L0-Norm-Based CycleGAN for Unpaired Brain CT-to-MR Synthesis

1 Introduction

• CT and MR provide complementary brain information, but cost, equipment access, and MR contraindications can leave patients with only CT.Synthetic MR could reduce the need for another MR acquisition when acceptable for the intended task.Source: §1 · p. 2
• CycleGAN enables unpaired synthesis when aligned CT–MR pairs are scarce because the scans come from different sessions, scanners, or fields of view.Its two generators use cycle consistency to encourage approximately invertible translations on the data support.Source: §1 · p. 2
• The study replaces the L1-norm cycle penalty with a smooth approximation of the L0-norm, increasing sensitivity to small residuals while suppressing the influence of large ones.Its redescending influence function has a bounded per-residual gradient that rises and then declines as residuals increase.Source: §1 · pp. 2–3
• On 9 held-out test images, the proposed model had a lower maximum MSE and higher minimum PSNR and SSIM, with aggregate performance and each metric’s extreme-case value reported.The evaluation used 102 images for training and 9 held out for testing.Source: §1 · p. 3

2.1 Problem setting

• For a brain CT image, the model learns a mapping to a synthetic MR image.The source and target domains are denoted X and Y, respectively.Source: §2.1 · p. 3

2.2 The cycleGAN model

• CycleGAN translates between unpaired CT and MR domains with two generators and two discriminators; cycle consistency encourages approximately invertible translations.The full objective combines the adversarial losses with the cycle-consistency term, weighted by λ.Source: §2.2 · pp. 3–4

2.3 L0-norm cycle-consistency penalty

• The L0 cycle penalty counts nonzero image residuals, making every nonzero residual contribute equally rather than weighting it by magnitude.The paper notes that this quantity is a cardinality function, though it follows the common convention of calling it an L0-norm.Source: §2.3 · pp. 4–5

2.4 Smooth approximation and gradient properties

• The proposed cycle-consistency term uses a smooth, bounded L0 approximation whose gradient peaks near residual magnitude σ and declines for larger residuals.The gradient is bounded by 0.787/σ and is most sensitive near |Iij| ≈ 1.02σ.Source: §2.4 · p. 5

2.5 Least-squares adversarial loss

• The method uses least-squares adversarial loss as the generator-discriminator constraint, which the paper says can alleviate vanishing gradients and improve training stability.The stated properties are attributed to the loss formulation.Source: §2.5 · p. 6

2.6 Full objective and optimisation

• The final objective combines the loss terms in a CycleGAN residual architecture, and the L0 model and L1 baseline share the training configuration except for the cycle penalty.The generators use nine residual blocks; the discriminators are 70 × 70 PatchGANs, and training uses λ = 10 and σ = 0.1.Source: §2.6 · pp. 6–7, Figure 1 · p. 7

3.1 Data and split

• The image-level split selected 111 paired images with similar structures, using 102 for training and 9 for testing.The dataset contains CT and MR images cropped into 256 × 256 images; white curvilinear artifacts in the original CT images were manually removed.Source: §3.1 · pp. 7–8

3.2 Evaluation metrics

• MSE, PSNR, and SSIM evaluate pixel-level agreement; lower MSE, higher PSNR, and SSIM closer to one indicate closer agreement with the reference.SSIM compares luminance, contrast, and structure.Source: §3.2 · pp. 8–9

3.3 Quantitative results

• Mean MSE was 20.7% lower and mean SSIM was higher by 0.0282 for the proposed model than the L1-norm baseline on the test set.The comparison is reported in Table 1.Source: §3.3 · p. 9, Table 1 · p. 9

3.4 Extreme-case performance

• The proposed model lowered maximum MSE from 0.04770 to 0.04048, raised minimum PSNR from 13.22 to 13.93 dB, and raised minimum SSIM from 0.3599 to 0.3776.The baseline retained a higher maximum SSIM: 0.6850 versus 0.6419.Source: §3.4 · p. 9

3.5 Qualitative comparison

• In two representative cases, the proposed model is described as visually closer to the reference MR, preserving sharper tissue boundaries while suppressing unnatural artifacts.Figure 2 places input CT, proposed synthetic MR, baseline synthetic MR, and reference MR in columns from left to right.Source: §3.5 · pp. 9–10, Figure 2 · p. 10

4 Discussion

• Replacing the L1-norm cycle penalty with the L0-norm approximation improved mean reconstruction accuracy and extreme-case metric values in this dataset.The models shared the CycleGAN architecture and other training components.Source: §4 · p. 10

4.1 Interpretation of the observed improvement

• The approximation's bounded gradient peaks near |Iij| ≈ 1.02σ and decays for larger residuals, reducing their relative influence while retaining sensitivity near the transition band.For σ = 0.1, its gradient exceeds the L1 penalty below approximately 0.21 residual magnitude.Source: §4.1 · p. 10

4.2 Relation to previous work and metric interpretation

• Unlike many CycleGAN improvements that retain the original cycle penalty, this work modifies the penalty itself, and the scheme can be transferred to other CycleGAN variants.MSE and PSNR measure global error, whereas SSIM compares local luminance, contrast, and structure.Source: §4.2 · p. 11

• The smooth, bounded approximation gives the L0-based cycle-consistency term a redescending influence function, with a bounded gradient that decays for much larger residuals.Source: §5 · p. 11
• 0.02746 versus 0.03464 mean MSE, 15.80 versus 14.77 dB mean PSNR, and 0.5689 versus 0.5407 mean SSIM were reported for the proposed model and comparator, respectively, on 9 test images.The models were trained on 102 images.Source: §5 · p. 11
• The proposed model also had lower maximum MSE and higher minimum PSNR and SSIM; the paper concludes that changing the penalty can improve reconstruction accuracy in this dataset.Source: §5 · p. 11
• Future work will include systematic (σ, λ) sensitivity analysis, multi-centre data, and downstream-task evaluation.Source: §5 · p. 11