Anatomy-Aware Frequency Loss in 3D MRI

• The paper introduces AFL as an attention-guided loss that selectively targets diagnostically critical high-frequency structures to overcome over-smoothing in latent MRI reconstructions.
• AFL extracts high-frequency residuals and weights them using a joint map from anatomical gradients and semantic attention, ensuring precise preservation of fine details.
• Empirical results on BraTS and IXI datasets demonstrate improved metrics and visual fidelity, highlighting AFL’s effectiveness in refining 3D MRI latent representations.

Searching arXiv for the named AFL paper and closely related frequency-aware methods to ground the article in current arXiv records. Anatomy-aware Frequency Loss (AFL) is an attention-guided high-frequency reconstruction loss introduced for 3D MRI latent modeling in “Recover Semantics First, Generate Better: Improved Latent Modeling for 3D MRI Reconstruction and Cross-Contrast Synthesis” (Chen et al., 16 Jun 2026). In that framework, AFL addresses a specific failure mode of latent reconstruction pipelines: pixel-wise objectives favor low-frequency accuracy, producing over-smoothed volumes and suppressing fine anatomical boundaries, small lesions, and other diagnostically important high-frequency structures. AFL therefore does not enforce uniform spectral fidelity; instead, it extracts a high-frequency residual and weights it by joint anatomical and semantic importance so that sharp detail is preserved where it is most clinically meaningful (Chen et al., 16 Jun 2026).

1. Definition and terminological scope

In the source paper, AFL is one of three cooperating components in a semantics-first latent modeling framework for 3D MRI reconstruction and cross-contrast synthesis: the Latent Harmonization Encoder (LHE), the Semantic Recovery Block (SRB), and the Anatomy-aware Frequency Loss (AFL) (Chen et al., 16 Jun 2026). LHE is used for long-range anatomical coherence, SRB for semantic recovery and contrast-aware separability, and AFL for preserving diagnostically relevant high-frequency structure. AFL is therefore not a standalone denoising or synthesis loss; it is the part of the framework that specifically counters over-smoothing during latent compression.

The term should be distinguished from several adjacent but non-identical usages of “frequency-aware” in the literature. In fluid reconstruction, “Frequency-Aware Reconstruction of Fluid Simulations with Generative Networks” introduces a Fourier-based frequency-aware loss with a shift-towards-low (STL) weighting heuristic, but that work is explicitly about fluid simulation velocity fields rather than anatomy-aware medical imaging (Biland et al., 2019). In PET denoising, “AnF-DiffPET: Anatomy- and Frequency-Guided Diffusion for PET/CT Denoising” uses Frequency-Contrastive Hard Mining (FCHM) as the frequency-domain supervisory term and Anatomical-Frequency Guidance (AFG) for anatomy-aware conditioning; it does not use the exact term AFL (Liu et al., 1 Jul 2026). In dialogue generation, “Improving Neural Response Diversity with Frequency-Aware Cross-Entropy Loss” employs “frequency-aware” to mean token-frequency reweighting rather than spatial or anatomical frequency preservation (Jiang et al., 2019).

2. Motivation in 3D MRI latent modeling

The motivating claim is that standard reconstruction losses such as MSE emphasize average intensity matching and therefore bias optimization toward low-frequency components, which can obscure “clinically vital fine-grained details, such as subtle boundaries or small lesions” (Chen et al., 16 Jun 2026). In a latent-compression pipeline, this is particularly consequential because once the compressor has blurred away high-frequency information, the downstream cross-contrast generator inherits a less informative representation.

Within the broader “recover semantics first, generate better” philosophy, AFL is presented as the remedy for the third major challenge in 3D MRI latent modeling: over-smoothing from pixel-wise objectives. The paper’s framing is specific. LHE addresses under-preserved long-range anatomical coherence; SRB mitigates semantic degradation by injecting high-level priors from a self-supervised semantic teacher; AFL adaptively preserves diagnostically relevant high-frequency structures (Chen et al., 16 Jun 2026). This suggests that the loss is intended not merely to sharpen outputs, but to preserve detail in a way that remains consistent with both anatomy and semantic tissue relevance.

