Two systems, one question: how far can AI own a decision before a human needs to step in?
One answers it in a hospital. The other answers it in an insurance claim. Same underlying bet — that confidence, quantified honestly, is what makes autonomy trustworthy.
A graph neural network that predicts five molecular biomarkers for pediatric brain tumors directly from routine H&E tissue slides — no additional lab work required. Validated on 182 real, IRB-approved cases from Dell Medical School.
The problem. Standard multiple-instance-learning treats each tissue patch as independent — discarding the spatial topology a pathologist actually relies on. Molecular testing gives ground truth, but it's frequently missing outright: in this cohort, GFAP was available for only 54% of cases and H3K27M for just 28%.
The approach. NeuroGraph treats each slide as a Delaunay-triangulated graph over UNI2-h patch embeddings (400–6,000 nodes per slide), reasoning over it with 3× EdgeConv blocks — computing explicit feature differences between neighboring patches, so the model reads tissue gradients and infiltrative edges, not just local averages. A custom masked BCE loss lets it learn from whichever biomarkers are actually annotated per case, instead of discarding partially-labeled slides.
Interpretability. Predictions alone aren't useful without a reason to trust them. GNNExplainer produces biomarker-specific tissue heatmaps, feeding a clinical dashboard qualitatively validated by the collaborating neuropathologist — for INI1, the model's highlighted regions correctly localized AT/RT rhabdoid morphology. Click a biomarker below to see its heatmap on a real case:
Where it's going next. I'm continuing this as an active UT Austin lab member — extending NeuroGraph into a multimodal framework fusing WSIs, tumor coordinates, and patient demographics via a joint embedding layer, with a multi-task objective decoding molecular biomarkers, genomic signatures, DNA methylation profiles, and histologic tumor type simultaneously. I've also proposed a biomedical-LLM branch via cross-attention fusion, so the model can generate a structured clinical explanation, not just a prediction.
