(Wednesday, 8/1/2018) 10:15 AM - 12:15 PM
Room: Room 207
Purpose: The purpose of this work was to develop a deep learning method to generate volumetric scout CT image volumes from conventional projection scout radiographs.
Methods: 751 clinically indicated chest-abdomen-pelvis (CAP) CT exams from 667 patients were retrospectively collected. Inclusion criteria were: 1) CT exams acquired with or without contrast media and 2) scanned with anatomical ranges in six different regions (chest-alone, abdomen-alone, pelvis-alone, chest-abdomen, abdomen-pelvis or CAP). In order to account for differences in image reconstruction parameters, CT image volumes were interpolated to a 1.0x1.0x1.0 mm³ isotropic voxel size. To avoid potential data inconsistency, large patients with significant truncation artifacts were excluded. Finally, interpolated CT image volumes were registered to the CT radiograph localizers using the patient’s positioning information in the DICOM header. A total of 211,659 radiograph localizers and corresponding CT images were divided into 163,840 images from 476 patients for training and 47,819 images from 191 patients for testing a 36-layer deep neural network which inputs the AP and lateral localizers and outputs a volumetric CT scout. In order to test the generalization error of the trained deep neural network, the total patient attenuation was measured in a randomly selected subset of 1000 CT and volumetric scout images for both the training and testing datasets and the absolute percent difference in total patient attenuation was calculated.
Results: Mean±SD for the total patient attenuation difference between the diagnostic CT and volumetric scout images were 1.5±1.1% and 3.3±2.3% (n=1000 images) for the training and testing datasets.
Conclusion: A deep learning method to reconstruct a volumetric scout CT image volume from projection radiograph localizers was developed. The proposed method could enable more accurate radiation dose estimates and scanning parameter prescription prior to the CT acquisition and help to overcome the current limitations of automatic exposure control schemes in diagnostic CT.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