The paper further states that AFL preserves high-frequency content selectively rather than uniformly. That selectivity is central to its rationale. A uniform high-frequency penalty could also amplify noise or background texture, whereas AFL is designed to focus preservation pressure on regions that are simultaneously edge-relevant and semantically meaningful.

3. Mathematical construction

AFL is defined through three steps: construction of an anatomy-aware joint attention map, extraction of a high-frequency residual, and computation of an attention-weighted residual loss (Chen et al., 16 Jun 2026).

The joint attention weight is

$A_j = A_a \odot A_s,$

where $\odot$ denotes element-wise multiplication. The anatomical attention term is derived from volumetric spatial gradients:

$A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$

The paper describes this as a local edge detector that highlights morphological discontinuities, sharp boundaries, and subtle textures. The semantic attention map $A_s$ is obtained by aggregating and upsampling the final-layer self-attention responses from the pre-trained DINO teacher network used in the SRB. Although no separate closed-form equation is given for $A_s$, it is described as a global semantic filter that emphasizes clinically meaningful tissues and suppresses irrelevant background regions.

The high-frequency component is defined by subtracting a smoothed version of the volume:

$\mathcal{H}(x) = x - \mathcal{S}_v(x),$

where $\mathcal{S}_v(\cdot)$ is a local volumetric smoothing operator. The paper does not specify the exact kernel size or filter type in the provided text.

The final AFL objective is

$$\mathcal{L}_{\text{afl}} = \left\| A_j \odot \mathcal{H}(\hat{x}) - A_j \odot \mathcal{H}(x) \right\|_1.$$

Equivalently, AFL computes an $L_1$ distance between the attention-weighted high-frequency residuals of the reconstructed volume $\hat{x}$ and the ground-truth volume $\odot$0 (Chen et al., 16 Jun 2026).

Two features of this construction are technically important. First, AFL is not defined in terms of a full Fourier transform; it operates in a frequency-like residual space built from local smoothing subtraction. Second, the weighting is multiplicative and joint: a region receives strong preservation pressure only when it is salient under both the gradient-based anatomical prior and the teacher-derived semantic prior. The paper characterizes this combined effect as adaptive preservation of diagnostically relevant high-frequency structures.

4. Anatomical and semantic priors

The anatomy-aware character of AFL comes from the interaction of two complementary priors rather than from frequency analysis alone (Chen et al., 16 Jun 2026).

The spatial prior is encoded in $\odot$1, which is computed from first-order partial derivatives along the three spatial axes. Because

$\odot$2

the weighting increases at edges, boundaries, and rapid local intensity transitions. In the paper’s description, these are the loci where important anatomical detail typically resides.

The semantic prior is encoded in $\odot$3, derived from the final-layer self-attention responses of a frozen DINO-pretrained ViT-B/16 teacher used in the SRB. The attention responses are aggregated and upsampled to form a global importance signal. The role of this signal is to emphasize structurally and clinically meaningful tissue while suppressing irrelevant background regions (Chen et al., 16 Jun 2026).

By multiplying the two maps,

$\odot$4

AFL requires both local edge relevance and global semantic relevance for a voxel region to exert strong influence on the loss. The paper’s own interpretation is that AFL is neither purely spectral nor purely edge-based; it is best understood as a frequency-domain residual penalty guided by spatial gradients and teacher-derived semantic attention. A plausible implication is that this gating structure is intended to reduce the risk of indiscriminate sharpening.

5. Integration into the optimization pipeline

AFL is incorporated into the total objective used during the first training stage, 3D MRI reconstruction:

$\odot$5

where $\odot$6 is the baseline reconstruction loss used in 3D FSQ, and

$\odot$7

aligns student latent features with the DINO teacher (Chen et al., 16 Jun 2026).

The training procedure is explicitly two-stage. Stage 1 performs 3D MRI reconstruction and learns compact latent representations using $\odot$8. Stage 2 performs cross-contrast synthesis, where the compression modules are frozen and a latent translation model is trained with an $\odot$9 translation loss. AFL therefore belongs to the latent compression and reconstruction stage, not to the downstream latent translation loss.

