M-Net: Multi-Scale Encoder–Decoder Architectures

• M-Net is a lineage of U-Net–derived, multi-scale encoder–decoder architectures that integrate additional cross-scale pathways and domain-specific modules.
• It has been adapted for diverse applications such as glaucoma analysis, fingerprint restoration, image denoising, and MRI segmentation, often leveraging modifications like bidirectional ConvLSTM or polar-coordinate preprocessing.
• While M-Net variants achieve high segmentation metrics (e.g., Dice coefficients) and improved boundary delineation, they can be computationally intensive and sensitive to pre-processing choices.

M-Net is a reused designation rather than a single invariant architecture. In the literature it most often denotes a U-Net–derived, multi-scale encoder–decoder family in which the standard skip-connected hierarchy is augmented by additional cross-scale pathways, joint supervision, or domain-specific fusion modules. In ophthalmic imaging, M-Net denotes a U-Net–type encoder–decoder for joint optic disc and optic cup segmentation, extended in one influential variant with bidirectional ConvLSTM and polar-coordinate preprocessing for glaucoma analysis; the same label is later reused for fingerprint restoration, real-image denoising, low-light enhancement, MRI brain tumor sequential segmentation, and anisotropic 2D/3D medical segmentation (Khan et al., 2021, V et al., 2018, Fan et al., 2022, Fan et al., 2022, Lu et al., 28 Jul 2025, Odendaal et al., 13 Jun 2026).

1. Terminological scope and historical lineage

A recurrent source of ambiguity is that “M-Net” names several related but non-identical designs. In the fingerprint denoising literature, M-net is described as the architecture proposed by Mehta et al. for 3D brain structure segmentation: a modified U-net that uses 3D information for segmentation via a 3D-to-2D converter block and four pathways, namely two main encoding and decoding paths plus a left leg and a right leg. In the optic disc–optic cup literature, the canonical M-Net referenced from Fu et al. is instead a U-shaped encoder–decoder with multi-scale input or side paths, skip connections, and joint multi-label supervision for coherent cup–disc segmentation. Later papers retain the name while re-centering it around restoration backbones, sequential slice modeling, or hybrid 2D/3D fusion rather than a fixed layer specification (V et al., 2018, Khan et al., 2021, Lu et al., 28 Jul 2025, Odendaal et al., 13 Jun 2026).

This reuse implies that M-Net is best understood as a lineage of hierarchical architectures rather than a single canonical graph. The stable motif is an encoder–decoder core with explicit multi-scale feature exchange; the unstable component is the inductive bias inserted into that core, such as anatomical nesting, wavelet-domain attention, temporal-like slice ordering, or anisotropy-aware 2D/3D fusion. A plausible implication is that the label survives because it denotes a recognizable design style—multi-path, multi-scale, and heavily skip-connected—even when the operational semantics differ substantially across domains.

2. Canonical architectural pattern

The earliest formulation summarized in the provided literature defines M-net as a modified U-net with four pathways. Beyond the main encoder and decoder, the left leg downsamples the input and feeds it to corresponding encoder layers, while the right leg upsamples outputs of decoding layers to the original size so that the final layer combines the outputs of the right leg and decoder layer. This arrangement is described as having similar functionality to deep supervision. In that same account, the original M-net uses a 3D-to-2D converter block, categorical cross entropy loss, and multi-class segmentation of brain structures (V et al., 2018).

The ophthalmic formulation preserves the U-shaped encoder–decoder logic but emphasizes multi-scale input or side paths and joint OD–OC prediction. Encoder blocks apply convolution, nonlinearity, and downsampling; decoder blocks use up-convolutions and skip connections at matching resolutions. Intermediate decoder stages may be supervised directly, and disc, cup, and background are predicted together as a multi-label or multi-class problem. This design targets a specific anatomical constraint: the cup lies inside the disc, so independent segmentation of the two structures ignores nesting and can degrade performance when boundaries are fuzzy or low contrast (Khan et al., 2021).

In this family, the Dice coefficient is a central segmentation measure:

$\mathrm{Dice}(S,G)=\frac{2|S\cap G|}{|S|+|G|}.$

The same family later reappears in restoration papers, but there the segmentation ancestry is repurposed as a structural prior rather than a literal class-prediction objective. That shift is especially visible in fingerprint denoising, where the problem is posed as segmenting fingerprint from noisy background, even though the final output is a continuous-valued grayscale image rather than a discrete mask (V et al., 2018).

3. M-Net for optic disc and optic cup segmentation

In fundus-image analysis, M-Net is used for joint optic disc and optic cup segmentation because glaucoma enlarges the optic cup relative to the optic disc, and the cup-to-disc ratio is a clinically significant biomarker. The cited ophthalmic variant proposes a modified M-Net with bidirectional convolutional long short-term memory, termed M-BLSTM, for joint cup and disc segmentation. The model localizes the disc region, converts disc-crop images and corresponding labels into polar coordinates, performs segmentation in polar space, maps the result back to spatial coordinates, and computes the vertical CDR from the reconstructed disc–cup mask. A vertical CDR $> 0.5$ is described as commonly suspicious for glaucoma (Khan et al., 2021).

