Click here to


Are you sure ?

Yes, do it No, cancel

Quantitative Evaluation of Image Quality of Deep-Learning-Based CT Reconstruction Using Structural SIMilarity (SSIM)

K Yang*, A Parakh, R Gupta, A Kambadakone, X Li, B Liu, Massachusetts General Hospital, Harvard Medical School, Boston, MA


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

Room: AAPM ePoster Library

Purpose: To compare image quality between CT images reconstructed with three different algorithms, FBP, iterative (ASiR-V), and deep-learning (DLIR, TrueFidelity), using a quantitative metric of Structural SIMilarity (SSIM).

Methods: Three different clinical exams (routine head, abdomen in delayed phase, CTA head and neck) acquired from a GE Revolution scanner were reconstructed using all three algorithms available on the scanner, while keeping all other parameters unchanged, including the standard reconstruction filter and two sets of slice thickness (1.25 mm and 5 mm). For the DLIR algorithm, three different levels of strength were used (low, medium, and high). For visual observation, direct image subtraction was applied between each image-set and the FBP (as baseline). Quantitative comparison was performed using Structural SIMilarity (SSIM) as a figure of merit. All CT images were initially segmented to exclude air outside the anatomy within the image FOV. Using FBP as the baseline, a SSIM matrix was calculated for each image pair. The average value of each SSIM matrix was used as a single numeric figure to evaluate the similarity between two CT images. SSIM of 1 indicates 100% similarity and 0 indicates complete difference.

Results: SSIM shows clear dependency on reconstruction algorithms. For the abdomen delay case with 1.25 mm slice thickness, SSIM was the highest for ASiR-V (level 40%, mean of 0.98) and lowest for DLIR-high (mean of 0.82). For the head case with 1.25 mm slice thickness, SSIM was the highest for DLIR-low (mean of 0.97) and the lowest for DLIR-high (mean of 0.92).

Conclusion: SSIM is a very sensitive quantitative metric to compare CT image similarities reconstructed with different algorithms. The current observation shows that ASiR-V (with level below 60%) is more similar to FBP than DLIR. DLIR-low was more similar to FBP than DLIR high for all exams.

Download ePoster [PDF]


CT, Image Analysis, Reconstruction


IM- CT: Quality Control and Image Quality Assessment

Contact Email