U-MAN: Multi-scale Adaptive KAN Network

• The paper introduces U-MAN, which enhances U-Net segmentation by integrating PAGF for attention-guided fusion and MAN for adaptive multi-scale feature extraction.
• It replaces conventional skip connections with a Progressive Attention-Guided Feature Fusion module to bridge semantic gaps and reduce noise in segmentation predictions.
• Empirical results on BUSI, GLAS, and CVC-ClinicDB datasets show consistent improvements in IoU and F1, underlining better boundary definition and fine-detail preservation.

Searching arXiv for the target paper and closely related KAN/U-Net segmentation works to ground the article in recent literature. U-MAN, short for U-Net with Multi-scale Adaptive KAN Network, is a medical image segmentation architecture that modifies the classical U-Net paradigm to address two limitations identified in conventional U-Net-based systems: the semantic gap in simple skip connections and insufficient multi-scale deep feature extraction. It does so by augmenting an encoder–decoder backbone with two specialized modules, Progressive Attention-Guided Feature Fusion (PAGF) and Multi-scale Adaptive KAN (MAN), thereby coupling attention-guided encoder–decoder fusion with adaptive multi-scale feature processing in deep layers. Reported experiments on the BUSI, GLAS, and CVC datasets indicate that the model improves boundary definition and fine-detail preservation relative to prior baselines, including U-Net and U-KAN (Huang et al., 26 Sep 2025).

1. Problem formulation and design rationale

Medical image segmentation is especially sensitive to boundary quality, small structures, and pathological heterogeneity. U-MAN is explicitly framed around two shortcomings of conventional U-Net architectures. First, standard skip connections directly fuse shallow encoder features with decoder features while ignoring the semantic gap between spatially detailed shallow representations and abstract deep representations. In the formulation used for U-MAN, this direct fusion can introduce noise and artifacts into predicted masks. Second, conventional U-Net, and even U-KAN, are described as lacking strong mechanisms for adaptive multi-scale processing in deep layers, which limits performance when target objects vary substantially in shape and size (Huang et al., 26 Sep 2025).

The architecture therefore separates the two issues. The PAGF module is introduced to replace simple skip connections and to perform attention-guided fusion between encoder and decoder features. The MAN module is introduced in deeper stages to extract and combine contextual information across scales while retaining the representational strengths of Kolmogorov-Arnold Networks. This division of labor is central to the design: PAGF addresses feature compatibility across the encoder–decoder bridge, whereas MAN addresses feature richness and scale adaptivity within deep semantic processing (Huang et al., 26 Sep 2025).

A common simplification is to describe U-MAN as merely “U-KAN with attention.” That description is incomplete. The architecture is organized around a specific pairing of attention-guided skip fusion and dual-branch multi-scale adaptive KAN processing, and these are treated as complementary rather than interchangeable components.

2. Network topology and stage allocation

The overall structure remains U-shaped. The encoder contains three convolutional stages for shallow spatial feature extraction, followed by two MAN stages for deep, multi-scale, non-linear processing. The decoder performs progressive upsampling to reconstruct the segmentation mask. The critical deviation from classic U-Net lies in the skip pathways: skip connections are not passed directly from encoder to decoder, but are intercepted by PAGF modules that perform attention and gating-based fusion (Huang et al., 26 Sep 2025).

The architecture overview described for the model presents a left-to-right encoder–decoder structure in which shallow layers are conventional convolutional blocks, deeper layers are MAN modules, and all skip pathways are mediated by PAGF rather than by simple copying or concatenation. In this respect, U-MAN preserves the macro-structure of U-Net while changing the computational meaning of both deep encoder stages and skip fusion.

This arrangement suggests a deliberately hybrid allocation of operations. Convolutional stages are retained where shallow spatial detail is dominant, while MAN is reserved for deeper feature maps where multi-scale contextual reasoning and non-linear function modeling are more consequential. The paper does not recast the entire backbone into KAN blocks; instead, it inserts them where higher-level semantic structure is expected to matter most (Huang et al., 26 Sep 2025).

3. Multi-scale Adaptive KAN (MAN)

The MAN module is the principal deep feature extractor in U-MAN. Its stated purpose is to adaptively extract and fuse multi-scale contextual features using both KAN and multi-scale attention, with the goal of handling complex anatomical variability. The module begins from features embedded by a patch embedding layer and then follows a dual-branch design (Huang et al., 26 Sep 2025).

