Building AI that sees what the eye misses.

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.

The problem. Real biomarker data for adult GBM is scarce — invasive to sample, and slow and expensive to test for. This project asked how much prognostic signal could be recovered from imaging alone, standing in for the biomarkers that couldn't be obtained directly.

The approach. A 3D U-Net trained on the BraTS 2020 dataset segments each scan into three tumor subregions across four MRI modalities (T1, T1CE, T2, FLAIR). Synthetic MGMT methylation and EGFR amplification values, derived from tumor and edema volume, are then fused with the imaging features to predict clinical outcomes.

Where it led. The synthetic biomarkers were never going to be a clinical substitute for the real thing — that limitation is the whole reason NeuroGraph exists. NeuroGraph picked up the same underlying question on pediatric tumors, where real biomarker labels were actually available to train and validate against.

The premise. Most enterprise workflows are human-default: an AI system assists a person, and a person makes the call. NEXUS argues for the opposite in autonomous insurance processing — a decentralized network of domain-specific agents handles claims end-to-end, with humans routed in only where the system's own confidence says they're needed.

The mechanism. Coordination runs through the Truth Score Engine, a weighted, penalty-adjusted consensus metric built on evidential deep learning principles:

T_s = [ Σ (w_i · C_i) ] · Π(1 − H_penalty)

Click to simulate a claim moving through the engine — each run generates a random agent-consensus score and routes it accordingly, exactly as NEXUS does in production.

Why this framing matters. Most "AI + human review" systems treat every decision the same way — full automation or full review. Confidence-aware routing spends human attention only where it's actually warranted; that's the difference between "AI-assisted" and genuinely AI-native operations.

Download on Google Play