In silico Evaluation Vs Standard Phantom Evaluation of a Deep Learning Reconstruction Algorithm šŸ“

Author: Naruomi Akino, Kirsten Lee Boedeker, Ilmar Hein, Dylan Mather, Akira Nishikori, Daniel W Shin šŸ‘Øā€šŸ”¬

Affiliation: Canon Medical Systems Corporation, Canon Medical Research USA šŸŒ

Abstract:

Purpose: To validate the performance a deep learning reconstruction (DLR) algorithm in an anatomical background compared to a uniform phantom background.
Methods: An analytic forward projection model based on the specifications of a wide volume CT scanner (Canon Medical Systems, Aquilion ONE PRISM, Otawara, Japan), including detector, source, and geometry, was used for simulation. The sinogram of low contrast rods, 5mm in diameter and 7HU in contrast, were simulated and inserted into an existing patient sinogram, with varying levels of added noise, and reconstructed with a commercial denoising deep learning reconstruction algorithms (AiCE and PIQE). A phantom with uniform background containing the same test object was scanned at the same corresponding SSDEs and the detectability index (d’) assessed via non-prewhitening model observer. The Modulation Transfer Function (MTF) was first measured using a standard edge-based approach on images acquired of the CATPHANTM 600 with uniform background and reconstructed with DLR and a hybrid iterative reconstruction algorithm (HIR). The cylindrical test objects used for assessment were then simulated via forward projection using the system model and inserted into patient raw data.
Results: The Pearson correlation of the detectability index in phantom vs anatomical background ranged from .87 to .99 for DLR across the dose range examined. The relative MTF behavior across dose/noise levels in anatomical background vs phantom was also highly correlated for DLR vs HIR, within 10%.
Conclusion: This work demonstrates that for the DLR examined, low contrast detectability performance in phantom with a uniform phantom was well correlated with corresponding test objects inserted into an anatomical background. Similarly, the relative MTF performance of DLR vs hybrid iterative reconstruction was well represented in a uniform phantom relative to the same test object inserted into an anatomical background.

Back to List