Room: Room 202
Purpose: To reduce noise in both low-dose and routine-dose CT images while preserving anatomical details using a deep convolutional neural network (CNN).
Methods: A CNN was trained to identify noise textures in low-dose abdominal CT images using supervised learning. Network training was performed using image patches that were cropped from abdominal patient scans reconstructed using a medium-sharp kernel with 3 mm slices. Patches cropped from full-dose (FD) CT scans were used as the ground truth, whereas corresponding patches cropped from simulated quarter-dose (QD) scans were used as the network input. Input patches had a shape of (N,N,3), where N is an arbitrary number of pixels, and the channel dimension contains 3 adjacent longitudinal slices. The training data set consisted of 250,000 randomly selected image patches with N=64, and 20,000 randomly selected patches with N=256 for additional fine-tuning. The CNN architecture was based on the ResNeXt design, which employs aggregated residual transformations with 32 convolutional layers. The CNN output was subtracted from the input image to produce the final output. This resulted in the CNN learning to isolate the noise signal in the input images, which reduces training time and may improve the generality of the results. The performance on both QD and FD images was assessed using a set of reserved patient and phantom data that were not used in network training.
Results: The noise-subtracted images were perceptually smooth without compromising sharpness. The contrast-noise-ratio measured from phantom data was improved by 149% for QD images, and by 77% for FD images.
Conclusion: A deep CNN was able to reduce noise in CT images while maintaining a high level of anatomical detail. The method was robust enough to apply to images acquired with different dose levels, provided the kernel and slice thickness match those of the training data.
Not Applicable / None Entered.