The framework is trained on the BraTS and IXI 3D MRI datasets. Volumes are resampled to a uniform spatial resolution. For the reconstruction stage, the paper reports the use of Adam with learning rate $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$0, batch size $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$1, $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$2 epochs, and $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$3 RTX4090 GPUs (Chen et al., 16 Jun 2026). The semantic attention component uses the DINO ViT-B/16 teacher, which the paper reports as superior to ResNet-50 and MAE-based alternatives in ablation. No explicit AFL-specific weighting coefficient is reported in the provided text; the total objective is written as a plain sum of the three terms.

6. Empirical findings and reported effects

The ablation study on BraTS reports the following progression for reconstruction performance: baseline FSQ yields Avg PSNR $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$4, SSIM $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$5, and LPIPS $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$6; adding LHE yields Avg PSNR $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$7, SSIM $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$8, and LPIPS $A_a = |\nabla_x x| + |\nabla_y x| + |\nabla_z x|.$9; adding LHE and SRB yields Avg PSNR $A_s$0, SSIM $A_s$1, and LPIPS $A_s$2; adding LHE, SRB, and AFL yields Avg PSNR $A_s$3, SSIM $A_s$4, and LPIPS $A_s$5 (Chen et al., 16 Jun 2026). The paper explicitly states that after introducing AFL, the model reaches the best reconstruction performance overall.

The qualitative claims are aligned with the loss design. In reconstruction visual comparisons, the model is reported to better preserve sharp tissue boundaries, structurally consistent anatomy, and subtle fine details after AFL is included. In synthesis results, the improved latent representation yields more faithful generated contrasts. This is presented as indirect evidence that AFL benefits downstream cross-contrast synthesis by making the compressed representation more detailed and less over-smoothed.

The paper positions AFL primarily against standard pixel-wise losses, especially MSE. The conceptual comparison is that pixel-space losses optimize overall intensity similarity but emphasize low frequencies and can oversmooth anatomy, whereas AFL targets high-frequency components and restricts them to anatomically and semantically important regions (Chen et al., 16 Jun 2026). No separate quantitative comparison against a generic frequency loss such as Focal Frequency Loss is reported in the provided text.

7. Relation to earlier and adjacent frequency-aware methods

AFL belongs to a broader family of losses motivated by the observation that uniform training objectives often privilege dominant low-frequency or high-frequency-occurrence content. The closest antecedent in reconstruction learning among the cited papers is the frequency-aware fluid reconstruction loss of “Frequency-Aware Reconstruction of Fluid Simulations with Generative Networks,” which supervises band-specific Fourier coefficients and applies STL weighting to emphasize lower and mid bands rather than very high frequencies (Biland et al., 2019). That work shares the premise that ordinary spatial-domain $A_s$6 supervision is not frequency-discriminative, but it differs materially from AFL in domain, supervision space, and weighting mechanism. Its target is fluid simulation velocity fields; its weighting is heuristic over Fourier bands; and it is not anatomy-aware.

AFL is also distinct from the PET denoising supervision used in “AnF-DiffPET: Anatomy- and Frequency-Guided Diffusion for PET/CT Denoising” (Liu et al., 1 Jul 2026). There, the explicit frequency-domain loss is Frequency-Contrastive Hard Mining:

$A_s$7

with the frequency term acting on FFT magnitude spectra of predicted high-dose PET, ground-truth high-dose PET, and low-dose PET negatives. The anatomy-aware aspect in that framework is indirect, arising from CT-conditioned Anatomical-Frequency Guidance and Multi-Scale Cross-Transformer Reconstruction rather than from an AFL-style joint attention map. This comparison underscores that “anatomy-aware frequency loss” is not a generic label for all anatomy-guided spectral supervision.

Finally, the NLP loss “Frequency-Aware Cross-Entropy” in dialogue generation is conceptually unrelated to medical image frequency preservation, even though it uses similar vocabulary (Jiang et al., 2019). In that paper, “frequency-aware” refers to token-frequency reweighting to reduce over-confidence in common words. The shared terminology reflects a general optimization principle—rebalancing dominant modes—but not a shared mathematical construction.

Taken together, these neighboring works suggest a common pattern: plain objectives often overfit dominant components, while frequency-aware reweighting can redirect learning toward underrepresented but task-relevant structure. In AFL, that redirection is made anatomy- and semantics-selective through

$A_s$8

which is the method’s defining feature (Chen et al., 16 Jun 2026).