E1938. Convoluted Neural Network Denoising for Whole Body Low Dose CT
  1. Francis Baffour; Mayo Clinic
  2. Tara Anderson; Mayo Clinic
  3. Mark Adkins; Mayo Clinic
  4. Brian McCollough; Mayo Clinic
  5. Lifeng Yu; Mayo Clinic
  6. Michael Bruesewitz; Mayo Clinic
  7. Katrina Glazebrook; Mayo Clinic
Whole body low dose CT (WBLDCT) is the preferred imaging modality to evaluate osteolytic bone disease in patients with multiple myeloma. Images can be acquired with the patients’ arms up or down. Convoluted neural networks (CNN) can be utilized to reduce image noise and improve the diagnostic quality of images while decreasing the total radiation dose. The goal of this project is to compare the diagnostic quality of a variety of WBLDCTs acquired at different radiation doses and with different reconstruction methods.

Materials and Methods:
Three readers with 32 years of experience blindly compared 12 sets of images at different acquisition and reconstruction methods: 1) arms up full dose (UFD) with iterative reconstruction (IR); 2) arms up quarter dose (UQD) with IR; 3) arms up one-tenth dose (UTD) with IR; 4) UFD with CNN; 5) UQD with CNN; 6) UTD with CNN; 7) arms down full dose (DFD) with IR; and 8) DFD with CNN. Images were reconstructed and reviewed in bone and soft-tissue kernels. Reviewers evaluated the image sets on the ability to visualize bony and soft-tissue structures, and also provided a rank preference based on the ability to visualize bone detail, soft-tissue detail, and overall diagnostic quality. Results were analyzed using the Wilcoxon two-tailed significance test with statistical significance set at p </=0.05.

When comparing UFD + IR and DFD + IR, there was no significant difference in the readers’ perception of bone and soft-tissue visualization, readers’ rank order of preference for bone visualization, soft-tissue visualization, and overall image quality. Similar results were observed in a comparison of UFD +CNN and DFD + CNN. When comparing UFD + IR and UQD + CNN, there was no significant difference in the readers’ perception of bone visualization; however, there was a significant difference in soft-tissue visualization (UFD+IR = 2.2; UQD + CNN = 2.5, P = 0.03). Ranking of these image sets based on bone visualization (UFD+IR = 2; UQD + CNN = 3, P = 0.02) and overall image quality (UFD+IR = 2; UQD + CNN = 2.5, P = 0.01) were statistically significant.

These initial studies demonstrate similar diagnostic image quality between CTs obtained in arms up and arms down positions. Overall image quality of UFD+IR images is marginally better than UQD+CNN images.