Various AI-based algorithms have been introduced more recently with the aim to overcome shortcomings of conventional denoising methods.
AI-based denoising methods:
Current DL algorithms use Convolutional Neural Networks (CNN) for noise suppression. Generally, DL algorithms can be applied at currently three levels of the image reconstruction pipeline:
1. Channel combination (in-line processing):
DL algorithms synthesize individual k-space information from each coil element including coil sensitivity profiles with the advantage of access to unaltered data from each coil element allowing production of complex imaging data with high signal fidelity.
2. Image generation (in-line processing):
DL algorithms applied to complex imaging data use the phase information and allow iteration between the AI processed imaging data and the k-space data that is generated by the inverse Fourier transform. Here the individual coil element information is already combined and integrated within the image. Raw k-space data of the individual coil elements is not available and can no longer be used for optimization such as checking for data consistency in the iterative reconstruction.
3. Postprocessing (off-line processing)
DICOM data is processed as an additional step after complete image reconstruction.
However, information removed in the earlier steps of the reconstruction, such as phase information, is lost and can no longer be used for image optimization.
In-line - based denoising
Access to raw data is an advantage of in-line reconstruction as no or little loss of data occurs at the time of processing, depending on the location of implementation within the reconstruction pipeline. Depending on vendor solution, the algorithms may be input-agnostic, which means they may operate with any pulse sequence and various anatomies.
However, some algorithms may be limited to certain pulse sequence designs.
Another advantage of in-line processing is that the user can select various levels of denoising or a combination of denoising and other enhancements such as increased resolution with only one acquisition.
. These may be selected before or after the acquisition, depending on the implementation.
Caution is advised when selecting denoising levels, as it may be difficult to predict the best level for different patients given that body habitus and motion dictate the imaging outcome. Overprocessing may result in smear depending on the pulse sequence and type of AI algorithm.
Postprocessing - based AI denoising
Postprocessing - based denoising programs are algorithms that operate on the image domain rather than on the raw data domain. They are usually vendor agnostic and can be applied to any pulse sequence on any system.
Some DICOM - based denoising AI-algorithms are limited to prospective applications only; acquisition parameters have to be adjusted upfront depending on the desired level of denoising after postprocessing, as postprocessing algorithms are input-agnostic just as in-line methods are.
The advantage lies in their vendor agnostic architecture, where the software can be applied to any existing MR system, given that is operates based on DICOM images. This offers a high degree of flexibility within any imaging enterprise and the ability to denoise images without the need to upgrade vendor-specific existing hard- or software.
Strategies for the implementation of AI-based noise reduction
The significance of computational rather than hardware-based denoising lies in the flexibility of its application: improving SNR can be done with the goal to:
- improve image quality without compromising acquisition parameters while maintaining acquisition time
- sparse sampling and reduce acquisition time to reduce motion artifacts and increase patient throughput
- aim for hybrid goal of improving image quality and increase throughput at the same time
[
Specifically, strategies B and C may play an increasing role in the future as cost-efficient imaging is crucial in modern healthcare systems. More seamless implementations of powerful in-line and post-processing noise reduction methods can be expected with the advent of increasingly sophisticated convolutional networks and computational capacity, allowing faster acquisitions with sparse sampling resulting in better image quality, increased patient comfort and throughput.
Challenges persist due to variable vendor-specific options and implementations of current DL- based denoising algorithms that require some adjustment for radiologists and decision making for various imaging applications that cannot be generalized for all patients or indications.
Flexible DL-based denoising options “on-the-fly” at the user end with seamless implementation in the PACS environment would be a desirable goal and can be envisioned for the future.