Cancer Research in NEC Labs America

1
Cancer Researchin NEC Labs AmericasMachine
Learning Dept.
Matt Miller Research Staff Member NEC Labs America
2
Modern Tools for Oncology

• Detailed diagnostics
• Targeted therapies
• Methods of mapping diagnostic results to
  therapies (personalized medicine)
• Methods of testing these tools

3
How Machine Learning Fits In

• Detailed diagnostics
• Pattern recognition
• Targeted therapies
• Machine-assisted drug design
• Methods of mapping diagnostic results to
  therapies (personalized medicine)
• Cocktail design
• Methods of testing these tools
• Statistical learning theory

4
How Machine Learning Fits In

• Detailed diagnostics
• Pattern recognition
• Targeted therapies
• Machine-assisted drug design
• Methods of mapping diagnostic results to
  therapies (personalized medicine)
• Cocktail design
• Methods of testing these tools
• Statistical learning theory

5
NECs Digital Pathology System

• Given a digitized pathology image, determine
  whether benign or malignant.
• Initial application double-check diagnoses of
  human pathologists.
• First implemented for gastric cancer.
• Current development work
• Colon cancer
• Breast cancer
• Prostrate cancer

6
NECs Digital Pathology System
ROIs
Image of whole tissue
Lo-res analysis (decision-tree like method)
Malignant/ Benign
Hi-res analysis (SVMs, CNNs)
7
NECs Digital Pathology System

• Gastric system tested on 1905 biopsies.
• Compared against diagnoses by three human
  pathologists.
• 100 malignant cases
• Results
• 227 / 1805 (12.6) false positives
• 1 / 100 false negative

8
Cocktail Design

• Assume there is an unknown function
• Find a function
• Such that
• is good

9
Cocktail Design We Have

• Some theoretical results due to Vapnik.
• Proposal for an algorithm.
• Methods for handling real-world problems
• Limits on what machine may try while learning.
• Paucity of real cases.
• No data to work on.

10
NEC ML People Here Today