To reduce radiation risk for patients, Dose reduction in computed tomography (CT) is essential. Iterative reconstruction algorithms turn out to be useful when it comes to deal with noise increased due to reduction in photon flux. The model uses K-sparse autoencoder (KSAE) for unsupervised feature learning and reconstruct a denoised image. The model performance was checked on 2016 Low-dose CT Grand Challenge dataset, containing projection and image data for chest and abdomen from 10 patients. Using a normal-dose image, a manifold was created and the distance between this manifold and reconstructed image was minimized along with data sanity during reconstruction. Results demonstrate that KSAE prior could perform better than dictionary learning and total variation for quarter-dose data, with good noise suppression and structural preservation. Model performance is acceptable until 1/6 dose.
Input variables : Low dose CT image
Output Variables : Denoised image
Visit Model : github.com
Additional links : ncbi.nlm.nih.gov
| Model Category | : | Public |
| Date Published | : | November, 2017 |
| Healthcare Domain | : | Provider |
| Code | : | github.com |
| Image Denoising |