Frequency-consistent synthesis and selective transfer for medical image analysis
Authors: Zhuhui Wang, Jianwei Zuo, Xin Peng, Kai Zhong, Xuliang Deng, Jiajia Luo
Organizations: Institute of Medical Technology, Peking University Health Science Center, Beijing, 100191, China. · Biomedical Engineering Department, Institute of Advanced Clinical Medicine, Peking University, Beijing, 100191, China. · Department of Oral and Maxillofacial Surgery, Peking University School and Hospital of Stomatology, Beijing, 100081, China. · National Clinical Research Center for Oral Diseases, National Engineering Laboratory for Digital and Material Technology of Stomatology, Beijing Key Laboratory of Digital Stomatology, Beijing, 100081, China. · National Engineering Research Center of Oral Biomaterials and Digital Medical Devices, NMPA Key Laboratory for Dental Materials, Beijing Laboratory of Biomedical Materials, Peking University School and Hospital of Stomatology, Beijing, 100081, China. · Department of Geriatric Dentistry, Peking University School and Hospital of Stomatology, Beijing, 100081, China.
Complementary imaging is not always available, and realistic synthesis does not necessarily improve downstream analysis. Here we introduce a training framework linking medical image synthesis to task learning through adaptive frequency consistency and selective feature transfer. Image-specific frequency constraints encourage consistency of structure and detail. Features are transferred only when a teacher using synthetic images achieves a lower task-error score than the student by a task-specific margin on a labelled training sample. Evaluations across five datasets show improved translation fidelity with three adversarial and diffusion-based translators trained without paired supervision, alongside gains in brain-tumour segmentation on tumour-containing slices and lesion centre-point detection. All 105 reported paired comparisons yielded point estimates favouring the proposed methods, including 36 external translation comparisons on post?treatment MRI without retraining. These findings support synthetic modalities as task-oriented complements to acquired scans when complementary imaging is unavailable.
Figures & tables
Figure 1: Frequency-consistent image translation and selective feature distillation. (a) During FDDT training, the shared base translator processes full-, low-, and high-frequency inputs. A single shared AFD module predicts sample-specific cutoffs for the source and generated images, and frequency-consistency losses align translations of the decomposed inputs with the corresponding components of the full-image translation. The frequency pathways and AFD module are removed after training. (b) The generated modality is combined with the source modality for downstream segmentation and lesion centre-point detection. (c) During SFD training, feature distillation is activated only when the teacher has a lower task-specific risk than the student by the predefined margin, while a low-weight translation GAN objective remains active to regularize the jointly fine-tuned translator. At inference, the translator and student are retained, while the teacher and distillation branches are removed.
Figure 2: Cross-modal image-translation performance. (a) Representative T1 → T1ce and T2 → FLAIR translations and case-level SSIM, PSNR and LPIPS results for the 219-case BraTS 2021 official validation cohort, used here exclusively for final evaluation. (b) Representative CT → MR translations, case-level SSIM and PSNR results, and dataset-level FID for the independently held-out 103-case SynthRAD2025 test partition. GT denotes the paired target-modality image, used to evaluate the corresponding held-out source image; source–target correspondence was not provided to the training objective. For the case-level metrics in (a) , bars show cohort means and error bars indicate standard deviations (SDs) across cases. In (b) , SSIM and PSNR bars show case-level cohort means without error bars, while FID is reported as a dataset-level value. Raw and Holm-adjusted results are reported in Supplementary Tables S1 and S2 .
Figure 3: Adaptive frequency decomposition. (a) Pixelwise error maps of the high-frequency (left) and low-frequency (right) components for bidirectional T1 ↔ T1ce translation on the held-out 219-case official BraTS 2021 validation cohort (T1ce → T1, top; T1 → T1ce, bottom). Each row shows a representative axial slice from a different case. For CycleGAN, CycleGAN+FDIT and CycleGAN+FDDT, maps show the absolute difference between the corresponding frequency component of the synthesized image and that of the ground truth, on a shared colour scale (dark, low error; bright, high error) common to all panels. GT columns show the ground-truth frequency component in grayscale for reference. (b) SSIM, PSNR and FID for the base CycleGAN translator, fixed-threshold FDDT and FDDT with AFD on the 88-case IXI test cohort for T2 → PD and PD → T2 translation. Bars show cohort means for SSIM and PSNR, which were calculated at the case level; FID was calculated at the dataset level. (c) Distributions of the sample-specific high- and low-frequency cutoffs predicted for the IXI T2 and PD images. Case-level SSIM and PSNR were compared using paired Wilcoxon signed-rank tests with Holm correction across the eight comparisons. Supplementary Table S3 reports mean paired differences to align with the bars and separately reports median paired differences with 95% percentile-bootstrap CIs obtained using 5,000 resamples. FID was reported descriptively.
