. 2022 Jul 18;49(9):6019–6054. doi: 10.1002/mp.15840

TABLE 4.

Miscellaneous approaches for cone‐beam CT (CBCT) correction

Author and year	Anatomic site	Model	Loss function	Augmentation	Preprocessing	(train/val/test)	Training configuration	Image similarity (input CBCT)	Dose similarity
Hansen et al. 2018 ⁷³	Pelvis	U‐Net	MSE	Linear combination of two random inputs (Mixup)	A priori scatter correction for target projections	15/8/7	Paired projection 2D	MAE: 46 (144). ME: −3 (138)	GPR2: 100%. GPR1: 90%. GPR2: 53%
Jiang et al. 2019 ⁷⁴	Pelvis	U‐Net	MSE		MC scatter correction for target CBCTs	15/3/2	Paired axial 2D	RMSE: 18.8 (188.4). SSIM: 0.9993 (0.9753)
Landry et al. 2019 ⁷⁵	Pelvis	U‐Net 1	MSE	Mixup	A priori scatter correction for target projections	27/7/8	Paired projection 2D	MAE: 51 (104). ME: 1 (30)	DPR2 > 99.5%. DPR1 > 98.4%. GPR2 > 99.5% GPR3 > 95%. GPR2 > 85%. DPR3 > 75%. DPR2 > 68%
		U‐Net 2		Random left‐right flips Random position shifts Random HU shifts	Intra‐patient DIR CT resampled to CBCT Voxels outside body set to −1000 HU CT cropped to CBCT cylindrical FOV CT and CBCT cropped to remove conical ends of CBCT		Paired axial 2D	MAE: 88 (104). ME: 2 (30)	DPR2 > 99.5%. DPR1 > 98.4%. GPR2 > 99.5% GPR3 > 97%. GPR2 > 89%. DPR3 > 81%. DPR2 > 76%
		U‐Net 3			Voxels outside body set to −1000 HU		Paired axial 2D	MAE: 58 (104). ME: 3 (30)	DPR2 > 99.5%. DPR1 > 98.4%. GPR2 > 99.5%. GPR3 > 98%. GPR2 > 91%. DPR3 > 85%. DPR2 > 79%
Nomura et al. 2019 ⁷⁹	HN	U‐Net	MAE	Random left‐right flips Random 90° rotations	MC simulation of training, validation and testing data Voxels outside body set to −1000 HU Anatomy segmented into air, adipose, soft tissue, muscle, rib bone	Training: 5 phantoms. Validation: HN phantom. Testing: 1 HN, 1 Thorax patient	Paired projections 2D	MAE: 17.9 ± 5.7 (21.8 ± 5.9). SSIM: 0.9997 ± 0.0003 (0.9995 ± 0.0003). PSNR: 37.2 ± 2.6 (35.6 ± 2.3)
Nomura et al. 2019 ⁷⁹	Thorax	U‐Net	MAE	Random left‐right flips Random 90° rotations			Paired projections 2D	MAE: 29.0 ± 2.5 (32.5 ± 3.2). SSIM: 0.9993 ± 0.0002 (0.9990 ± 0.0003). PSNR: 31.7 ± 0.8 (30.6 ± 0.9)
Lalonde et al. 2020 ⁸⁰	HN	U‐Net	MAPE	Vertical and horizontal flips	Projections downsampled to 256 × 256 Projection intensities normalized against flood field	29/9/10	Paired projections 2D	MAE: 13.41 (69.64). ME: −0.801 (−28.61)	GPR2: 98.99% (68.44%)
Rusanov et al. 2021 ⁸¹	HN	U‐Net	MAE	Random vertical/horizontal flips	Bowtie filter removal via projection normalization using flood field scan	4/0/2	Paired projection 2D	MAE: 74 (318) ^a . SSIM: 0.812 (0.750)

Note: GPR3 = 3%/3 mm; GPR2 = 2%/2 mm; GPR1 = 1%/1‐mm criteria. DPR2 = 2% DD threshold; DPR1 = 1% DD threshold.

Abbreviations: DIR, deformable image registration; HN, head and neck; MAE, mean absolute error; MC, Monte Carlo; ME, mean error; MSE, mean squared error; PSNR, peak signal‐to‐noise ratio; RMSE, root mean square error; SSIM, structural similarity.

^{^a}

Image similarity metrics computed within body contour.