Encrypted login | home

Program Information

High-Resolution CT Image Reconstruction Using Sparse-Coding-Based Deep Learning

X Yang

X Yang1*, Y Lei2 , K Higgins3 , Z Zhou4 , X Jiang5 , W Curran6 , (1) Emory University, Atlanta, AA, (2) Emory University, Atlanta, GA, (3) Emory University, Atlanta, Georgia, (4) Nanjing University, Nanjing, Jiangsu, (5) Emory University, Atlanta, GA, (6) Emory University, Atlanta, GA


SU-K-FS4-12 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: Four Seasons 4

Purpose: Due to CT x-ray exposure and respiratory motion, a routine planning CT (lung or abdomen) is usually captured with a large slice thickness (e.g. 3-4mm). Such CT images with distinctly low out-of-slice resolution will affect both contouring and dose calculation in treatment planning. The purpose of this study is to develop a deep-network-based method to reconstruct high-resolution CT images from routine CT images for radiotherapy treatment planning.

Methods: We propose to integrate the domain expertise of sparse coding and the merits of deep learning to learn a mapping between high- and low-resolution CT images for reconstructing high-resolution CT from routine CT. Firstly, the patches of each pixel are extracted from the low-resolution CT and each patch is represented as a high-dimensional vector including a set of feature maps. Second, each high-dimensional vector is nonlinearly mapped onto another high-dimensional vector using deep convolutional neural network (CNN). Each mapped vector is the representation of a high-resolution patch. Finally, the high-resolution patch-wise representations are combined to generate the final high-resolution CT. We used different low-resolution CT down-sampled from original CT as the inputs and compared the retrieved high-resolution CT (output) with the original CT to evaluate the reconstruction accuracy.

Results: We used a dataset with 250 coronal lung CT images to test the proposed method. In order to get a quantitative evaluation, mean absolute error (MAE) and structural similarity (SSIM) indexes were used to quantify the reconstruction accuracy. Overall the mean MAE and SSIM between the retrieved and original CT were 18.23±3.05 and 0.95±0.01, which demonstrated the restoration accuracy of our method.

Conclusion: We have investigated a new approach to retrieve high-resolution CT using sparse-coding-based CNN and demonstrated its reliability. The proposed method has great potential in improving radiation dose calculation and delivery accuracy and decreasing CT radiation exposure of patients.

Contact Email: