MENU

Click here to

Ă—

Are you sure ?

Yes, do it No, cancel

Material Decomposition of X-Ray Radiography for Accurate Target Tracking and Patient Setup Using Deep Neural Network

Y Lin*, D Lam, S Zhou, Y Tan, Q Chen, B Sun, T Zhang, Washington University School of Medicine, St. Louis, MO

Presentations

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

Room: AAPM ePoster Library

Purpose: Dual-energy subtraction (DES) material decomposition may improve the visibility of soft tissue target in x-ray images. However, it requires special hardware that are not commonly available. The purpose of this study is to generate tissues- and bone-only decomposition from conventional X-ray radiography (CXR) using a deep-learning model for image-guided radiotherapy.

Methods: Due to the lack of clinical DES images, we generated DES digitally reconstructed radiography (DRR) images from CT volumetric data as training images. Each CT dataset was segmented into bone and soft tissue images, followed by upsampling using an edge-preserving algorithm. After high resolution bone- and tissues-only DRRs were generated, patches were randomly cropped from each DRR to be used as training samples. The model adopted a U-net architecture with a residual block at the center, with the original DRR as input while the corresponding bone- and tissues-only DRRs were set as the outputs. To create full-sized decomposed radiographic images, the patch predictions from the model were stitched together using an image stitching algorithm developed in Python.

Results: After training for 120 epochs, the model achieved peak signal-to-noise ratios (PSNR) of 34.2dB and 32.8dB for the AP and lateral models, respectively. The stitching algorithm removed some irregular localized errors by averaging the overlapping regions between neighboring patches, though losing some resolution in the process. Overall, the model produced decomposed bone and soft tissue images with improved soft tissue contrast.

Conclusion: The work suggests the potential of AI-based image processing for tissues-bone decomposition based on traditional single energy X-ray images. The decomposed images enable easier target tracking and patient setup without the need for additional DES hardware.

Funding Support, Disclosures, and Conflict of Interest: This project is funded by the Varian Research Grant.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email