One branch is the KAN branch, which processes features through a sequence of LayerNorm $\rightarrow$ KAN layer $\rightarrow$ DropPath. The other branch is the MSAB branch, where MSAB denotes a Multi-Scale Attention Block. This branch is composed of cascading Conv-BN-ReLU layers and includes mixed-kernel $1\times1$/$3\times3$/$5\times5$ depthwise convolutions to capture features at multiple receptive fields. The outputs of the two branches are then fused using learnable scalar weights, allowing the model to emphasize fine detail or broader context as needed (Huang et al., 26 Sep 2025).

Formally, with input feature tensor $$\mathbf{X}_{\text{embed}} \in \mathbb{R}^{H_{\text{embed}} \times W_{\text{embed}} \times C}$$, the module is described by

$$\mathbf{F}_{\text{kan}} = \text{KANBlock}(\mathbf{X}_{\text{embed}})$$

$$\mathbf{F}_{\text{msab}} = \text{MSAB}(\mathbf{X}_{\text{embed}})$$

$$\mathbf{F}_{\text{man}} = w_1 \otimes \mathbf{F}_{\text{msab}} + w_2 \otimes \mathbf{F}_{\text{kan}},$$

where $w_1$ and $\rightarrow$0 are learnable weights (Huang et al., 26 Sep 2025).

Within this construction, the KAN block is described as a grid-based neural operator, distinct from convolutional or transformer blocks, and associated with high-capacity, interpretable, non-linear function approximation. The target paper further states that, inside MAN, stacked KAN blocks with three layers are optimal, with normalization and stochastic regularization via DropPath used for robust training. This makes MAN not just a multi-branch feature mixer, but a module in which KAN-based non-linear function modeling and multi-scale attention-based spatial processing are explicitly co-optimized (Huang et al., 26 Sep 2025).

This also distinguishes MAN from other KAN-derived units in the literature. U-KAN inserts tokenized KAN layers into deeper semantic stages of a U-Net pipeline, but does not define the dual-branch KAN-plus-multi-scale-attention mechanism used in MAN (Li et al., 2024). Implicit U-KAN 2.0, by contrast, combines second-order neural ODEs with MultiKAN layers in a SONO-MultiKAN phase, emphasizing continuous-time feature evolution rather than the PAGF/MAN decomposition (Cheng et al., 5 Mar 2025).

4. Progressive Attention-Guided Feature Fusion (PAGF)

The PAGF module is the skip-fusion mechanism of U-MAN. Its function is to bridge the semantic gap between encoder and decoder representations by using data-driven attention and gating to suppress noise and irrelevant information before fusion. In contrast to standard U-Net skip connections, shallow encoder features and upsampled decoder features are not merged indiscriminately (Huang et al., 26 Sep 2025).

PAGF takes as input an upsampled decoder feature map $\rightarrow$1 and a shallow encoder feature map $\rightarrow$2. The two are concatenated to gather context. An attention map $\rightarrow$3 is then used to focus on salient regions, and a gating map $\rightarrow$4 learns spatially adaptive fusion weights. The final fusion is defined as

$\rightarrow$5

after which the result passes through a final convolutional refinement layer (Huang et al., 26 Sep 2025).

The semantic interpretation of this formulation is straightforward. Rather than treating encoder features as inherently beneficial high-resolution detail, PAGF evaluates both decoder and encoder activations under a shared attention map and then uses gating to determine the proportion contributed by each source at each location. The target paper’s diagram summary characterizes PAGF as dual-attention followed by gated fusion, and the ablation description further notes that channel attention and the gating mechanism each significantly contribute, with gains of up to 1% IoU increase reported in the ablation setting (Huang et al., 26 Sep 2025).

This makes PAGF more than a routine attention gate. Its role is specifically to manage incompatibility between feature abstractions at different depths. In that sense, PAGF operationalizes the claim that skip connections are not merely transport pathways but semantic reconciliation mechanisms.

5. Empirical evaluation and ablation evidence

U-MAN is evaluated on three public datasets: BUSI for breast ultrasound segmentation (647 images), GLAS for gland segmentation in histopathology (165 images), and CVC-ClinicDB for colon polyp segmentation (612 endoscopy images) (Huang et al., 26 Sep 2025).

The reported main results are given in terms of IoU and F1. On BUSI, U-Net achieves 57.2 / 71.9, U-KAN 65.0 / 78.6, and U-MAN 66.1 / 79.1, corresponding to an improvement of +1.1 IoU and +0.6 F1 over U-KAN. On GLAS, U-Net achieves 86.7 / 92.8, U-KAN 87.7 / 93.5, and U-MAN 89.1 / 94.2, an improvement of +1.3 IoU and +0.7 F1 over U-KAN. On CVC-ClinicDB, U-Net reaches 83.8 / 91.1, U-KAN 82.7 / 90.2, and U-MAN 86.1 / 92.5, yielding +3.5 IoU and +0.8 F1 over U-KAN (Huang et al., 26 Sep 2025).

