UltraLBM-UNet: Ultralight Lesion Segmentation

• The paper introduces UltraLBM-UNet, an ultralight U-Net architecture that integrates bidirectional Mamba-based global context and multi-branch local feature perception to achieve high segmentation fidelity.
• Its six-stage encoder–decoder design, featuring variable kernel sizes and shared Mamba weights, delivers robust performance with over 79% IoU and 88% DSC on benchmarks like ISIC17 and PH².
• The architecture supports real-time deployment in constrained environments with sub-0.06 GFLOPs and a memory footprint under 0.14 MB, validated through extensive ablation studies and hybrid knowledge distillation.

UltraLBM-UNet is an ultralight variant of the U-Net architecture designed for high-performance and resource-efficient skin lesion segmentation. It incorporates a bidirectional Mamba-based global modeling mechanism with multi-branch local feature perception, resulting in a model that delivers robust segmentation accuracy with extremely low computational complexity. The architecture supports deployment in point-of-care scenarios, where memory and inference latency are critical constraints, without sacrificing segmentation fidelity (Fan et al., 25 Dec 2025).

1. Architectural Principles and Configuration

UltraLBM-UNet utilizes a six-stage encoder–decoder configuration. The encoder channels are [8, 16, 24, 32, 48, 64], mirrored in the decoder, enabling hierarchical feature extraction. Early encoder stages (I–III) employ conventional Conv–ReLU–Conv–ReLU blocks with max-pooling for local representation learning. Shallow encoder stage III integrates a Local Multi-Branch Perception (LMBP) module comprising three DwConv branches (kernel sizes: 3, 5, 7 dependent on depth) and an identity branch, emphasizing local feature detail.

Deep encoder stages (IV–VI) and the first three decoder stages feature Global–Local Multi-Branch Perception (GLMBP) modules. These modules merge bidirectional Mamba-based global context with depthwise-separable convolution branches, balancing spatial reach and edge preservation. Skip-connections are implemented via element-wise addition with a learnable scalar scale factor $k$, expressed as $X_i = \hat{X}_i + k t_i$. Bilinear interpolation is used for upsampling in the decoder.

The distilled model, UltraLBM-UNet-T, retains an identical topology but halves all channel widths ([4, 8, 12, 16, 24, 32]), lowering both parameter count and FLOPs for further resource efficiency.

2. Bidirectional Mamba-Based Global Modeling

UltraLBM-UNet exploits the Mamba state-space model (SSM) for linear-time, long-range dependency modeling within feature maps. Given a feature map $X \in \mathbb{R}^{B \times C \times H \times W}$, the tensor is flattened to a sequence of length $N = H \cdot W$. LayerNorm is applied, and channels are split into four parts, $$\{X_1, X_2, X_3, X_4\} \in \mathbb{R}^{B \times N \times (C/4)}$$.

Branches 1 and 2 serve global context extraction via Mamba modules in bidirectional configuration: $$M_{i,\text{forward}} = M(X_i), \quad M_{i,\text{backward}} = \text{Flip}(M(\text{Flip}(X_i)))$$ Feature fusion is performed by $$X_i^{g} = M_{i,\text{forward}} + M_{i,\text{backward}} + \gamma X_i$$, with $\gamma$ a learnable scalar. Shared Mamba weights ensure parameter efficiency while doubling contextual information. This strategy enables non-causal global modeling over the spatial domain of the image.

3. Multi-Branch Local Feature Perception and Multi-Receptive-Field Design

Branch 3 processes local features via DwConv, reshaping $X_3$ to 2D, then computing $$X_3^l = \text{DwConv}_\text{2d}(\text{reshape}(X_3)) + \gamma X_3$$. Branch 4 employs an identity shortcut: $X_4^i = X_4 + \gamma X_4$. Final multi-branch fusion concatenates outputs along the channel dimension: $$X_{\text{fuse}} = [X_1^g \| X_2^g \| X_3^l \| X_4^i]$$.

Multi-receptive-field design is enforced via variable kernel sizes in GLMBP modules:

• Encoder stages IV, V, VI: DwConv kernels $\{3, 5, 7\}$
• Decoder stages I, II, III: kernels $\{7, 5, 3\}$ This composition maximizes localization detail while maintaining parameter compactness.