Figure 4: Selective feature distillation improves tumour-containing-slice segmentation and lesion centre-point detection. (a) Representative input images, predicted whole-tumour binary masks, ground-truth masks and T1ce MRI, together with case-level precision, recall, Dice and IoU on the 249-case held-out BraTS 2021 test partition. Predicted probabilities were thresholded at 0.5, and the metrics were calculated in three dimensions over the selected tumour-containing slices of each case. (b) Comparison of ungated full feature distillation (FD) and selective feature distillation (SFD) using the independently trained 80/16/16 ablation split. (c) Ablation of the base model and its combinations with SFD, FDIT and FDDT on the same 16-case test partition. (d) Representative student (S) and teacher (T) feature maps during SFD training, shown alongside the ground-truth mask. (e) Representative DeepLesion CT images and detection heatmaps, together with patient-wise macro-averaged sensitivity at FP budgets of 0.5, 1 and 2 false positives per image, KL divergence and conditional mean localization error for CenterNet with and without SFD. Sensitivity and KL divergence were evaluated in all 663 test patients; localization error was evaluated in the 528 patients with valid matched detections for both methods. Raw and Holm-adjusted results for panels (a) and (e) are reported in Supplementary Tables S5 and S6 , respectively. Bars show case-level means in (a–c) and patient-level means in (e).
Figure 5: External validation of image translation on BraTS 2024. BraTS 2021-trained translators were evaluated without retraining on 87 patients from BraTS 2024, comprising 187 longitudinal post-treatment MRI examinations. (a) Representative T1 → T1ce translations and patient-level SSIM, PSNR and LPIPS. (b) Representative T2 → FLAIR translations and the corresponding patient-level metrics. For patients with multiple examinations, each metric was first calculated per examination and then averaged within patient, such that each patient contributed one value to the cohort-level summary and statistical analysis. GT denotes the paired target-modality image used only for evaluation. Bars show means across the 87 patients, and error bars show standard deviations (SDs) across patients. Raw and Holm-adjusted results are reported in Supplementary Table S4 .
Task
Comparison
Metric
n
Mean Δ
95% CI of Δ
Raw p
Holm-adjusted p
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN
SSIM
219
0.02088
[0.01812,0.02363]
p<0.001
p<0.001
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN
PSNR
219
1.40474
[0.91617,1.89331]
p<0.001
p<0.001
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN
LPIPS
219
−0.01665
[−0.01909,−0.01421]
p<0.001
p<0.001
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN+FDIT
SSIM
219
0.03591
[0.03262,0.03921]
p<0.001
p<0.001
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN+FDIT
PSNR
219
2.43778
[1.88086,2.99470]
p<0.001
p<0.001
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN+FDIT
LPIPS
219
−0.02240
[−0.02506,−0.01973]
p<0.001
p<0.001
Supplementary Table S1 : Case-level paired analyses for the held-out 219-case official BraTS 2021 validation cohort used exclusively for final evaluation shown in Fig. 2 a. Each case contributed one value per metric. Paired t -tests were used. The paired mean difference was defined as Δ=FDDTvariant−reference ; positive values favor FDDT for SSIM and PSNR, whereas negative values favor FDDT for LPIPS. The 95% CIs are pointwise CIs for the mean paired differences. All comparisons belong to family S1 ( m=36 ).
