High-fidelity geometry generation from CT data using convolutional neural networks

Aniket Tekawade; Brandon A. Sforzo; Katarzyna E. Matusik; Alan L. Kastengren; Christopher F. Powell

doi:10.1117/12.2540442

10 September 2019 High-fidelity geometry generation from CT data using convolutional neural networks

Aniket Tekawade, Brandon A. Sforzo, Katarzyna E. Matusik, Alan L. Kastengren, Christopher F. Powell

Proceedings Volume 11113, Developments in X-Ray Tomography XII; 111131X (2019) https://doi.org/10.1117/12.2540442
Event: SPIE Optical Engineering + Applications, 2019, San Diego, California, United States

Abstract

The X-ray Fuel Spray research at Argonne National Laboratory is aimed at utilizing synchrotron X-ray diagnostics for providing insights into automotive fuel injection. One important task is to generate high-fidelity geometries or iso-surfaces of steel fuel injector nozzles from X-ray Computed Tomography measurements, to be used as inputs to realistic CFD simulations of fuel injector flow. These fuel nozzles contain 3D features between 5 - 500 micron and are imaged at a pixel resolution of 1 micron. The main bottleneck to automated generation of an STL geometry from X-ray CT data is the segmentation or surface determination process – conversion of the CT volume into a binary map that classifies each voxel as belonging to either injector metal or flow domain, accurately locating the metal surface at the transitions between these domains. Here, we describe our recent success in automating the segmentation process itself, which is challenging because various artifacts that arise from X-ray imaging and CT reconstruction confound the identification of threshold values needed for traditional segmentation algorithms. A convolutional neural network (CNN) coupled with a tailored loss function is implemented to achieve state-of-the-art accuracy in surface localization with limited manual intervention. Through data augmentation, the model can be trained on less than 30% of the slices drawn from two CT scans of different automotive injectors that were manually segmented and is tested on a third. Our architecture achieves state-of-the-art accuracy at lower computation time and GPU memory requirement compared to U-net, one of the most popular architectures for image segmentation.

Citation Download Citation

Aniket Tekawade, Brandon A. Sforzo, Katarzyna E. Matusik, Alan L. Kastengren, and Christopher F. Powell "High-fidelity geometry generation from CT data using convolutional neural networks", Proc. SPIE 11113, Developments in X-Ray Tomography XII, 111131X (10 September 2019); https://doi.org/10.1117/12.2540442

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available