Matthew Vaishnav
Computational Pathology Researcher & Research Engineer
I develop and study PA-NF, an end-to-end computational pathology pipeline spanning imaging, whole slides and institutions.
About
I developed Paired-Acquisition Neural Factorization (PA-NF) as an end-to-end computational pathology research pipeline. Its starting point is a measurement problem: the same tissue can produce different digital images when scanned using different instruments, and those differences can influence models further along the analysis process.
The pipeline begins with a neural model designed to distinguish tissue-oriented information from scanner-related variation. It extends to whole-slide analysis through TransnnMIL and to institutional learning through PathologyFL and FAIR-WEIGHTS-H. These are components of the same research programme, developed and tested using distinct experiments.
I build the models, training software and reproducibility systems used throughout PA-NF. My work examines not just whether the models perform well, but what their results actually establish about the tissue and its observation.
Research
SCORPION provides the primary paired-scanner experiment: 48 human H&E slides scanned on five instruments. In 175 registered model fits, the PA-NF factorization model reduced recoverable scanner identity by 31.1 percentage points against an equal-capacity neural control while meeting its same-region retrieval requirement. Independent canine SCC experiments and stronger simple scanner corrections establish additional limits.
The same PA-NF programme includes the PCam patch benchmark, TransnnMIL experiments on 10,611 PANDA slides, and PathologyFL studies of institutional weighting using simulated PANDA sites and real centre differences from CAMELYON17. Each result supports its own specific experimental conclusion.
Background
I conduct independent computational pathology research from Kitchener–Waterloo, Ontario, while studying Computer Programming at Conestoga College. My earlier work in systems engineering continues to inform how I develop and maintain research software.