Article Figures & Data
Figures
- FIG 1.
Overview of GAN for producing synthetic STIR volumes. Each patient’s 3 acquired MRI volumes (T1, T2, STIR) are utilized during training of a GAN. The 2 models, generator and discriminator, compete to improve their own performance by either generating (more realistic) synthetic STIR (sSTIR) images or discriminating between sSTIR and acquired STIR (aqSTIR).
- FIG 2.
Top, Example of an acquired STIR volume slice from 1 patient (blue) and the same patient’s equivalent synthetic volume slices generated from GANs with 4 different loss functions (yellow). Bottom, Three zoomed images that allow a comparison of the quality of the basi vertebral venous plexus between the acquired volume slice and 2 synthetic volume slices (SSIM+Sobel+MAE and Foreground-MAE).
- FIG 3.
A slice from a randomly selected volume from the test set, which was assigned a quality metric of 5/5 from the 3 reviewers. Left, Slice from the acquired STIR volume. Right, Same slice from the synthetic STIR volume.
- FIG 4.
Examples of a variety of imaging features. Each pair shows a sample sagittal slice from the acquired STIR volume with its corresponding model-generated synthetic STIR volume.
Tables
Variables Data Set Train Validation Test Patients 1633 84 100 Sex (% female) 53% 54% 47% Age (mean ± SD) 57.66 ± 17.7 55.05 ± 19.5 61.7 ± 17.8 Volumes 4899 252 300 T1/T2 slices 42,688 2121 2702 aNumber of patients (including sex and age information), number of image volumes, and number of slices in each randomized data set.
- Table 2:
Summary statistics of radiologists’ assessments of acquired and synthetic volumesa
Acquired STIR Volumes Synthetic STIR Volumes R1 R2 R3 R1 R2 R3 Quality (1–5) Average 3.92 3.38 3.53 3.83 3.01 3.31 SD 0.84 1.17 0.79 1.18 1.27 1.08 Median 4 3 4 4 3 4 Min-max 1–5 1–5 1–5 1–5 1–5 1–5 Motion artifact (0–3) Average 0.69 0.74 1.2 0.68 1.14 1.4 SD 0.81 0.88 0.653 0.89 1.04 0.8 Median 1 0.5 1 0 1 1 Min-max 0–3 0–3 0–3 0–3 0–3 0–3 Acquired vs synthetic (1–5) Average 4.16 3.85 3.48 1.11 1.11 1.63 SD 0.73 1.14 1.05 0.34 0.34 0.89 Median 4 4 3 1 1 1 Min-max 2–5 1–5 1–5 1–4 1–3 1–5 Fat suppression Y/N Yes % 100% 93% 100% 98% 81% 96% ↵a Summary statistics of radiologists’ assessments of acquired and synthetic STIR volumes from test set of 100 patients. Three radiologists (R1, R2, R3) each reviewed in a blinded fashion 1 synthetic STIR volume and 1 acquired STIR volume from 100 patients and evaluated each volume’s quality (1 = Unacceptable, 2 = Poor, 3 = Acceptable, 4 = Good, 5 = Excellent), degree of motion artifacts (0 = absent, 1 = mild, 2 = moderate, 3 = severe), whether they believed the volume to be acquired or synthetic (1 = definitely synthetic, 2 = probably synthetic, 3 = unsure, 4 = probably acquired, 5 = definitively acquired) and whether the fat suppression is homogeneous (Y/N) in the volume.
- Table 3:
Radiologist-assessed occurrence of pathologies present in acquired and synthetic volumesa
Pathologies Acquired STIR Synthetic STIR 2/3 Reviewers 3/3 Reviewers 2/3 Reviewers 3/3 Reviewers Bone marrow lesion 3% 2% 2% 2% Spinal cord signal change 0% 0% 0% 0% Paraspinous muscle edema 24% 9% 9% 3% Facet arthritis with synovitis 13% 2% 4% 1% Fracture with edema 2% 1% 0% 0% Modic type 1 change 41% 21% 19% 4% Synovial cyst 0% 0% 0% 0% None 25% 9% 50% 21% ↵a Percentage of acquired and synthetic volumes in the test set (n = 100) with each pathology (or none) as determined by the 3 radiologists during the blinded review. Rows detail pathologies available to the 3 radiologist reviewers (including none). Columns detail the percentage of the 100 patients in the test set with each pathology according to 2/3 reviewers or 3/3 reviewers.
Pathologies PPV Sensitivity NPV Specificity Ratio % Ratio % Ratio % Ratio % Bone marrow lesion 2/2 100% 2/3 66.6% 97/98 98.9% 97/97 100% Spinal cord signal change 0/0 – 0/0 – 100/100 100% 100/100 100% Paraspinous muscle edema 9/9 100% 9/24 37.5% 76/91 83.5% 76/76 100% Facet arthritis with synovitis 3/4 75% 3/13 23% 86/96 89.5% 86/87 98.8% Fracture with edema 0/0 – 0/2 0% 98/100 98% 98/98 100% Modic type 1 change 18/19 94.7% 18/41 43.9% 58/81 71.6% 58/59 98.3% Synovial cyst 0/0 – 0/0 – 100/100 100% 100/100 100% None 30/50 60% 30/34 88.2% 46/50 92% 46/66 69.6% ↵a Predictive performance statistics of pathologies from synthetic STIR volumes versus acquired STIR volumes in the test set of 100 patients. Three radiologists each reviewed in a blinded fashion the synthetic and acquired STIR volumes and evaluated the pathologies present in each of them. The rows detail the pathologies available to the 3 radiologist reviewers (including none). Their evaluations of the presence of pathologies were aggregated by using majority-rule; a pathology was considered present in each acquired STIR volume if 2 or 3 reviewers identified it as present, and similarly for each synthetic volume. The columns detail predictive performance metrics (numerator, denominator, and percentage), comparing the prediction (from sSTIR) to the ground truth (from aqSTIR) for each pathology. 95% confidence intervals for the percentage values are available in Supplemental Data.
Jump to section
Related Articles
Cited By...
- No citing articles found.