Research
Paired-Acquisition Neural Factorization (PA-NF) is the name of my complete computational pathology research pipeline. The name began with a neural model for paired scanner observations, but the published work extends from image representations to whole-slide predictions and then to learning across institutions. TransnnMIL, PathologyFL and FAIR-WEIGHTS-H are parts of that same programme.
I designed these methods and the software used to test them. The studies below evaluate different stages of PA-NF; results from one stage are not treated as validation of the others.
Paired-acquisition model: SCORPION and canine SCC
I developed the paired-acquisition model to examine changes caused by scanning the same tissue on different instruments. It learns tissue-oriented and scanner-related representations, then uses both to reconstruct image features. Matching the same physical regions across scanners makes the proposed separation testable.
SCORPION provides 48 human H&E slides, 480 matched tissue regions and five scanner views per region. Across 175 registered model fits, PA-NF reduced recoverable scanner identity by 31.1 percentage points relative to an equal-capacity neural control, with same-region retrieval preserved within the predefined margin. I also tested the objective on DINOv2, Phikon and ResNet50 representations.
Independent experiments using five-scanner canine squamous cell carcinoma (SCC) tissue test the model outside SCORPION, while allocation experiments examine the effect of choosing and repeating tissue pairs. Simpler scanner corrections have matched or exceeded PA-NF on some endpoints. Its strongest established advantage is therefore the registered SCORPION neural comparison, not universal superiority or proof of independent control over acquisition.
Patch-classification foundation: PCam
For the patch-classification stage of the PA-NF programme, I trained a ResNet18-based model with a transformer encoder on PatchCamelyon (PCam). It achieved an ROC AUC of 0.9394 and 85.26% accuracy on the official test split. This is a patch-level result, distinct from whole-slide or clinical validation.
Whole-slide modeling: TransnnMIL and PANDA
TransnnMIL is the custom whole-slide architecture developed within PA-NF. It combines two approaches to selecting and aggregating information from image patches, and includes experimental spatial extensions. Its PANDA studies use 10,611 readable prostate pathology slides. The corrected architecture has been implemented, but a matched evaluation establishing its advantage over competing whole-slide methods remains pending.
Institutional learning: PathologyFL and FAIR-WEIGHTS-H
PathologyFL is the institutional-learning framework developed within PA-NF. It provides a way to study training across participating sites, while FAIR-WEIGHTS-H examines how much influence a site should have when its data become less reliable. PANDA experiments use simulated sites and test when changing institutional weights improves performance under controlled corruption.
CAMELYON17 provides a different setting with 455,954 image patches from five real centres. In one frozen-feature comparison, equal weighting of the training centres increased held-out-centre accuracy from 0.8312 to 0.9132 relative to sample-proportional weighting. This was a centralized experiment on institutional influence, not a deployed multi-hospital system.
Experimental software and reproducibility
I develop and audit the software used throughout PA-NF, including data preparation, model training, controlled comparisons and reproducible releases. The research record preserves negative results, implementation corrections and exact evidence for the claims retained in the manuscript.
Methods, manuscripts and evidence are linked above.