4247

View Abstract / View Presentation

Deep Learning-Based Needle Tracking Trained on Bloch-Simulated Data and Evaluated on Clinical Real-Time bSSFP Images

Ralf Vogel^1,2, Dieter Ritter², Jonathan Weine^2,3, Jonas Faust^2,4, Elodie Breton⁵, Julien Garnon^5,6, Afshin Gangi^5,6, Andreas Maier¹, and Florian Maier²
¹Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany, ²Siemens Healthcare, Erlangen, Germany, ³TU Dortmund, Dortmund, Germany, ⁴Universität Heidelberg, Heidelberg, Germany, ⁵ICube UMR7357, University of Strasbourg, CNRS, FMTS, Strasbourg, France, ⁶Imagerie Interventionnelle, Hôpitaux Universitaires de Strasbourg, Strasbourg, France

Synthetic training data was successfully used to train a U-Net for image-based needle tracking, utilizing automatically generated ground-truth labels. The evaluation on clinical patient data indicates that synthetic MR images can replace patient data and animal experiments.

Figure 1: Needle predictions on clinical interventional images (red overlays), generated by U-net trained with synthetic data only.

Figure 5: (a) random selection of simulated images generated using the AustinMan model. All images were simulated in the transversal plane, tilted by up to ±45° to the coronal plane. The needle was either placed in liver tissue from a reasonable angle or completely random (11, 12, 14, 15). (b) Image including the automatically generated needle label (green).