This Deep Learning CT presentation was given at AAPM 2021, and is presented here with permission from the AAPM.
For more information on the basics behind x-ray and CT acquisitions and reconstruction please see: howradiologyworks.com.
Deep Learning in CT (Computed Tomography) is a new technology which replaces iterative reconstruction methods. At GE CT the specific name of the product offering for Deep Learning in CT is TrueFidelity. Deep learning offers similar advantages of improved contrast to noise ratio as iterative reconstruction and model based iterative reconstruction, but deep learning can offer improved image texture and improved reconstruction time.
The GE Healthcare white paper is available here:
www.gehealthcare.com/-/jssmed...
Since we strive for bite size content on topics in Radiology on this channel we have split up the talk into three sections.
The aim of applying deep learning for CT is to improve over the state of the art for reconstruction which is iterative reconstruction. The improvement requested most frequently for iterative reconstruction is not to reduce the image noise further, but rather to keep the image texture more similar to a higher dose Filtered Backprojection (FBP) reconstruction.
Since there is a noise texture penalty associated with existing iterative reconstruction methods users typically use a lower strength of iterative reconstruction.
In the work of Kim et. al. significant improvement in the contrast to noise ratio (CNR) is measured in clinically relevant soft tissue structures within the brain. These CNR values are a significant gain over their currently acceptable level of iterative reconstruction.
In addition to measuring the contrast to noise it is important to assess the spatial resolution. The modulation transfer function is the means of measuring the spatial resolution. Some iterative reconstruction methods have been shown to have varying spatial resolution depending on the material type. This was assessed for TrueFidelity images by Szczykutowicz et al. www.ajronline.org/doi/abs/10.... and was shown that the spatial resolution is not highly dependent on the tissue type.
As discussed above the main goal of using Deep Learning for CT reconstruction in the TrueFidelity design is to improve upon the noise texture of iterative reconstruction techniques. In early experiments this was demonstrated in a pig liver phantom where the noise correlation yields a less plasticy visual impression.
The noise texture is quantified here using the Noise Power Spectrum (NPS) and the normalized NPS for DLIR generated TrueFidelity images closely matches Filtered BackProjection (FBP) whereas the iterative reconstruction yields a left shift in the noise power spectrum (i.e. the lower frequencies make up a larger proportion of the noise).
There are many ways to train the neural networks and one possible way is to use Nobel based iterative reconstruction (MBIR) as the ground truth. In this case the network can well approximate MBIR and the deep learning results will have a similar noise power spectrum as MBIR.
Another advantage of DLIR is the reduced noise for thin slice images. This is an area for future research on the potential for improving practice if thinner images can be reviewed more regularly.
The GE approach for DLIR on Rev CT and Revolution Apex also offers reduced motion artifacts since the significant noise reduction may be used in conjunction with a smoother view weighting for improvement in motion artifacts.
DLIR can technically encompass many parts of the image chain including the physics corrections and even enhance the image contrast. The current GE approach is to build upon the years of existing well tested physics models. The GE approach is also to use the algorithm for noise reduction and not contrast modification.
The Deep Learning approach can also be applied to dual energy reconstruction as long as careful consideration is used.