MFEnNet: MetaFormer Encoding for Med Segmentation

• MFEnNet is a MetaFormer-driven Encoding Network that refines the U-Net encoder with pooling-based token mixing, Swish activation, and SPP for efficient and robust medical segmentation.
• The network employs a modified MetaFormer block to combine global context with multi-scale feature extraction, significantly reducing computational cost compared to conventional transformer-based methods.
• MFEnNet achieves competitive IoU and Dice scores on benchmarks like CASIA and ISBI while cutting parameters by 64% and FLOPs by nearly 68% versus standard U-Net architectures.

Searching arXiv for the named model and closely related work. MFEnNet, short for MetaFormer-driven Encoding Network, is a medical image semantic segmentation framework that reformulates the encoder of a vanilla U-Net according to the MetaFormer paradigm in order to combine global context modeling with reduced computational cost. It is designed for robust medical semantic segmentation, with an explicit emphasis on resource-constrained clinical settings in which many strong segmentation models, especially transformer-based ones, are too expensive to deploy. The model replaces conventional transformer self-attention with pooling-based token mixing, uses Swish activation for smoother optimization, and inserts Spatial Pyramid Pooling (SPP) at the bottleneck for multi-scale feature extraction. In the reported experiments, MFEnNet attains competitive accuracy while significantly lowering computational cost relative to several state-of-the-art baselines (Tran et al., 1 Jan 2026).

1. Definition and intended scope

MFEnNet denotes a MetaFormer-driven Encoding Network whose main goal is to provide accurate yet computationally efficient medical image semantic segmentation (Tran et al., 1 Jan 2026). The motivating observation is that semantic segmentation is crucial for medical image analysis because it enables precise disease diagnosis and treatment planning, while many advanced models employ complex architectures that limit their use in clinical environments with constrained compute budgets.

The model is therefore positioned around three simultaneous design targets: high segmentation accuracy, low computational complexity, and better suitability for real-world medical deployment. This positioning is specific: MFEnNet is not presented as a general-purpose vision backbone, but as a U-Net-derived segmentation architecture whose encoder is redesigned to preserve the benefits of long-range context modeling without incurring the full cost of standard transformer self-attention (Tran et al., 1 Jan 2026).

A plausible implication is that MFEnNet belongs to a broader class of hybrid segmentation architectures that retain the deployment pragmatics of encoder–decoder designs while importing selected transformer-era abstractions only where their benefit-to-cost ratio is highest.

2. Architectural organization and MetaFormer integration

MFEnNet is built on a vanilla U-Net backbone, that is, a classic encoder–decoder segmentation architecture with skip connections (Tran et al., 1 Jan 2026). Its high-level organization comprises an encoder for multi-level feature extraction, a bottleneck that includes SPP, a decoder that reconstructs the segmentation map through upsampling and skip fusion, and a final $1 \times 1$ convolution that produces the segmentation mask.

The network uses five hierarchical stages:

• $H \times W \times 32$
• $\frac{H}{2} \times \frac{W}{2} \times 64$
• $\frac{H}{4} \times \frac{W}{4} \times 128$
• $\frac{H}{8} \times \frac{W}{8} \times 256$
• $\frac{H}{16} \times \frac{W}{16} \times 512$

The defining modification occurs in the encoder. Instead of ordinary CNN-only encoding blocks, the encoder is reformulated according to the MetaFormer template. The paper describes the standard ViT-style formulation as

$Z = \mathrm{InputEmb}(X)$

where $X$ is the input feature map and $Z \in \mathbb{R}^{N \times C}$ is the token sequence with $N$ tokens and $H \times W \times 32$0 channels. The two standard sub-blocks are

$H \times W \times 32$1

$H \times W \times 32$2

MFEnNet preserves this MetaFormer-like two-stage block structure, but replaces the token mixer with pooling:

$H \times W \times 32$3

This preserves the sequence of normalization, token mixing, residual connection, normalization, feed-forward network, and residual connection, while altering the mixing primitive itself (Tran et al., 1 Jan 2026).

Several implementation choices are explicit. The input embedding is a $H \times W \times 32$4 convolution. The FFN uses two fully connected layers with expansion ratio $H \times W \times 32$5. The activation inside the FFN is Swish. Skip connections are used in each sub-block. MetaFormer blocks are inserted only in the encoder, because the authors state that long-range context is most critical there (Tran et al., 1 Jan 2026).

