ViGNet Reported 82.55% Discrimination Predicting Immunotherapy Response in NSCLC
A Hong Kong Polytechnic University team reports that a network fusing whole slide pathology images with gene expression data reached 82.55% discrimination in classifying which non-small cell lung cancer patients respond to immunotherapy. The same announcement introduces a separate digital twin platform now entering commercialisation. The two arrived together and read as one thing. Only one has a published number attached.
Multimodal deep learning study. ViGNet, the Visual-Global Relation Fusion Network, pairs a multi-scale visual encoder operating on histopathology whole slide images with a gene-driven encoder taking gene expression profiles and cancer-type text, then fuses the two to classify immunotherapy response in non-small cell lung cancer. Image preprocessing included colour normalisation and tissue masking. Evaluated against baseline approaches on response classification. Kong L, Wu S, Cai C and colleagues, in a group led by Lawrence Chan of the PolyU Department of Health Technology and Informatics. Published in Medical Image Analysis, 2026.
- ViGNet reported 82.55% discrimination performance on immunotherapy response classification, ahead of baseline comparators. The institutional materials use that phrase without naming the metric.
- The gain is attributed to fusing modalities rather than to either encoder alone. Most deployed tools read a single CT scan, a single genomic report or static clinical data, and the case made for ViGNet is that pathology and expression carry response signal neither holds by itself.
- The separately announced Virtual Patient Simulation System is a digital twin combining a clinician platform with a patient mobile application, moving records between sites through an encrypted Deep Feature QR code. No performance figure is published for this system.
- The platform has been conditionally accepted into the Hong Kong Science and Technology Park incubation programme and was shown at Mobile World Congress 2026, where it was shortlisted for a Global Mobile Awards category.
The number belongs to ViGNet, a retrospective classifier. It does not belong to the digital twin platform, which is the thing being commercialised and which carries no published performance figure at all.
The problem ViGNet is aimed at is real and expensive. PD-L1 immunohistochemistry remains the working selection biomarker for checkpoint inhibitors in NSCLC despite being a weak discriminator, and tumour mutational burden adds less than the enthusiasm around it suggested. A model that pulls signal out of haematoxylin and eosin slides already cut for every patient, at close to zero marginal collection cost, is attacking the right target.
What is missing is everything needed to act on it. Anyone tracking patient selection in checkpoint inhibitor programmes should establish three things before treating this as a competitive development: what the 82.55% actually measures, how many patients it was measured on, and whether any of them sat outside the training institution.