4. Hybrid Knowledge Distillation to UltraLBM-UNet-T

UltraLBM-UNet-T is trained using a hybrid knowledge distillation regime from the full model. The total loss is: $$L_{\text{distill}} = \lambda_h L_{\text{hard}} + \lambda_s L_{\text{DKD}} + \lambda_a L_{\text{AT}} + \lambda_g L_{\text{grad}}$$ Where:

• $L_{\text{hard}}$ is the standard BCE + Dice loss on segmentation outputs,
• $L_{\text{DKD}}$ (Decoupled Knowledge Distillation) aligns pixel probability distributions via KL divergence,
• $L_{\text{AT}}$ (Attention Transfer) aligns spatial attention maps,
• $L_{\text{grad}}$ matches gradient boundaries via Sobel edge operator between student and teacher predictions.

Ablations on the ISIC2017 dataset demonstrate incremental improvements: base student (w/o distillation) achieves 77.30% IoU / 87.20% DSC, DKD only increases to 78.19%/87.76%, and full loss yields 78.57% IoU / 88.00% DSC.

5. Computational Complexity and Resource Profiles

Model complexity is strictly constrained:

• UltraLBM-UNet: 0.034M parameters (≈0.14 MB), 0.060 GFLOPs for $256 \times 256$ inputs.
• UltraLBM-UNet-T: 0.011M parameters (≈0.044 MB), 0.019 GFLOPs.

Module breakdown: | Module | Params (M) | FLOPs (G) | |------------------------------------|:----------:|:---------:| | Stem & shallow encoder (I–III) | 0.012 | 0.020 | | Encoder/decoder GLMBP (6 total) | 0.016 | 0.030 | | Skip-scale, upsampling, LayerNorm | 0.006 | 0.010 |

Editor's term: "Resource profiles" — UltraLBM-UNet and its student variant require exceedingly small memory and compute, making them suitable for real-time operation on embedded GPUs (e.g., NVIDIA Jetson, mobile NPUs) with sub-10 ms inference and <1 MB footprint.

6. Segmentation Performance Benchmarks

Extensive evaluation was conducted on ISIC2017, ISIC2018, and PH² datasets. The following table presents segmentation accuracy and resource usage for existing lightweight models and UltraLBM-UNet variants:

Model	ISIC17 IoU/DSC	ISIC18 IoU/DSC	PH² IoU/DSC	Params (M)	FLOPs (G)
MAL-UNet	78.71/88.09	79.42/88.53	83.83/91.20	0.178	0.083
EGE-UNet	78.32/87.84	79.45/88.55	83.36/90.93	0.053	0.072
UltraLight VM-UNet	77.93/87.59	78.93/88.23	83.31/90.89	0.045	0.069
UltraLBM-UNet-T	78.57/88.00	78.82/88.15	84.92/91.85	0.011	0.019
UltraLBM-UNet	79.82/88.78	79.94/88.85	84.41/91.54	0.034	0.060

On ISIC2017, UltraLBM-UNet exhibits the highest IoU (79.82%) and DSC (88.78%) among all sub-1M parameter models. On PH², the distilled student model achieves the top average IoU/DSC (84.92%/91.85%) (Fan et al., 25 Dec 2025). This suggests that ultra-compact architectures can match or surpass larger models in segmentation fidelity when equipped with efficient context modeling and knowledge transfer strategies.

7. Deployment Advantage and Design Rationale

UltraLBM-UNet is explicitly designed for point-of-care scenarios requiring high throughput, low latency, and minimal compute resources. Model weights of <0.14 MB and GFLOPs of ≤0.06 enable stable operation on low-power devices, including embedded platforms and mobile processors. Absence of matrix-heavy self-attention ensures deterministic and predictable resource usage. Multi-branch identity shortcuts stabilize gradients and preserve detail, while learnable skip-scales and fixed branching favour fixed-point inferencing (e.g., for INT8 quantization).

Balanced fusion between bidirectional Mamba-based global modeling and multi-kernel depthwise convolution supports robust edge detection and contextual reasoning. Shared Mamba weights maintain model simplicity despite bidirectional traversal. Hybrid knowledge distillation further compresses the student model without runtime trade-off, transferring structural, spatial, and boundary cues from the teacher.

UltraLBM-UNet thus establishes a new Pareto frontier for model accuracy versus resource cost in skin lesion segmentation, specifically tailored for real-time clinical and embedded deployment contexts (Fan et al., 25 Dec 2025).