The qualitative interpretation given for these results is that U-MAN produces smoother, more accurate boundaries and better detail preservation, especially in challenging images with complex backgrounds or small or irregular targets. The paper explicitly links the observed gains to boundary accuracy and fine-detail retention (Huang et al., 26 Sep 2025).

The ablation study on CVC-ClinicDB isolates the contribution of the two principal modules. Starting from U-KAN at 82.69 IoU / 90.24 F1, adding MAN only yields 83.82 / 90.95, adding PAGF only yields 84.91 / 91.89, and the full U-MAN reaches 86.14 / 92.47. The reported interpretation is that the modules are complementary and jointly produce the best result. The ablation also states that multi-scale attention adds approximately 0.5–0.7 IoU, that three KAN layers is optimal, and that channel attention and gating each significantly contribute (Huang et al., 26 Sep 2025).

These results are narrowly defined but methodologically informative. They indicate that the largest isolated gain comes from revising skip fusion, while the best overall performance requires both revised fusion and revised deep multi-scale feature extraction. A plausible implication is that boundary-sensitive medical segmentation benefits from simultaneous control of inter-level feature compatibility and intra-level scale diversity.

6. Position within the KAN-based U-shaped segmentation literature

U-MAN belongs to a broader family of architectures that incorporate Kolmogorov-Arnold Networks into U-shaped segmentation backbones, but it occupies a specific niche within that literature. The earlier U-KAN framework integrates KAN layers into tokenized intermediate representations in a U-Net backbone and emphasizes learnable non-linear activations on tokenized deep features; U-MAN inherits the general KAN orientation but adds the MAN and PAGF modules to target multi-scale deep feature extraction and skip-connection semantics more directly (Li et al., 2024).

Other related variants address different bottlenecks. Implicit U-KAN 2.0 replaces discrete deep processing with second-order neural ODEs and MultiKAN, emphasizing continuous feature evolution, theoretical approximation properties independent of input dimension, and constant-memory adjoint training rather than attention-guided skip fusion (Cheng et al., 5 Mar 2025). UKAN-EP extends U-KAN to 3D multi-modal MRI brain tumor segmentation and incorporates Efficient Channel Attention and Pyramid Feature Aggregation, together with a dynamic loss weighting strategy, thereby focusing on volumetric multi-modality and efficiency (Chen et al., 2024). KANDU-Net instead uses a dual-channel KAN-convolution structure with an auxiliary fusion network to combine local convolutional and KAN features (Fang et al., 2024).

A second cluster of works combines KAN-style processing with alternative global-modeling mechanisms. KM-UNet integrates KANs with state-space models and a Selective-Scan Efficient Multi-scale attention module, explicitly targeting long-range dependency modeling and efficiency (Zhang, 5 Jan 2025). PGM-UNet uses a prompt-guided CNN-Mamba dual-path UNet and a KAN-based multi-scale information extraction module at the bottleneck, emphasizing parallel local–global fusion rather than U-MAN’s attention-guided skip replacement (Zhang et al., 25 Mar 2025).

There is also a distinct line of research on explicit multi-scale U-Net modifications that do not rely on KAN. Deep Multi-Scale U-Net Architecture and Label-Noise Robust Training Strategies for Histopathological Image Segmentation adds multi-scale encoder feature maps and annotation-noise-aware training via fuzzy boundaries and bootstrapping, showing that explicit scale handling and robustness can materially improve segmentation even without KAN-based operators (Kurian et al., 2022). Conversely, U-KABS extends the KAN family by combining Bernstein polynomials and B-splines within a U-shaped encoder–decoder, framing multi-scale adaptivity in terms of the basis functions used in learnable activations (Bhattacharyya et al., 7 Feb 2026).

Within this landscape, U-MAN is most precisely understood as a U-KAN-derived segmentation architecture specialized for two specific failure modes of classical U-Net: semantic inconsistency in skip fusion and insufficient multi-scale semantic processing in deep layers. It does not primarily target 3D multi-modal fusion, continuous-time dynamics, state-space sequence modeling, or annotation-noise robustness. Its contribution lies in pairing PAGF and MAN so that boundary delineation and fine-detail preservation become explicit architectural objectives rather than indirect side effects of deeper or more global models (Huang et al., 26 Sep 2025).