Matthew Vaishnav
Computational pathology research engineering
Research-only machine learning for whole-slide histopathology, scanner/site robustness, paired-acquisition representation learning, and reproducible ML evaluation
Work
I build research-only computational pathology experiments around whole-slide modeling, pathology foundation-model features, scanner/acquisition robustness, simulated federated pathology stress tests, and reproducible evaluation infrastructure. My current strongest research line is Paired-Acquisition Neural Factorization: testing whether paired acquisitions of the same tissue can reduce scanner/acquisition signal in pathology embeddings while preserving tissue identity.
Current Research
- 1. Paired-Acquisition Neural Factorization
Scanner/acquisition signal vs tissue identity preservation. SCORPION, external canine SCC, cross-backbone, and baseline controls.
- 2. External multi-scanner validation
Canine SCC validation package with independent scanner conditions.
- 3. Pair-repeat allocation
Unique pair diversity vs repeated exposure tradeoffs.
- 4. Mechanism audits
Acquisition-branch audit: branch separation is measurable.
Pair-structure boundary test: biological pairing structure matters. - 5. Earlier research
PANDA / PCam / MIL experiments. Simulated federated pathology stress tests.
Research Metrics
- SCORPION scanner probe
- 0.7825 → 0.3989
- Reduction in scanner recoverability
- Canine SCC scanner probe
- 0.7529 → 0.3614
- External validation reduction
- PANDA readable features
- 10,611
- Verified slide feature vectors
- PCam validation AUC
- 95.37%
- Full validation set
- PANDA gated AttentionMIL
- QWK 0.8100
- Slide-level baseline
- PANDA tuned TransnnMIL
- QWK 0.8155 / 0.8225
- Repeated-seed results
Claim Boundary
Research-only. Not clinically validated. Not diagnostic software. Not intended for clinical deployment or patient-care use. Results are from controlled experiments and simulated federations over pathology-derived feature vectors, not real hospital deployments. All findings are bounded by the specific datasets, backbones, and experimental conditions described in each report.
Bio
I ♥
Computational pathology, whole-slide image modeling, scanner/site robustness, reproducible experiments, security engineering, failure analysis, home labs, clean interfaces, coffee, music, art, and my dog.
Inspired by Takuya Matsuyama's homepage