegMamba-V2IEEE TMI / 2025
ContributionDeveloped a tri-orientated Mamba backbone for long-range 3D anatomical modeling across large datasets.
Why it mattersProvides a reusable and computationally efficient perception layer for volumetric medical imaging.
- SOUSAMedical Image Analysis / 2022
ContributionCombined sparse scribble labels and unlabeled images in a unified teacher-student segmentation framework.
Why it mattersLowers the expert annotation burden that limits clinical imaging model development.
SegMamba-V2
A general 3D medical image segmentation backbone for long-range volumetric modeling across organs and modalities.
- CRC-2000 benchmark
- Tri-orientated spatial Mamba
- 4 large-scale datasets
SOUSA
A unified weakly and semi-supervised segmentation framework that learns from sparse labels and unlabeled medical images.
- Teacher-student consistency
- Scribble supervision
- Multi-angle reconstruction loss
Clinical data engine
One disease-focused platform across the full care continuum
More than 50,000 structured CRC records connect routine CT and MR, digitized pathology, selected endoscopy and ultrasound, molecular profiling, treatment, and follow-up. ICGC-ARGO provides the molecular reference layer for a platform that keeps growing through routine clinical care.
- 10,000+
- CRC surgeries expected in 2026
- 10,000+
- patients with digitized pathology
- ~12,000
- pathology slides digitized each month
Emerging program
Medical cognitive agents for coordinated clinical reasoning
We are extending validated multimodal capabilities into a modular architecture that can coordinate evidence retrieval, perception, boundary checks, and workflow execution. This is an active research and prototyping program, not a deployed autonomous clinical system.
Decision orchestration
Break down clinical questions and coordinate specialist modules.
Multimodal perception
Connect structured records with imaging, pathology, and omics.
Evidence and safety
Retrieve guidance, grade evidence, and surface uncertainty.
Workflow execution
Support clinicians, researchers, and longitudinal follow-up.
Beyond CRC & open science
Evidence of generalization
Colorectal cancer is our flagship clinical setting, while the underlying methods are tested across cancers, modalities, and real clinical workflows.
Clinical Gastroenterology and Hepatology / 2026
INTELCAPE
A two-center capsule-endoscopy system for Crohn's disease diagnosis and reader assistance.
Nucleic Acids Research / 2026
PCsRNAdb
An open pan-cancer resource covering 190 datasets and 11,114 small-RNA samples.
Genomics, Proteomics & Bioinformatics / 2025
Biomedical multimodal data fusion
A methods map for multimodal representation, missing data, interpretation, and clinical translation.