Diagnosis of Bone Tumors Using Radiographs

Team:

  • Bao H. Do, MD
  • Rui Shu
  • Curtis P. Langlotz, MD, PhD
  • Christopher Beaulieu, MD, PhD

Training Details

  • 811 annotated cases, 66 unique diagnoses
  • Evaluated 3 overlapping subsets:
    • (Rare) Minimum 5 examples of each diagnosis: 710 cases, 29 unique diagnoses
    • (Intermediate) Minimum 18 examples of each diagnosis: 559 cases, 14 unique diagnoses
    • (Common) Minimum 30 examples of each diagnosis: 478 cases, 10 unique diagnoses

Results

Future Work

  • Test against other machine learning algorithms and against human performance
  • Incorporate published priors, e.g. Dahlin’s Bone Tumors
  • Analysis of correlations between features
  • Compare to image classification using deep learning
  • Pair with feature extraction using deep learning
  • Increase sample size, ideally N > 30 per diagnosis