(Wednesday, 7/17/2019) 7:30 AM - 9:30 AM
Room: 221AB
Purpose: To simultaneously reduce noise in dual energy CT spectral images while maintaining CT number accuracy using a deep convolutional neural network with an application-specific architecture and regularization function.
Methods: A convolutional neural network (CNN) was trained to subtract noise from dual energy computed tomography (DECT) images. The CNN architecture consisted of two parallel branches of 8 convolutional layers that operated on 80 kVp and 150 Sn kVp images separately, followed by 5 convolutional layers that operated on combined image features from both spectral images. Activity regularization was used to ensure that noise corrections do not locally bias the CT number. The training and testing datasets were constructed from 13 contrast-enhanced abdominopelvic CT exams acquired with a DECT scanner. A validated projection-based noise insertion was used to simulate low-dose (LD) exams at 40% of the full dose (FD) used for the original exam. Training was performed using 250,000 randomly centered patches cropped from 10 scans, with 3 scans reserved for validation and testing. The CNN was optimized for 50 epochs using the LD image patches as inputs and the FD image patches as the target outputs. After training, the denoising performance was assessed using root mean square (RMS) measurements at specific regions of interest (ROIs) and visual inspection of both spectral images. The mean pixelwise difference from the original images was used to check CT number accuracy after denoising.
Results: After CNN noise subtraction, the RMS in uniform ROIs was reduced by 46% for the LD images and by 40% for the FD images. The measured bias was less than 0.027 HU.
Conclusion: A deep CNN was able to reduce noise in DECT images for both high and low kVp images and across multiple dose levels while maintaining spatial resolution and CT number accuracy.
Keywords
Noise Reduction, Dual-energy Imaging
Taxonomy
IM- CT: Machine learning, computer vision
Contact Email
