Artificial Intelligence in Breast Imaging

Data Concepts

TermDefinition
Training dataDataset used to develop/train the algorithm
Validation dataSmaller training set used for cross-validation of model to test performance robustness and/or boost model performance during training stage
Test dataUnseen data withheld from training — used to assess real-world performance
External validationTesting the algorithm using data sets from a different population (different from the one used for initial development)

Key Principles

  • AI algorithms should perform equally well on diverse populations (age, race, density, risk, etc.)
  • Patient acceptance: a recent study shows 77% of patients do NOT support being solely screened by AI
  • AI is promising for:
    • Screening accuracy
    • Breast cancer risk prediction
    • Diagnostic radiology workflows
    • Clinical efficiencies

Current Performance

  • AI has been found to be superior or not inferior to human double reading in some studies
  • Limitations: performance may vary across populations different from the training population

Clinical Applications in Breast Imaging

  • Screening mammography: detection of masses, calcifications, asymmetries
  • BC risk prediction: using imaging features + patient data
  • Workflow optimization: triage, prioritization of worklists
  • Supplemental screening: decision support for supplemental US or MRI