Low Dose CT Noise Artifact Reduction Based on Multi-scale Weighted Convolutional Coding Network

Low dose computed tomography becomes increasingly important to alleviate X-ray radiation damage. However, low tube current or short exposure time scanning protocols often lead to degraded CT images with increased amplified spot noise and radial streak artifacts. In this paper, we aims to obtain high...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2021 7th International Conference on Systems and Informatics (ICSAI) s. 1 - 6
Hlavní autoři: Liu, Jin, Xia, Zhenyu, Kang, Yanqin, Qiang, Jun
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 13.11.2021
Témata:
Computed tomography
Convolution
Convolutional codes
convolutional coding
Deep learning
Image coding
learning reweighted
low dose CT
Magnetic resonance imaging
multi-scale
Ultrasonic imaging
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Low dose computed tomography becomes increasingly important to alleviate X-ray radiation damage. However, low tube current or short exposure time scanning protocols often lead to degraded CT images with increased amplified spot noise and radial streak artifacts. In this paper, we aims to obtain high-quality images from LDCT image using a multi-scale weighted convolutional coding network (MW-CCN). Referenced by the iterative shrinkage thresholding algorithms (ISTA), we transformed the traditional convolutional sparse coding to its reweighted convolutional representation and learning update form. Further, to extract multi-scale image features we use a series of dilated convolution layers, which promote our model to capture the hierarchy features of different scales. To automatically correct the weights of the feature maps code, the channel attention mechanism is adopted. At last, our method combined the multi-scale image features extraction and weighted convolutional coding with a learning-based deep CNN. The whole network architecture yields better interpretation. Compared with the recent state-of-the-art methods, our network's processing effect is more significant in the experiments.
DOI:10.1109/ICSAI53574.2021.9664134