Back to Portfolio

Paired-Acquisition Neural Factorization 2025–

Overview

Research-only computational pathology experiments centered on testing whether paired acquisitions of the same tissue can reduce scanner/acquisition signal in pathology foundation-model embeddings while preserving tissue identity. The work spans SCORPION (human prostate), external canine SCC validation, cross-backbone checks, baseline stress tests, pair-repeat allocation, and mechanism-hardening audits.

The core approach, Paired-Acquisition Neural Factorization, uses paired same-region acquisitions (different scanners, same tissue block) to train a factorization that separates biological signal from acquisition/scanner signal in frozen foundation-model feature spaces. This is a research investigation, not a clinical tool.

Current Research Packages

Main indexResearch documentation and results index
Canine SCCExternal multi-scanner validation PDF
AllocationPair-repeat allocation study PDF
SCORPIONGitHub repository
GitHubMain research repository

Key Results

SCORPION DINOv2Scanner-probe accuracy reduced from 0.7825 to 0.3989 while mean paired cosine improved from 0.8476 to 0.8789.
Canine SCC DINOv2Scanner-probe accuracy reduced from 0.7529 to 0.3614 while paired cosine improved from 0.6960 to 0.7300.
Cross-backbonePhikon and ResNet50 backbones reproduce the scanner-suppression / tissue-preservation pattern.
Baseline stressLinear scanner projection and PCA do not match the scanner-suppression / tissue-preservation tradeoff.
Acquisition-branch auditAcquisition branch retains higher scanner recoverability while carrying much lower tissue-identity retrieval than biological branch.
Pair-structure boundaryTrue same-region pairs preserve tissue identity best. Looser/random pairings suppress scanner but do not preserve tissue identity as well.

Mechanism-Hardening Audits

Post-freeze audits strengthen the mechanism interpretation without expanding the claim surface. The acquisition-branch audit confirms that branch separation is measurable: the acquisition branch retains higher scanner recoverability while the biological branch preserves tissue-identity retrieval. The pair-structure boundary test shows that biological pairing structure matters: true same-region pairs preserve tissue identity best, and looser or random pairings do not recover true-pair behavior even though they can still suppress scanner signal.

These are research findings under controlled experimental conditions. They support the interpretation that the factorization is doing more than simple scanner suppression, but they do not prove full factorization or establish clinical validity.

Earlier Work

PANDA10,611 readable slide-level Phikon feature vectors. Gated AttentionMIL QWK 0.8100. Tuned TransnnMIL QWK 0.8155 / 0.8225 / 0.8086.
PCam95.37% validation AUC. 85.26% test accuracy and 0.9394 test AUC on the full 32,768-sample test set.
Federated stress15-seed simulated-federation PANDA studies of FedAvg under dominant-site label corruption and systematic ordinal threshold bias. Cross-site blending improved robustness when the dominant simulated site became unreliable.
TransnnMILCustom multiple-instance learning architecture direction for WSI modeling — earlier slide-level MIL research prototype work.

Claim Boundary

Research-only. Not clinically validated. Not diagnostic software. Not intended for clinical deployment, patient care, or medical decision-making. Results are from controlled experiments on specific datasets (SCORPION, canine SCC, PANDA, PCam) with specific foundation-model backbones (DINOv2, Phikon, ResNet50). Findings use language like “supports,” “suggests,” and “across audited settings” — they do not prove, solve, or establish clinical validity. Simulated federated experiments are not real hospital deployments.