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 True Fidelity. 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:
[ Ссылка ]
And general information on x-ray and CT education is available here:
[ Ссылка ]
Since we try to have bite size content on topics in Radiology on this channel we have split up the talk into three sections.
1) What are the aspects that are critical to neuro CT acquisition quality?
2) How does Deep Learning for CT compare with iterative reconstruction?
3) Evaluation of Deep Learning for CT compared with iterative reconstruction.
This video is focused on (2) and the other topics will be linked in a playlist at the end of this video.
Chapters:
00:00 AOCNR 2021 Introduction
00:50. FBP vs IR vs DLIR background
02:00. Reconstruction speed vs CNR (contrast to noise)
02:56. Image texture vs CNR (contrast to noise)
03:40 AI vs ML vs DL
05:03. FBP vs IR vs DLIR images
05:40. Training Deep Learning CT
07:32. Supervised Learning Deep Learning CT
Deep learning is a subset of machine learning which is a subset of artificial intelligence (AI), where the power of deep learning is that the engineers do not need to develop specific filters for image characteristics but rather then network can learn this itself.
The power of deep learning comes in the ability to well leverage large amounts of training data and in comparison with iterative reconstruction can use models which are orders of magnitude larger. This enables deep learning to better estimate the correlated noise in CT images.
We go through the progression from filtered backprojection (FBP) in 1972 to iterative reconstruction in 2008 to deep learning image reconstruction in 2018. Where FBP can generate images quickly and as long as the dose is relatively high the quality is very good. On the other hand when there is the desire to reduce the radiation dose iterative reconstruction algorithms enable reduced noise compared with FBP. However, the image texture of iterative reconstruction algorithms has been a reason slowing the implementation of iterative reconstruction.
Deep learning algorithms can be trained to achieve high dose FBP image quality and reconstruction times are significantly faster than model based iterative reconstruction frameworks.
The networks which are employed today on clinical scanners are trained at the factory and deployed in the field, i.e. the network is not learning at the clinical site.