Architecturally, the starting backbone is an M-Net or U-Net–like encoder–decoder with convolutional layers using $3\times 3$ kernels and ReLU activation, max pooling in the encoder, and $2\times 2$ up-convolution in the decoder. The principal modification is the replacement of simple skip connections by bidirectional ConvLSTM blocks. The paper describes these modules as combining features of encoding and decoding sides, so the skip path becomes a context-aware fusion operator with gating and memory rather than a static concatenation. Bidirectional fusion is summarized by

$$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$

This is used to preserve the spatial relationships among cup, disc, and background and to improve boundary delineation in low-contrast regions (Khan et al., 2021).

The preprocessing pipeline is equally central. The REFUGE2 training set contains 400 training images and 400 test images, with 40 glaucoma and 360 non-glaucoma images in each split. The 400 training images are converted into 2,000 disc-crop images and 2,000 label images in polar coordinates, each of size $400\times 400$; 80% of the polar-transformed disc-crop images are used for training and 20% for validation. The polar transformation is described as making the fundus image flat based on optic disc center, turning cup, disc, and background into layered ordered structures, obtaining spatial constraints, helping in better optic disc and cup segments, yielding a more balanced representation, and helping avoid overfitting. After inverse transformation, the authors note that interpolation can reduce Dice values slightly by deforming mask boundaries (Khan et al., 2021).

Quantitatively, on REFUGE2 the reported optic cup average Dice is $0.86$, the optic disc average Dice is $0.92$, and the validation accuracy is $98.99\%$. The paper also notes that comparison with CUHKMED, Masker, BUCT, and NKSG is indirect because those baselines are reported on REFUGE rather than REFUGE2. Qualitatively, the narrative claims improved boundary adherence and fewer misclassifications than plain M-Net without ConvLSTM, although no numerical ablation on the same dataset is given (Khan et al., 2021).

4. Restoration-oriented reinterpretations

A major second lineage treats M-Net as a restoration backbone. In FPD-M-net, the architecture is directly adapted from M-net for fingerprint denoising and inpainting. The 3D-to-2D conversion block is removed because fingerprint data are inherently 2D; the layer order is changed from Conv–ReLU–BN to Conv–BN–ReLU; dropout with $p=0.2$ is inserted between pairs of CONV–BN–ReLU blocks; and the loss is replaced by a weighted combination of per-pixel $> 0.5$0 loss and MS-SSIM with $> 0.5$1. The network is fully convolutional, uses $> 0.5$2 convolutions, $> 0.5$3 maxpooling in the encoder, upsampling in the decoder, and a final $> 0.5$4 convolution with sigmoid. On the test set, FPD-M-net reports MSE $> 0.5$5, PSNR $> 0.5$6, and SSIM $> 0.5$7, outperforming U-net on all three metrics; in the Chalearn LAP Inpainting Competition it ranks 3rd overall and 2nd in SSIM (V et al., 2018).

In real-image denoising, Selective Residual M-Net redefines M-Net as a hierarchical multi-scale U-Net–like topology with four layers, an initial channel count of 96, dual downsampling paths, and SK-based fusion. At each scale, bilinear downsampling supplies smoother semantic information, while pixel unshuffle preserves local detail by rearranging spatial resolution into channels. Classical concatenation is replaced by Selective Kernel Feature Fusion, and local feature extraction is handled by the Selective Residual Block rather than a residual dense block. The training objective is the Charbonnier loss,

$> 0.5$8

with $> 0.5$9. SRMNet reports $3\times 3$0 dB / $3\times 3$1 SSIM on SIDD, $3\times 3$2 dB / $3\times 3$3 on DND, and $3\times 3$4 G FLOPs for $3\times 3$5 inputs, compared with $3\times 3$6 G for MIRNet (Fan et al., 2022).

HWMNet extends this restoration line through M-Net+, an improved hierarchical model for low-light enhancement. M-Net+ replaces max pooling in the trunk path with pixel unshuffle, uses bilinear downsampling in gatepost paths, and replaces concatenative decoder fusion by SKFF. Its distinctive block is the Half Wavelet Attention Block, which splits the channels into an identity half and a transformed half, applies DWT, a Dual Attention Unit, and IWT only to the transformed half, then recombines the two branches with residual learning. On LOL, HWMNet reports PSNR $3\times 3$7, SSIM $3\times 3$8, and LPIPS $3\times 3$9, with $2\times 2$0 T FLOPs for a $2\times 2$1 color image; on MIT-Adobe FiveK it reports PSNR $2\times 2$2 and SSIM $2\times 2$3 (Fan et al., 2022).

