Early computational pathology research notes

Where I Started

I began working with computational pathology in late 2024, starting from public benchmark datasets and open-source tooling. The goal was to learn the fundamentals of whole-slide image modeling, multiple-instance learning, and pathology ML evaluation — not to build clinical software.

The early project, which I called HistoCore, was a research learning vehicle. It went through many iterations as I learned PyTorch, WSI processing pipelines, MIL architectures (AttentionMIL, CLAM, TransMIL), and the challenges of working with gigapixel pathology images on consumer hardware.

What I Explored

The work covered several areas that remain relevant to my research today:

PCam benchmark. The PatchCamelyon dataset provided a manageable starting point for learning patch-level classification, data loading, training loops, and evaluation. Results: 95.37% validation AUC, 85.26% test accuracy (0.9394 AUC) on the full 32,768-sample test set with bootstrap confidence intervals.

PANDA prostate grading. Moving from patches to whole-slide images required learning HDF5-based feature extraction, slide-level aggregation, and multi-class ordinal regression. Using mean-pooled Phikon features and gated AttentionMIL, I achieved QWK 0.8100. TransnnMIL experiments reached QWK 0.8155–0.8225 in repeated-seed runs.

Training optimization. On an RTX 4070 laptop GPU, I learned to use torch.compile, mixed precision (AMP), channels-last memory format, and persistent DataLoader workers. GPU utilization improved from ~17% to ~85%, and PCam training time dropped from 20–40 hours to 2–3 hours. These optimization patterns carry forward into my current work.

Federated learning simulations. I built a simulation framework to study how FedAvg behaves under realistic pathology-lab conditions: uneven site sizes, label noise concentrated at dominant sites, and systematic grading bias. The key finding — that FedAvg becomes vulnerable when the largest simulated client is unreliable — was a genuine research result, though it came from simulations, not real hospital deployments.

Infrastructure testing. Property-based testing with Hypothesis and systematic evaluation pipelines taught me the difference between model evaluation and infrastructure verification. This discipline now underlies how I structure experiments and validate results.

Important Context

In earlier versions of these pages, I used language that overstated what this work was. Phrases like “AI framework for cancer detection,” “production-ready,” “clinical PACS integration,” and “HIPAA compliance” were inaccurate. I have corrected that language.

This work was and is: research-only learning and prototyping on public benchmark datasets. Not diagnostic software. Not clinically validated. Not deployed in any hospital. Not a medical device. Not used for patient care.

The value of this period was the engineering experience it built: how to structure reproducible experiments, how to validate data pipelines, how to run systematic hyperparameter studies, and how to write honest research claims with clear boundaries.

Performance Notes (Research Benchmarks)

PCam Test Set (32,768 samples, public benchmark)
- Accuracy: 85.26% (95% CI: 84.83%-85.63%)
- AUC: 0.9394 (95% CI: 0.9369-0.9418)
- F1: 0.8507 (95% CI: 0.8464-0.8543)
- Bootstrap: 1,000 resamples
- Hardware: RTX 4070 Laptop (8 GB VRAM)

Research infrastructure notes:
- GPU utilization improved from 17% to 85%
- Training time reduced from 20-40 hours to 2-3 hours
- Property-based testing for data pipeline correctness
- All results on public benchmark data only

Current Direction

My research has evolved from general-purpose pathology ML infrastructure to a focused investigation of Paired-Acquisition Neural Factorization. The question is whether paired acquisitions of the same tissue (different scanners, same block) can be used to reduce scanner/acquisition signal in pathology foundation-model embeddings while preserving tissue identity.

That work — across SCORPION, canine SCC, cross-backbone checks, baseline stress tests, pair-repeat allocation, and mechanism audits — is documented in the main research repository.

Full documentation: matthewvaishnav.github.io/computational-pathology-research