Our deep learning model quantifies tertiary lymphoid structures (TLS) on standard H&E slides and has been clinically validated as a predictor of immunotherapy response in a phase III trial. No additional staining is needed, and every result is explained.
The platform detects and quantifies every TLS, then compares TLS features with your own cohort data (treatment response, survival, PD-L1 or any custom analysis) to find new biomarkers. It starts from the routine H&E, with no extra sections, staining or manual annotation. The video uses mRECIST response as one example.
The tissue and cohort data in this video are illustrative. The results in the sections below come from real annotated images.
You share routine H&E slides; we return a quantitative TLS profile for every case.
Each result is traceable to the tissue, so pathologists can verify it.
A novel deep learning model we developed for tertiary lymphoid structures on routine H&E.
Move the divider to compare the H&E with the AI map. The dashed line is the pathologist's annotation; the panel compares the AI's measurements with the pathologist's. Each map was produced by a model that had not seen that image during training.
We compared the AI's TLS score on H&E with a multi-marker IHC assessment of TLS on slides aligned to the same tissue. The model never saw the IHC. The same check works with any IHC marker or panel, including your own.
Left to right: H&E · AI TLS score · IHC staining (orange) with the AI's TLS regions outlined
Left to right: H&E · AI TLS score (high scores outlined) · two IHC markers of the multi-marker assessment
Left to right: H&E · AI TLS score (high scores outlined) · two IHC markers of the multi-marker assessment
The model's decisions align with features a pathologist can verify: dense, small nuclei, and a pattern that holds in every case. Hover over or tap the charts for values.
Median inside annotated TLS divided by the median in the surrounding tissue. Values right of 1 are higher in TLS.
Median nuclear density inside TLS and in the surrounding tissue of each image, sorted by the difference.
Tell us where predicting immunotherapy response is hardest in your work, and whether the TLS analysis on this page would help. Your answers guide what we build through the ICURe Discovery programme; only your role is required.
Work in pharma, biotech, a CRO or a digital pathology company? Take the industry survey instead (about 8 minutes) →
We bring together AI research, clinical histopathology and technology transfer to take TLS analysis from the lab to the people who need it.
Senior Research Data Scientist
Clinical Researcher, Histopathologist
Technology Transfer Officer
We validated our TLS model on routine H&E slides from nivolumab-treated patients in the phase III CONFIRM trial (NCT03063450). Patients with TLS-high tumours showed significantly greater tumour shrinkage on mRECIST than those with TLS-low tumours, and TLS-high status was independently associated with longer overall and progression-free survival.
Barmpoutis P, Moazzen H, Poile C, Ibrahim TS, Ewings SM, Griffiths G, Fennell DA, Ng Kee Kwong F. Presented 26 October 2026.
Read the abstract on the ESMO websiteOne of the first approaches to automated identification and quantification of TLS on routine H&E slides.
Barmpoutis P, Di Capite M, Kayhanian H, Waddingham W, Alexander DC, Jansen M, Ng Kee Kwong F. (2021) PLoS ONE 16(9): e0256907. doi.org/10.1371/journal.pone.0256907
Read the article in PLoS ONEWe are partnering with early adopters: pathology groups, translational researchers and industry teams who want TLS measurements from their own H&E slides.