5. Sequential and hybrid volumetric M-Nets

A more recent redefinition appears in MRI brain tumor segmentation, where M-Net is no longer primarily a multi-path U-Net variant but a sequential image segmentation framework for adjacent 2D MRI slices. Here the input is a sequence

$2\times 2$4

with $2\times 2$5 contiguous slices and $2\times 2$6 modalities, typically T1, T1c, T2, and FLAIR. The architecture combines a 2D encoder–decoder backbone, a Vision Sequential Module using Cross-Scan serialization inside each slice, and the Mesh-Cast Sequential Module, which first models temporal-like correlations across slices and then transposes the representation so that channels become the sequential axis. Spatial dimensions are flattened with $2\times 2$7, so each slice becomes $2\times 2$8; the sequence is processed along $2\times 2$9, transposed into a channel sequence, processed again along $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$0, and reshaped back to $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$1. Training uses the Two-Phase Sequential schedule: shuffled sequence training followed by ordered sequence fine-tuning. On BraTS 2019, M-Net with Mamba SSM and TPS reports Dice WT $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$2, TC $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$3, ET $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$4, with Hausdorff95 WT $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$5, TC $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$6, ET $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$7; on BraTS 2023 it reports Dice WT $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$8, TC $$H_t^{\mathrm{bi}}=\mathrm{concat}\!\left(H_t^{\rightarrow},H_t^{\leftarrow}\right).$$9, ET $400\times 400$0 (Lu et al., 28 Jul 2025).

Another distinct lineage is the hybrid 2D/3D MNet for anisotropic medical image segmentation. In this formulation, MNet embeds 2D and 3D convolutions in each latent block and fuses their features across five encoder stages and five decoder stages, with six auxiliary output branches for deep supervision. Original feature fusion is handled by Feature Merging Units using fixed operations such as addition, subtraction, or concatenation followed by convolution. The reproduced model achieves DSC $400\times 400$1 on PROMISE and $400\times 400$2 for liver and tumor segmentation on LiTS. MNet++ introduces learned Fusion Gating for adaptive 2D–3D blending and a VMamba state-space module for long-range depth modeling; the Spatial Gating variant improves DSC by $400\times 400$3 with less than $400\times 400$4 inference overhead, VMamba yields $400\times 400$5 Dice for LiTS liver, and robustness to anisotropy is summarized by $400\times 400$6Dice $400\times 400$7 across $400\times 400$8–$400\times 400$9 mm voxel spacing (Odendaal et al., 13 Jun 2026).

6. Common design principles, limitations, and related nomenclature

Across these variants, several design principles recur. First, M-Net almost always preserves a hierarchical encoder–decoder with aggressive multi-scale information flow. Second, the architecture is usually specialized by injecting a domain prior directly into skip paths or resolution transitions: polar coordinates and bidirectional ConvLSTM in fundus analysis, left-leg and right-leg pathways in the original segmentation formulation, bilinear plus pixel-unshuffle dual downsampling in denoising, half-wavelet processing in low-light enhancement, temporal-like slice ordering in MRI, or explicit 2D/3D branch fusion in anisotropic segmentation (Khan et al., 2021, V et al., 2018, Fan et al., 2022, Fan et al., 2022, Lu et al., 28 Jul 2025, Odendaal et al., 13 Jun 2026).

The limitations are equally domain-specific. In the glaucoma model, ConvLSTMs are heavier than simple skip concatenations, training is partly reliant on accurate disc localization and polar-coordinate preprocessing, and inverse polar transformation can reduce Dice because of interpolation-induced deformation (Khan et al., 2021). In restoration-oriented M-Net derivatives, the original M-Net’s concatenative shallow layers are described as computationally expensive, motivating SKFF-based redesigns (Fan et al., 2022). In sequential MRI segmentation, Mesh-Cast adds computational overhead relative to a plain 2D backbone, depends on correct slice ordering, and approximates rather than fully replaces 3D receptive fields (Lu et al., 28 Jul 2025). In anisotropic 2D/3D segmentation, the tumor class remains noisy under restricted data and shorter training schedules, even when liver performance and variance improve (Odendaal et al., 13 Jun 2026).

Because the name is overloaded, adjacent terms require care. MMNet for Massive MIMO detection is an unfolded iterative detector for neural signal detection rather than an encoder–decoder image model, and S$0.86$0M-Net is a joint semantic segmentation and stereo matching framework for autonomous driving with a shared encoder, an FFA module, and a semantic consistency-guided loss (Khani et al., 2019, Wu et al., 2024). This suggests that “M-Net” now functions less as a unique architecture than as a historically layered naming convention whose precise meaning is fixed by the application and the paper-specific modules wrapped around a multi-scale backbone.