Building Clinical-Grade Cancer Detection AI on a Startup Budget

Client: CytoVeris — OncoSIGHT AI on the TumorMAP platform

Engagement: 18-month ML consulting partnership (2021-2022)

The Business Problem

Between 20-40% of breast conservation surgeries require a second operation because surgeons lack real-time tools to assess whether all cancer has been removed. Each re-excision costs the healthcare system thousands of dollars and puts patients through additional procedures, recovery time, and emotional distress.

CytoVeris had developed novel hardware (UV-excited autofluorescence imaging) but needed ML expertise to transform multispectral images into actionable surgical guidance. With a lean technical team and limited runway, they needed to accelerate time-to-validation and de-risk the regulatory pathway by demonstrating robust performance across diverse tissue.

What I Delivered Over 18 Months

Weekly iteration cycles on ML architecture, data strategy, and error analysis—establishing a tight feedback loop between model performance, data quality, and hardware improvements.

The Technical Challenge

• Mixed tissue heterogeneity - Invasive carcinoma, DCIS, and normal tissue often present in the same sample
• Registration uncertainty - Spatial alignment between autofluorescence and H&E images introduced label noise
• Limited labeled data - Small dataset across multiple device iterations with varying optical characteristics
• Real-time inference requirement - Clinical workflow demanded fast processing with no sample preparation

Core Deliverables

• Pathology-anchored labeling pipeline with targeted flagging of registration/labeling errors → cleaned training data for fairer comparisons
• Error analysis → data cleaning feedback loop → identified problematic samples and labeling issues, enabling targeted dataset improvements
• Method triage across spectral features, dictionary learning, CNNs, and weak supervision → focused effort on 2-3 viable approaches rather than exploring 10+ dead ends

Business Impact

Published validation: 82% sensitivity, 91% specificity, 0.92 AUC across 172 tissue blocks from 115 patients — peer-reviewed in Archives of Pathology & Laboratory Medicine (Nov 2023)

De-risked regulatory path: Data-quality processes and cross-validated results prepared CytoVeris for regulatory engagement

Accelerated product-market fit: Label-free, no-prep approach maintained clinical workflow compatibility — critical for adoption in time-pressured surgical and pathology settings

Capital efficiency: Method triage avoided 6+ months of scattered experimentation; focused ML roadmap gave investors confidence in technical feasibility

Why It Worked

CytoVeris had a lean technical team managing hardware integration, production infrastructure, product development, and ML research. I brought specialized cancer-imaging ML expertise and structured experimentation protocols that:

• Freed the internal team to focus on device integration and infrastructure while driving the ML research agenda
• Reduced cycle time from hypothesis to validated result through weekly iteration loops
• Caught data-quality issues (registration errors, label noise) that would have undermined credibility with clinical partners and regulators
• Translated ML errors into targeted data improvements—avoiding the “just collect more data” trap that burns runway

The leverage point: Their team knew the device and clinical workflow intimately; I brought deep pattern-recognition expertise for messy, heterogeneous biomedical images. This division of labor let them move from prototype to peer-reviewed validation in 18 months while staying capital-efficient.

Client Testimonial

“Working with Heather and Pixel Scientia Labs was a great choice for us… Her deep understanding of the diverse range of ML architectures and capabilities allows us focus on the data preprocessing, data augmentation, and validation techniques that produce the most robust results given our unique data and challenges.”

— Alan Kersey, President & CEO, CytoVeris

Evidence

Peer-reviewed publication: AI-Powered Biomolecular-Specific and Label-Free Multispectral Imaging Rapidly Detects Malignant Neoplasm in Surgically Excised Breast Tissue Specimens, Archives of Pathology & Laboratory Medicine, 2023