3. Pooling-based token mixing, Swish, and spatial pyramid pooling

The central efficiency mechanism in MFEnNet is the replacement of conventional self-attention with pooling-based token mixing. In a normal ViT block, token mixing typically relies on self-attention, which is expensive because it models pairwise interactions among all tokens. For high-resolution medical images, this is costly in both memory and computation. MFEnNet instead uses

$H \times W \times 32$6

in place of attention-based mixing (Tran et al., 1 Jan 2026).

The stated rationale for this substitution is computational. Pooling-based token mixing is lighter because it avoids quadratic token interaction cost, does not build a full attention matrix, performs localized aggregation more efficiently, and still enables contextual information flow across the feature map. The paper explicitly characterizes the result as global feature aggregation at substantially reduced computational cost (Tran et al., 1 Jan 2026). This suggests a design hypothesis associated with MetaFormer-style architectures: the performance contribution may derive substantially from the architectural scaffold, not only from self-attention as the mixer.

MFEnNet also replaces ReLU with Swish. The reported roles of Swish are smoother gradients, better optimization stability, mitigation of vanishing-gradient-related issues, and potentially stronger representational power than ReLU. Within the paper’s interpretation, this contributes to faster convergence and more stable training (Tran et al., 1 Jan 2026).

At the bottleneck, the model inserts Spatial Pyramid Pooling (SPP) to improve multi-scale feature extraction. SPP pools over regions of different sizes so that the network can capture local details, global semantics, and structures at multiple receptive-field scales. The paper states that SPP helps capture “both local details and global context across various spatial resolutions”. This is described as particularly useful when segmenting structures that vary in size and shape, including iris boundaries, lesion regions, and fine anatomical contours (Tran et al., 1 Jan 2026).

4. Experimental protocol, datasets, and metrics

MFEnNet is evaluated on two public medical segmentation benchmarks (Tran et al., 1 Jan 2026). The first is CASIA Iris Interval (CASIA), comprising 2,639 iris images from 395 eyes of 249 subjects, collected with a consistent sensor and split into 80% training and 20% validation. The second is ISBI 2016 skin lesion segmentation (ISBI), with 900 training images and 379 test images.

Input images are resized to $H \times W \times 32$7. Data augmentation consists of random flipping and random cropping. The implementation uses PyTorch on a Linux system with NVIDIA Tesla T4 GPUs. Training uses the Adam optimizer with batch size 16, learning rate $H \times W \times 32$8, 50 epochs, and binary cross-entropy (BCE) loss (Tran et al., 1 Jan 2026).

The evaluation protocol includes both accuracy and efficiency metrics. Accuracy is measured by IoU and Dice coefficient. Efficiency is measured by Params in millions and FLOPs in gigaflops for $H \times W \times 32$9 input. This is important for interpreting MFEnNet, because the paper’s claims are not restricted to accuracy; they concern the joint accuracy–efficiency trade-off.

5. Reported results and accuracy–efficiency profile

The reported comparisons include U-Net, ResUNet, U-Net++, TransUNet, ResUNet++, and ThinDyUNet (Tran et al., 1 Jan 2026). On CASIA, MFEnNet records IoU 0.9484 and Dice 0.9732. On the same benchmark, U-Net records IoU 0.9487 and Dice 0.9734, U-Net++ records IoU 0.9484 and Dice 0.9730, and TransUNet records IoU 0.9462 and Dice 0.9720. On ISBI, MFEnNet records IoU 0.8218 and Dice 0.8913, compared with U-Net++ at IoU 0.8238 and Dice 0.8947, U-Net at IoU 0.8199 and Dice 0.8911, and TransUNet at IoU 0.8195 and Dice 0.8894 (Tran et al., 1 Jan 2026).

