AMBER-AFNO: Lightweight 3D Segmentation

• The paper demonstrates that replacing multi-head self-attention with Adaptive Fourier Neural Operators in a fully 3D architecture yields competitive segmentation performance with an over 80% reduction in trainable parameters.
• AMBER-AFNO incorporates a hierarchical Transformer encoder with 3D patch embedding and Mix-FFN blocks to efficiently model global and local features in volumetric MRI and CT datasets.
• The model achieves state-of-the-art results on the ACDC benchmark while offering a lightweight alternative for medical segmentation, though its performance on heterogeneous datasets like Synapse remains lower.

Searching arXiv for AMBER-AFNO and related AMBER/AFNO papers to ground the article. I’m going to look up the relevant arXiv records for “AMBER-AFNO”, “AMBER hyperspectral SegFormer”, and AFNO weather validation. AMBER-AFNO is a lightweight 3D medical image segmentation model that adapts the AMBER architecture—originally introduced for multiband hyperspectral segmentation—to volumetric healthcare data and replaces multi-head self-attention with Adaptive Fourier Neural Operators (AFNO). Its stated objective is to preserve global-context modeling for MRI and CT datacubes while substantially reducing trainable parameters relative to heavy transformer baselines, especially UNETR++, and to do so with a fully 3D encoder–decoder pipeline that predicts volumetric masks directly rather than reducing the representation to a 2D output (Dosi et al., 3 Aug 2025).

1. Nomenclature, lineage, and scope

AMBER-AFNO inherits its name from AMBER, short for “Advanced SegFormer for Multi-Band Image Segmentation,” a SegFormer-derived architecture for hyperspectral image segmentation. In that earlier formulation, the central modifications were 3D convolutional patch embedding, SegFormer-style hierarchical encoding, and an all-MLP decoder; AFNO was not part of the model, and the paper does not mention Fourier neural operators explicitly (Dosi et al., 2024). AMBER-AFNO is therefore not the original AMBER model, but a methodological transfer from remote sensing to 3D medical segmentation in which AFNO becomes the primary global token-mixing mechanism (Dosi et al., 3 Aug 2025).

The term “AMBER” is overloaded across multiple unrelated arXiv contexts. It denotes, among other things, a hyperspectral segmentation architecture (Dosi et al., 2024), a neutron-spectroscopy background-estimation algorithm (Lass et al., 15 Jul 2025), a semi-numerical Epoch of Reionization code (Trac et al., 2021), and the VLTI/AMBER interferometric instrument in stellar-atmosphere studies (0906.4792). In the context of AMBER-AFNO, only the hyperspectral-segmentation lineage is directly relevant.

A recurrent misconception is that AMBER-AFNO is simply “AMBER plus AFNO” in the sense of a minor substitution inside the original hyperspectral model. The published framing is broader than that. The model is described as a domain transfer from multiband remote sensing to volumetric medical segmentation, with a fully 3D architecture, AFNO-based encoder blocks, a lightweight decoder, and direct volumetric output (Dosi et al., 3 Aug 2025). Another terminological ambiguity concerns the “funnalizer”: AMBER-AFNO contrasts its direct 3D output with the original AMBER “funnalizer,” whereas the hyperspectral AMBER paper does not explicitly define a dedicated Funnelizer module. This suggests a variant-specific reinterpretation rather than a uniformly named canonical component across both papers (Dosi et al., 3 Aug 2025, Dosi et al., 2024).

2. Architectural organization

AMBER-AFNO is organized into two principal subsystems: a hierarchical Transformer encoder and a lightweight MLP decoder. The encoder uses 3D patch embedding or overlapped patch merging, AFNO blocks for global mixing, and Mix-FFN blocks for local enhancement. The decoder performs channel projection of multiscale encoder outputs, trilinear upsampling to a common resolution, concatenation and fusion by $1\times1\times1$ convolution with batch normalization and ReLU, followed by transposed 3D convolution and a final $1\times1\times1$ convolution to produce class logits (Dosi et al., 3 Aug 2025).

The encoder follows the SegFormer MiT-style hierarchy already central to AMBER. Hierarchical features $F_1, F_2, F_3, F_4$ are produced by patch merging, with progressively reduced spatial size and increased channel dimension. The paper describes multiscale outputs of the form

$$\frac{D}{2^i} \times \frac{H}{2^i} \times \frac{W}{2^i} \times C_i,\quad i\in\{0,1,2,3\}.$$

This preserves early high-resolution structure while allowing deeper stages to represent more abstract volumetric semantics (Dosi et al., 3 Aug 2025).

Overlapped patch merging is used to reduce reliance on explicit positional encoding. The reported 3D kernel–stride–padding configurations include $K=3, S=1, P=1$ and $K=3, S=2, P=2$. As in the original AMBER formulation, the design intent is to retain local detail without excessive parameter growth (Dosi et al., 3 Aug 2025, Dosi et al., 2024).