Task
Comparison
Metric
n
Mean Δ
95% CI of Δ
Raw p
Holm-adjusted p
CT → MR
CycleGAN+FDDT vs. CycleGAN
SSIM
103
0.01745
[0.01522,0.01968]
p<0.001
p<0.001
CT → MR
CycleGAN+FDDT vs. CycleGAN
PSNR
103
0.42025
[0.30006,0.54044]
p<0.001
p<0.001
CT → MR
CycleGAN+FDDT vs. CycleGAN+FDIT
SSIM
103
0.02103
[0.01768,0.02438]
p<0.001
p<0.001
CT → MR
CycleGAN+FDDT vs. CycleGAN+FDIT
PSNR
103
1.32802
[1.17242,1.48361]
p<0.001
p<0.001
CT → MR
CUT+FDDT vs. CUT
SSIM
103
0.01489
[0.01218,0.01760]
p<0.001
p<0.001
CT → MR
CUT+FDDT vs. CUT
PSNR
103
0.45838
[0.22098,0.69579]
p<0.001
p<0.001
Supplementary Table S2 : Case-level paired analyses for the 103-case held-out SynthRAD2025 CT → MR test partition shown in Fig. 2 b. Each case contributed one value per metric. Paired t -tests were used. The paired mean difference was defined as Δ=FDDTvariant−reference ; positive values favor FDDT. All comparisons belong to family S2 ( m=12 ).
Task
Comparison
Metric
n
Mean Δ
Median Δ (95% CI)
Raw p
Holm-adjusted p
PD → T2
CycleGAN+FDDT+AFD vs. CycleGAN
SSIM
88
0.025449
0.028984[0.012512,0.042527]
p<0.001
p<0.01
PD → T2
CycleGAN+FDDT+AFD vs. CycleGAN
PSNR
88
1.287684
1.331045[0.425650,1.872452]
p<0.001
p<0.01
PD → T2
CycleGAN+FDDT+AFD vs. CycleGAN+FDDT
SSIM
88
0.013659
0.018983[0.015480,0.022335]
p<0.001
p<0.01
PD → T2
CycleGAN+FDDT+AFD vs. CycleGAN+FDDT
PSNR
88
0.718979
0.552907[0.354597,1.004461]
p<0.001
p<0.01
T2 → PD
CycleGAN+FDDT+AFD vs. CycleGAN
SSIM
88
0.007752
0.006950[0.004337,0.009257]
p<0.001
p<0.01
T2 → PD
CycleGAN+FDDT+AFD vs. CycleGAN
PSNR
88
0.281004
0.299682[0.161102,0.340695]
p<0.001
p<0.01
Supplementary Table S3 : Case-level paired analyses for the IXI AFD ablation experiments shown in Fig. 3 b. Each of the 88 cases contributed one value per metric. The mean paired difference, Δ=AFD-enabledvariant−reference , is reported to align the effect estimate with the cohort-mean bars. Paired Wilcoxon signed-rank tests were used; the median paired difference and its 95% percentile-bootstrap CI, calculated using 5,000 case-level paired resamples with replacement (seed 42), are reported separately as the nonparametric effect summary. Raw p -values are from the Wilcoxon tests; the final column reports the corresponding Holm-adjusted p -values. Positive values favor the AFD-enabled variant. All comparisons belong to family S3 ( m=8 ).
Task
Comparison
Metric
n
Mean Δ
95% CI of Δ
Raw p
Holm-adjusted p
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN
SSIM
87
0.00791
[0.00633,0.00949]
p<0.001
p<0.001
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN
PSNR
87
0.31588
[0.15110,0.48066]
p<0.001
p<0.01
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN
LPIPS
87
−0.01019
[−0.01259,−0.00779]
p<0.001
p<0.001
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN+FDIT
SSIM
87
0.00679
[0.00431,0.00927]
p<0.001
p<0.001
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN+FDIT
PSNR
87
0.43504
[0.19955,0.67053]
p<0.001
p<0.01
T1 → T1ce
CycleGAN+FDDT vs. CycleGAN+FDIT
LPIPS
87
−0.00588
[−0.00774,−0.00402]
p<0.001
p<0.001
Supplementary Table S4 : Patient-level paired analyses of SSIM, PSNR, and LPIPS for external validation of the BraTS 2021-trained translators, evaluated without retraining in 87 BraTS 2024 patients comprising 187 longitudinal MRI examinations (Fig. 5 ). Examination-level metrics were averaged within patient across available time points. Paired t -tests were used. The paired mean difference was defined as Δ=FDDTvariant−reference ; positive values favor FDDT for SSIM and PSNR, whereas negative values favor FDDT for LPIPS. The 95% CIs are pointwise CIs for the mean paired differences. All comparisons belong to family S4 ( m=36 ). All 36 mean differences favour FDDT; 35 pointwise CIs exclude zero and 30 comparisons remain significant after Holm correction.