Comparison target	MFEnNet	Selected baseline values
CASIA IoU / Dice	0.9484 / 0.9732	U-Net 0.9487 / 0.9734
ISBI IoU / Dice	0.8218 / 0.8913	U-Net++ 0.8238 / 0.8947
Params / FLOPs	11.14M / 17.13	U-Net 31.04M / 54.66

The efficiency numbers are central to the paper’s characterization of the model. MFEnNet uses 11.14M parameters and 17.13 GFLOPs. The comparison values reported are U-Net: 31.04M params, 54.66 GFLOPs; ResUNet: 13.04M params, 80.83 GFLOPs; U-Net++: 9.16M params, 34.87 GFLOPs; TransUNet: 3.63M params, 33.36 GFLOPs; ResUNet++: 14.48M params, 70.92 GFLOPs; and ThinDyUNet: 0.81M params, 13.56 GFLOPs (Tran et al., 1 Jan 2026).

The paper’s explicit interpretation is that MFEnNet achieves competitive or near-best accuracy, and on ISBI is very close to the top performers while being much more efficient than many baselines. It also emphasizes that MFEnNet cuts parameters by 64% vs U-Net and FLOPs by nearly 68% vs U-Net. Qualitatively, the reported outputs show cleaner iris boundaries on CASIA, better lesion contours on ISBI, less over-segmentation and fewer artifacts than competing methods, and stronger robustness on small or irregular lesions (Tran et al., 1 Jan 2026).

No formal module-by-module ablation that separately removes the MetaFormer/pooling block, Swish, or SPP is reported in the described study. Instead, the analysis attributes lower complexity to pooling blocks, improved optimization stability and gradient flow to Swish, and improved multi-scale representation to SPP (Tran et al., 1 Jan 2026).

6. Nomenclature, disambiguation, and relation to adjacent arXiv usage

A persistent source of confusion is that “MFEnNet” resembles several unrelated acronyms in the arXiv literature. In the context of medical semantic segmentation, however, the term denotes the MetaFormer-driven Encoding Network introduced in “MetaFormer-driven Encoding Network for Robust Medical Semantic Segmentation” (Tran et al., 1 Jan 2026).

Several adjacent names should be distinguished from it. “Mean-Field Networks” uses MFN to denote a feed-forward neural-network view of mean-field variational inference for pairwise MRFs and CRFs (Li et al., 2014). “Proactive Message Passing on Memory Factor Networks” introduces Memory Factor Networks (MFNs) together with Proactive Message Passing (PMP), and does not use “MFEnNet” as the paper’s model name (Eschenfeldt et al., 2016). “MFNets: Data efficient all-at-once learning of multifidelity surrogates as directed networks of information sources” uses MFNets for a multifidelity surrogate framework based on a directed acyclic graph (DAG) over information sources (Gorodetsky et al., 2020).

The acronym can also be confused with geometry- and manifold-oriented models. “Mesh-Informed Neural Networks for Operator Learning in Finite Element Spaces” introduces MINNs, a sparse neural-operator architecture built around mesh-based finite element spaces (Franco et al., 2022). “Manifold Filter-Combine Networks” and “Convergence of Manifold Filter-Combine Networks” introduce MFCNs, a filter-combine framework for manifold neural networks with graph-based point-cloud implementations and continuum-limit guarantees (Johnson et al., 2023, Johnson et al., 2024). “Manifold Function Encoder: Identifying Different Functions Defined on Different Manifolds” introduces MFE, an encoder for functions defined on varying manifolds and domains (Hu et al., 23 Dec 2025).

Outside medical imaging, still other similar names appear. “A Mask Free Neural Network for Monaural Speech Enhancement” introduces MFNet, a mask-free speech-enhancement model built from Global Local Former Blocks (GLFBs) (Liu et al., 2023). “A Markov Random Field Multi-Modal Variational AutoEncoder” proposes MRF MVAE, GMRF MVAE, ALMRF MVAE, and NN-MRF MVAE, and is explicitly not specifically about “MFEnNet” by name (Oubari et al., 2024). “Mesh Graph Neural Network Framework for Accelerating Finite Element Simulation for Arbitrary Geometries” presents a mesh graph network (MGN) for predicting von Mises stress fields across varying geometries, again unrelated to the MetaFormer-driven medical segmentation model (Kunz et al., 6 Jun 2026).

Accordingly, MFEnNet is best treated as a specific term of art for the MetaFormer-driven U-Net encoder used in medical semantic segmentation, not as a generic umbrella acronym for MFN-, MFNet-, MFCN-, MFE-, or mesh-based architectures.