The decoder is intentionally minimal. The paper emphasizes that it avoids a dedicated spectral-reduction stage, integrates the final MLP into the transposed convolution, reduces memory cost, and remains accurate despite being simpler than the original AMBER decoder (Dosi et al., 3 Aug 2025). In structural terms, this places most of the model’s representational burden on the encoder’s AFNO-plus-Mix-FFN stack rather than on a heavy reconstruction head.

3. AFNO substitution and frequency-domain token mixing

The defining modification in AMBER-AFNO is the replacement of multi-head self-attention with AFNO. The motivation is computational: conventional self-attention scales as $O(N^2)$ in the number of tokens, and the earlier reduced-attention AMBER formulation still scales as $O(N^2/R)$, which remains costly for 3D volumes (Dosi et al., 3 Aug 2025). AFNO replaces explicit token–token interaction with spectral mixing in the Fourier domain.

The AFNO input is a 5D tensor

$$x \in \mathbb{R}^{B \times D \times H \times W \times C},$$

where $B$ is batch size and $1\times1\times1$0 is embedding dimension. The first operation is a 3D real FFT: $1\times1\times1$1 The channel dimension is then partitioned into $1\times1\times1$2 blocks, and each block is processed by a learnable complex MLP. The mixed coefficients are regularized with soft shrinkage,

$1\times1\times1$3

after which the representation is mapped back to the spatial domain by inverse FFT with a residual connection: $1\times1\times1$4 The paper’s boxed summary is that AFNO applies $1\times1\times1$5, blockwise complex MLP mixing, $1\times1\times1$6, and $1\times1\times1$7, then adds the input through a residual path (Dosi et al., 3 Aug 2025).

The conceptual claim is not merely that AFNO is cheaper than attention, but that it captures global context differently. In Fourier space, each coefficient encodes information from the entire spatial volume, so frequency-wise learned transformations operate on globally aggregated structure rather than on explicit pairwise affinities (Dosi et al., 3 Aug 2025). This is the central reason AMBER-AFNO treats AFNO as a substitute for attention rather than as an auxiliary block.

Local inductive bias is reintroduced through Mix-FFN. As in SegFormer and the original AMBER, channel mixing is paired with a $1\times1\times1$8 convolutional operation inside the feed-forward path, so the encoder divides labor between spectral global modeling and local neighborhood refinement (Dosi et al., 3 Aug 2025, Dosi et al., 2024).

4. Problem setting, datasets, and evaluation protocol

AMBER-AFNO is evaluated on two standard 3D medical segmentation benchmarks: ACDC and Synapse. ACDC contains 200 labeled cardiac MRI volumes, with 160 used for training and 40 for testing, and three foreground classes: right ventricle, myocardium, and left ventricle. The images are resampled to $1\times1\times1$9, with depth varying roughly from 90 to 130 slices. Synapse contains 30 abdominal CT scans, with 18 training and 12 testing cases, and eight organ labels: spleen, right kidney, left kidney, gallbladder, liver, stomach, aorta, and pancreas. The spatial resolution is $F_1, F_2, F_3, F_4$0, with depth roughly 75 to 250 slices (Dosi et al., 3 Aug 2025).

The principal metrics are Dice Similarity Coefficient (DSC) and the 95th percentile Hausdorff Distance (HD95). For ACDC, the main reported metric is DSC; for Synapse, both DSC and HD95 are reported (Dosi et al., 3 Aug 2025). The training setup uses deep supervision with a hybrid Dice-plus-cross-entropy loss, and preprocessing is stated to follow the same standard pipeline used by UNETR++, UNETR, DS-UNETR++, and SAM-3D, including resampling, intensity normalization, Z-score normalization, cropping, and padding (Dosi et al., 3 Aug 2025).

The experiments were run on a single NVIDIA Tesla V100-SXM2 32G GPU with CUDA 12.2, learning rate $F_1, F_2, F_3, F_4$1, and weight decay $F_1, F_2, F_3, F_4$2. The comparison set includes CNN and transformer baselines such as U-Net, TransUNet, Swin-UNet, UNETR, MISSFormer, Swin-UNETR, nnFormer, UNETR++, LeVit-UNet, and PCCTrans (Dosi et al., 3 Aug 2025).

These protocol choices matter because AMBER-AFNO is presented not as a narrowly tuned proof-of-concept but as a benchmark-oriented lightweight alternative to prevailing 3D transformer systems. The comparison against both convolutional and transformer baselines is integral to that positioning (Dosi et al., 3 Aug 2025).

5. Quantitative results and efficiency profile

The reported quantitative results are split between segmentation accuracy and computational footprint. On ACDC, AMBER-AFNO achieves the best overall Dice at 92.85%, slightly above UNETR++ at 92.83%, with per-class scores of 91.60% for RV, 90.74% for Myo, and 96.21% for LV. On Synapse, AMBER-AFNO achieves 83.76% DSC and 16.96 HD95, which is competitive but below UNETR++ at 87.22 DSC and 7.53 HD95 (Dosi et al., 3 Aug 2025).

