title:: Real-time ultrasound transducer localization in fluoroscopy images by transfer learning from synthetic training data

publisher:: Elsevier

people:: Tobias Heimann, Peter Mountney, Matthias John, Razvan Ionasec

organization:: Siemens AG, Siemens Corporation

domain:: Medical Imaging, Machine Learning, Computer Vision

link:: https://pubmed.ncbi.nlm.nih.gov/24856101/

publications

Summary

This paper presents a machine learning approach to localize an ultrasound transducer in X-ray fluoroscopy images, which is important for image fusion between ultrasound and X-ray data during minimally invasive cardiac interventions. A key challenge is obtaining enough training data to cover the 6 degrees of freedom of the transducer pose. The proposed solution generates synthetic training data automatically from a single volumetric scan of the transducer. To adapt to real fluoroscopy data, unlabeled images are used to estimate and correct differences between the synthetic and real feature distributions via instance weighting techniques like probabilistic classification and Kullback-Leibler importance estimation. Experiments on over 1900 images show the approach reduces detection failures from 7.3% to 0% and improves the localization error from 1.5mm to 0.8mm compared to training only on real data.

Data Points

  • Addresses localizing ultrasound transducer in fluoroscopy X-ray images for image fusion
  • Uses discriminative learning but gathering enough training data is challenging
  • Automatically generates synthetic training data from single transducer CT scan
  • Employs domain adaptation via instance weighting to match real fluoroscopy data
  • Evaluates probabilistic classification and KLIEP instance weighting methods
  • Tested on over 1900 images, reduced detection failures from 7.3% to 0%
  • Improved localization error from 1.5mm to 0.8mm compared to real data training
  • Enables flexible adaptation to different medical devices with minimal effort