Comparison
Metric
n
Mean Δ
95% CI of Δ
Raw p
Holm-adjusted p
SFD vs. UNet-SE-CA
Precision
249
0.05722
[0.04784,0.06659]
p<0.001
p<0.001
SFD vs. UNet-SE-CA
Recall
249
0.00841
[0.00002,0.01680]
p<0.05
p<0.05
SFD vs. UNet-SE-CA
Dice
249
0.03315
[0.02672,0.03957]
p<0.001
p<0.001
SFD vs. UNet-SE-CA
IoU
249
0.04168
[0.03485,0.04850]
p<0.001
p<0.001
SFD vs. Zhang et al.
Precision
249
0.02356
[0.01536,0.03177]
p<0.001
p<0.001
SFD vs. Zhang et al.
Recall
249
0.02683
[0.01955,0.03411]
p<0.001
p<0.001
Supplementary Table S5 : Case-level paired analyses for the 249-case held-out BraTS 2021 tumour-containing-slice segmentation test cohort shown in Fig. 4 a. Each case contributed one value per metric. Paired t -tests were used. The paired mean difference was defined as Δ=SFD−reference ; positive values favor SFD. The 95% CIs are pointwise CIs for the mean paired differences. All comparisons belong to family S5 ( m=8 ). This table concerns the main 249-case segmentation comparison, not the separately trained 16-case FD-versus-SFD ablation.
Comparison
Metric
n
Mean Δ
95% CI of Δ
Raw p
Holm-adjusted p
CenterNet+SFD vs. CenterNet
Sensitivity at 0.5 FP/image
663
0.02817
[0.00416,0.05217]
p<0.05
p<0.05
CenterNet+SFD vs. CenterNet
Sensitivity at 1 FP/image
663
0.02893
[0.00474,0.05312]
p<0.05
p<0.05
CenterNet+SFD vs. CenterNet
Sensitivity at 2 FP/image
663
0.04142
[0.01925,0.06359]
p<0.001
p<0.01
CenterNet+SFD vs. CenterNet
KL divergence
663
−0.43922
[−0.55769,−0.32076]
p<0.001
p<0.001
CenterNet+SFD vs. CenterNet
Conditional mean localization error (px)
528
−0.21628
[−0.36810,−0.06446]
p<0.01
p<0.05
Supplementary Table S6 : Patient-level paired analyses for DeepLesion centre-point detection shown in Fig. 4 e. For each patient, sensitivity was calculated from confidence-ranked candidate peaks at false-positive budgets of 0.5, 1, and 2 FP per image, and the resulting patient-level values were used for paired analysis. Paired t -tests were used. The paired mean difference was defined as Δ=CenterNet+SFD−CenterNet . Positive values favor SFD for sensitivity, whereas negative values favor SFD for KL divergence and localization error. Sensitivity and KL-divergence analyses included all 663 test patients. Conditional mean localization error was evaluated in the 528 patients with valid matched detections for both methods; undetected lesions were excluded. The 95% CIs are pointwise CIs for the mean paired differences. All comparisons belong to family S6 ( m=5 ).
UFR SFA · University of Poitiers, CNRS, XLIM, France · I3M common laboratory CNRS-Siemens Healthinners, Poitiers University Hospital and University of Poitiers, France +4
Department of Diagnostics and Intervention, Radiation Physics, Biomedical Engineering, Umeå University, Universitetstorget, 4, Umeå, 90187, Sweden · Unit of Artificial Intelligence and Computer Systems, Department of Engineering, Università Campus Bio-Medico di Roma, Via Alvaro del Portillo, 21, Roma, 00128, Italy · Università Cattolica del Sacro Cuore, Largo Francesco Vito, 1, Roma, 00168, Italy +12