Dataset	AMBER-AFNO	Reference comparison
ACDC	92.85 DSC	UNETR++: 92.83 DSC
Synapse	83.76 DSC, 16.96 HD95	UNETR++: 87.22 DSC, 7.53 HD95

The efficiency claims are central to the model’s identity. On ACDC, the paper reports UNETR++: 81.55M params, 52.14 GFLOPs, DSC 92.83 and AMBER-AFNO: 14.77M params, 163.27 GFLOPs, DSC 92.85. On Synapse, it reports UNETR++: 42.96M params, 47.98 GFLOPs, DSC 87.22 and AMBER-AFNO: 14.86M params, 161.24 GFLOPs, DSC 83.76 (Dosi et al., 3 Aug 2025). The paper repeatedly emphasizes a reduction of trainable parameters by over 80% relative to UNETR++, while also stating that FLOPs remain comparable to other state-of-the-art architectures and that practical gains are seen in training efficiency, inference speed, and memory usage (Dosi et al., 3 Aug 2025).

The ablation study isolates the contribution of AFNO more directly. On ACDC, the paper reports AMBER (MHSA): 19.01M params, 132.07 GFLOPs, 92.03 DSC, AMBER-AFNO (light): 8.70M params, 58.29 GFLOPs, 92.83 DSC, and UNETR++: 81.55M params, 52.14 GFLOPs, 92.83 DSC (Dosi et al., 3 Aug 2025). This is the clearest internal evidence that replacing MHSA with AFNO simultaneously reduces parameter count and improves ACDC Dice over the attention-based AMBER variant.

The numerical pattern is therefore asymmetric across benchmarks. On ACDC, AMBER-AFNO is described as essentially state-of-the-art while remaining far smaller than heavy transformer baselines. On Synapse, it is lighter but not top-performing in accuracy (Dosi et al., 3 Aug 2025). This benchmark dependence is important for interpreting the model: its principal contribution is not uniform supremacy across datasets, but a favorable accuracy–efficiency trade-off, especially in settings where compactness matters.

6. Relation to broader AFNO work, limitations, and significance

Within the broader AFNO literature, AFNO has also been used as a function-to-function operator learner for regional short-range weather forecasting, where the attraction lies in efficient modeling of multiscale structure in the Fourier domain (Sunil et al., 22 Jan 2026). That work is not about AMBER-AFNO, but it places the medical-segmentation model within a broader family of Fourier-neural-operator approaches that replace or complement attention-style global interaction mechanisms. A plausible implication is that AMBER-AFNO should be understood less as a one-off architectural hybrid than as a domain-specific instantiation of a more general design pattern: global mixing through adaptive spectral operators rather than dense pairwise attention.

The limitations stated in the AMBER-AFNO paper are concrete. The main one is that performance on the more heterogeneous Synapse benchmark remains behind the best large transformer models, and the authors note that current efforts focus on reducing this gap and improving generalization across diverse anatomy and modalities (Dosi et al., 3 Aug 2025). The paper does not claim uniformly superior accuracy; instead, it frames the model as a lightweight alternative that can remain competitive while substantially reducing model size.

A second limitation is interpretive rather than numerical. The model’s efficiency profile is not reducible to a single scalar. The paper emphasizes parameter reduction, memory footprint, training efficiency, and inference speed, but the reported GFLOP counts show that low parameter count does not automatically imply minimal arithmetic cost on every benchmark (Dosi et al., 3 Aug 2025). This makes AMBER-AFNO particularly relevant to deployment regimes where parameter storage and memory behavior are critical constraints.

In the literature lineage, AMBER-AFNO marks a transition from SegFormer-style hyperspectral modeling to volumetric medical segmentation. The original AMBER established that hierarchical Transformer encoding with 3D convolutions can operate directly on high-dimensional multiband data without spectral dimensionality reduction (Dosi et al., 2024). AMBER-AFNO preserves that general architectural intuition—structured high-dimensional inputs require both global and local modeling—but substitutes AFNO for attention in order to compress the model substantially (Dosi et al., 3 Aug 2025).

The resulting significance is methodological. AMBER-AFNO demonstrates that frequency-domain token mixing can replace self-attention effectively in 3D medical image segmentation, at least on benchmarks such as ACDC, and that a carefully pruned encoder–decoder hierarchy can remain competitive with much heavier transformer systems (Dosi et al., 3 Aug 2025). Its place in the literature is therefore that of a lightweight benchmark model: not a universal replacement for all 3D transformers, but a concrete argument that spectral operator blocks can support strong volumetric segmentation with a substantially smaller trainable footprint.