Synthetic medical scans get judged by the task, not just the look: frequency consistency plus gated distillation lifts tumour segmentation and lesion detection
Synopsis
Medical image synthesis is often treated as filling in a missing modality, yet a realistic image need not help the downstream task. Across five public datasets, a training framework ties synthesis to task learning: the translator is constrained during training by image-adaptive frequency consistency on structure and detail, and the downstream network receives teacher features only when a teacher using synthetic images beats the student by a task-specific margin on labelled samples. Translation fidelity improved for CycleGAN, CUT and SynDiff unpaired translators, with gains in brain-tumour segmentation and lesion centre-point detection, and the training-time components are removed at inference.
Interpretation
Synthetic modalities are treated as task-oriented complements to acquired scans rather than independent measurements, with translation quality and downstream segmentation and detection gains examined inside one training pipeline. Earlier frequency constraints used a fixed band split that did not adapt to individual images, and neither image realism nor feature transfer alone established that a synthetic image helps every training sample. On the 219-case official BraTS 2021 validation cohort, T1-to-T1ce and T2-to-FLAIR translation raised mean SSIM and PSNR and lowered mean LPIPS relative to the unmodified CycleGAN, CUT and SynDiff backbones and their FDIT variants; pointwise 95% CIs excluded zero in the direction favouring the new method, and comparisons survived Holm correction.
Frequency consistency is enforced by making translations of low- and high-frequency inputs agree with the corresponding bands of the full-image translation, with band boundaries predicted per image. The adaptive decomposition module AFD regresses image-specific low and high cutoffs from spatial and spectral representations through cross-attention, replacing a fixed threshold. On bidirectional PD–T2 translation in IXI, paired Wilcoxon tests over 88 cases showed AFD over the fixed-threshold version raised mean SSIM by 0.0137 and PSNR by 0.719 dB for PD to T2, and by 0.0116 and 1.115 dB for T2 to PD.
Feature transfer becomes a per-sample risk comparison: the student receives teacher features only when the teacher using synthetic images has a lower task-error score than the student by a task-specific margin on a labelled training sample. Earlier knowledge distillation did not check whether the teacher actually holds an advantage on each sample, and the gate makes that check explicit while acting only during training. On the 249-case held-out BraTS 2021 tumour-containing-slice segmentation test partition, SFD raised mean Dice by 0.0331 and 0.0261 over UNet-SE-CA and the cycle- and shape-consistency method of Zhang et al.; on 663 DeepLesion test patients, patient-level sensitivity rose by 0.0282, 0.0289 and 0.0414 at 0.5, 1 and 2 false positives per image.
Perspective
The demonstrated scope is evaluation on five public de-identified datasets: BraTS 2021, SynthRAD2025, IXI and DeepLesion, plus BraTS 2024 post-treatment MRI as external validation; segmentation results cover tumour-containing slices only, and the external evaluation assessed translation fidelity, not downstream tasks on post-treatment MRI.
Frequency decomposition uses the discrete Fourier transform with isotropic Gaussian masks and does not explicitly model spatially localized or direction-specific frequency patterns; SFD currently uses a hard gate derived from supervised task risk; the 16-case segmentation ablation limits evidence separating the gate from ungated distillation; reported case-level CIs do not quantify variability across repeated training runs.