Abstracts

RETURN TO ABSTRACT LISTING


1775. Deep Learning Image Reconstruction Versus Filtered Back Projection on Computed Tomography Image Quality and Coronary Artery Calcium Score
Authors * Denotes Presenting Author
  1. Zhaoying Xian *; Houston Methodist Hospital
  2. Justin Schmidgall; Houston Methodist Hospital
  3. John Nance; Houston Methodist Hospital
  4. Pamela Mager; Houston Methodist Hospital
  5. Lienard Chang; Houston Methodist Hospital
  6. Pankaj Patel; Houston Methodist Hospital
  7. Nakul Gupta; Houston Methodist Hospital
Objective:
To compare image quality and coronary artery calcium score (CACS) obtained from computed tomography (CT) scans reconstructed with deep learning image reconstruction (DLIR) vs. filtered back projection (FBP).

Materials and Methods:
38 CACS scans were reconstructed using both DLIR (TrueFidelity®; GE healthcare) and FBP. Scans were obtained on a 256-detector CT (Revolution CT; GE Healthcare) at 120kVp using automated exposure control and reconstructed at 2.5 mm thickness. DLIR was set at ‘medium’ strength. CACS scores were obtained semi-automatically by a single operator (SmartScore 4.0; GE Healthcare). CACS severity was categorized as none (0), mild (1 - 100), moderate (101 - 400) or severe (> 400). Noise was assessed as the SD of a circular ROI over the ascending aorta. Size specific dose estimate (SSDE) was recorded from radiation dose tracking software (Imalogix; King of Prussia, PA). Noise values were compared with a paired t-test. A Bland-Altman difference plot was used to assess agreement of CACS. Agreement between severity categories was assessed with kappa statistics. Multiple linear regression was used to predict DLIR CACS from FBP CACS and SSDE.

Results:
Noise was significantly lower on DLIR images (16.6 +/- 4.9 HU vs. 26.1 +/- 4.1 HU, p < 0.001). For CACS between 0 - 100, the mean difference +/- SD between FBP and DLIR was 0.5 +/- 2.9 with 95% limits of agreement (LOA) of -5.1 - 6.2. For CACS > 100, the percent difference was assessed due to wide dispersion, and was 3.3 +/- 2.9 %, with 95% LOA of -2.4 - 8.9%. A significant regression equation was found on multiple linear regression analysis (F(2, 32) = 24561.71, p < 0.001), with an R2 of 0.999. SSDE was not a significant predictor (p = 0.729). DLIR score was equal to -0.4 + 0.992*(FBP CACS). There was perfect agreement among risk categories (k = 1.0), with no patients re-categorized by DLIR vs. FBP.

Conclusion:
DLIR significantly reduced noise without recategorization of coronary artery calcium severity scores. The 95% LOA for scores between 0 - 100 was -5.1 - 6.2, which may be acceptable for many applications. Further studies may explore if the reduced noise of DLIR may allow dose reduction without sacrificing CACS accuracy.