MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Deep Learning-Based End-To-End CT Reconstruction Method

K Lu*, L Ren, F Yin, Duke University, Durham, NC

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: To develop a deep learning-based end-to-end CT reconstruction method and compare it with traditional CT reconstruction methods of filtered-backprojection (FBP) and simultaneous-algebraic-reconstruction-technique (SART).

Methods: A deep learning-based CT reconstruction method was developed to perform CT reconstruction from X-ray projection data (sinogram). The method contains a custom-made deep learning architecture, which incorporates both Convolutional Neural Networks (CNN) with fully connected layer and Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN). Two-thousand training data with simulated varying internal structures and one test data were generated from Shepp-Logan phantom (128x128 pixels) for 32 and 64 parallel projections spread evenly over 180 degrees. The detector size was 128x1 pixels with the source-to-isocenter distance of 750mm and the source-to-detector distance of 120mm. Three methods (deep learning-based, FBP, and 50 iterations of SART methods) were compared for reconstructing the testing data with the same geometric setup. The performance of the three methods was assessed by comparing the reconstruction speed and quality of the reconstructed images, quantified by peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) index.

Results: Both the deep learning-based method and the FBP method completed the image reconstruction within 1 second (0.48s for the deep learning-based method and 0.04s for the FBP method using 32 projections; 0.50s and 0.05s respectively for 64 projections), while the SART took 149s for 32 projections, and 223s for 64 projections. PSNR (and SSIM) values of images reconstructed by deep learning-based, FBP, and SART methods are 29.553 (0.977), 11.891 (0.366), and 22.958 (0.918) for 32 projections; 29.120 (0.973), 17.366 (0.467), and 23.571 (0.934) for 64 projections.

Conclusion: Comparing to the traditional FBP and SART methods, the deep learning-based CT reconstruction method achieved better image quality and fast reconstruction, demonstrating great potential for practical application in real-time high-quality CT image reconstruction.

Keywords

Reconstruction, CT

Taxonomy

IM- CT: Machine learning, computer vision

Contact Email