MedLiteNet: Efficient Hybrid Segmentation

• MedLiteNet is a lightweight hybrid CNN–Transformer model designed for efficient skin lesion segmentation on small-sample dermoscopic datasets.
• It combines convolution-based local feature extraction with a bottleneck-located Transformer for global context, maximizing accuracy while keeping parameter counts low.
• The architecture integrates boundary-aware attention and ASPP modules to sharpen contours and aggregate multi-scale features, ensuring robust performance even in low-contrast settings.

MedLiteNet is a lightweight hybrid CNN–Transformer model for skin-lesion segmentation in dermoscopic images, introduced in "MedLiteNet: Lightweight Hybrid Medical Image Segmentation Model" (Yu et al., 3 Sep 2025). It is designed for small-sample medical datasets in which lesions exhibit low contrast, irregular shapes, artifacts such as hair and rulers, and fuzzy boundaries. The model combines convolutional feature extraction for local detail encoding with Transformer-based global context modeling, while explicitly seeking to avoid CNNs’ limited receptive fields and Transformers’ quadratic complexity and high parameter counts. Its stated design goals are lightweight efficiency, high accuracy on small-sample medical data, hierarchical feature extraction, multi-scale context aggregation, and explicit boundary sharpening through boundary-aware attention (Yu et al., 3 Sep 2025).

1. Problem setting and design rationale

MedLiteNet is situated in the setting of dermoscopic skin-lesion segmentation, where accurate delineation is technically difficult because lesion appearance varies widely, contrast to surrounding skin can be low, and lesion borders are often blurred or irregular (Yu et al., 3 Sep 2025). The paper identifies a methodological tension between two dominant paradigms. Pure CNN architectures, including U-Net and its variants, encode local features effectively but have limited receptive fields, which makes long-range dependency modeling difficult and risks loss of fine-grained spatial information during downsampling. Pure Transformer architectures capture global context through self-attention, but their quadratic scaling with resolution, dependence on large parameter budgets, and typical reliance on pretraining are ill-suited to small datasets such as ISIC (Yu et al., 3 Sep 2025).

The architectural response is a hybridization strategy centered on the encoder bottleneck. Rather than applying self-attention throughout high-resolution stages, MedLiteNet places a global Transformer at the bottleneck, where spatial resolution is reduced and attention cost is therefore manageable. Around this core, the model uses a depth-wise Mobile Inverted Bottleneck encoder for local texture preservation, a bottleneck-level cross-scale local–global token-mixing block for interaction between convolutional and Transformer streams, a boundary-aware attention mechanism for contour sharpening, and ASPP plus SCSE in the decoder for multi-scale aggregation and fine-grained refinement (Yu et al., 3 Sep 2025).

This suggests that MedLiteNet is best understood not as a generic CNN–Transformer fusion, but as a compute-constrained lesion-segmentation design in which every major component is placed so as to preserve accuracy while minimizing the overhead typically associated with Transformer-heavy hybrids.

2. Network topology and constituent modules

The overall pipeline follows an encoder–bottleneck–decoder structure with skip connections akin to U-Net (Yu et al., 3 Sep 2025). The encoder extracts hierarchical features via MBConv blocks, introduces a global Transformer at the bottleneck, and then applies a local–global fusion block. ASPP follows to capture multi-scale context, after which the decoder progressively upsamples, fuses with encoder features through skip connections, and applies SCSE recalibration and boundary-attention refinement to generate the final lesion mask (Yu et al., 3 Sep 2025).

For an input of size $3\times256\times256$, downsampling begins with a stride-2 $3\times3$ convolution, followed by four MBConv stages with channel widths $[32, 64, 128, 256]$ (Yu et al., 3 Sep 2025). In each stage, the first block uses stride-2 and the remaining blocks use stride-1 for refinement. The convolutional operators are MobileNetV2-style inverted residual MBConv blocks comprising depthwise separable $3\times3$ convolution in the expanded channel space, together with pointwise $1\times1$ expansions and projections. The expansion ratio is $t=6$, and the paper states that BatchNorm and SiLU are used for all convolutions (Yu et al., 3 Sep 2025).

The Transformer operates on a bottleneck feature map

$F \in \mathbb{R}^{H \times W \times C},$

which is flattened into $N=H\times W$ tokens, projected to a $d$-dimensional embedding, and augmented with positional encoding. $L$ stacked Transformer layers consisting of MHSA and FFN blocks with residual connections and LayerNorm produce

$3\times3$0

which is reshaped to $3\times3$1, where $3\times3$2 is the bottleneck downsampling factor (Yu et al., 3 Sep 2025). The paper explicitly states that full MHSA is used at the bottleneck to keep complexity manageable, and does not specify window sizes, number of heads, or $3\times3$3 (Yu et al., 3 Sep 2025).

A local–global fusion block exchanges information between the convolutional stream and the Transformer stream. The paper gives the fusion as

$3\times3$4

where $3\times3$5 denotes channel-wise concatenation, $3\times3$6 and $3\times3$7 are the $3\times3$8 convolution weights and bias, and $3\times3$9 is a nonlinearity. The text suggests ReLU in this fusion, although SiLU is used elsewhere for convolutions (Yu et al., 3 Sep 2025). Functionally, this block performs lightweight pixel-wise aggregation of local and global features without introducing additional quadratic attention at higher resolutions.

3. Boundary-aware refinement and decoder operations

A distinguishing component of MedLiteNet is its boundary-aware attention module, abbreviated BAA in the paper (Yu et al., 3 Sep 2025). The paper does not implement boundary awareness by modifying MHSA’s $[32, 64, 128, 256]$0, $[32, 64, 128, 256]$1, and $[32, 64, 128, 256]$2 terms. Instead, it uses a boundary-aware attention gate that computes a boundary response map and uses that map to reweight features (Yu et al., 3 Sep 2025).

The mechanism is specified by the following equations. Boundary feature extraction is represented as

$[32, 64, 128, 256]$3

and optionally

$[32, 64, 128, 256]$4

where $[32, 64, 128, 256]$5 is a learnable matrix. The attention mask is then computed as

$[32, 64, 128, 256]$6

and feature refinement is

$[32, 64, 128, 256]$7

with $[32, 64, 128, 256]$8 denoting element-wise multiplication (Yu et al., 3 Sep 2025). According to the paper, BAA is inserted near the end of the encoder and at high-resolution stages in the decoder in order to improve contour fidelity (Yu et al., 3 Sep 2025).

At the bottleneck, MedLiteNet also incorporates ASPP for multi-scale context. The paper describes four parallel $[32, 64, 128, 256]$9 dilated convolutions with dilation rates $3\times3$0, together with a global average pooling branch. These branches are adjusted by $3\times3$1 convolutions, upsampled to a common resolution, concatenated, and fused through a $3\times3$2 convolution to produce a multi-scale enhanced map $3\times3$3 (Yu et al., 3 Sep 2025). ASPP is explicitly placed at low resolution to keep computation modest while increasing robustness to lesions of varying sizes.

The decoder is symmetric and U-Net-style, progressively upsampling and merging with encoder skips at matching resolutions (Yu et al., 3 Sep 2025). The paper states that it upsamples but does not specify whether the operator is bilinear interpolation, transposed convolution, or another mechanism. Boundary attention modules are added in each decoding stage except the final one, and SCSE recalibration is used to emphasize salient features while suppressing noise (Yu et al., 3 Sep 2025). The final output layer is a $3\times3$4 convolution followed by sigmoid to produce a probability map, and thresholding at $3\times3$5 yields the binary mask $3\times3$6 (Yu et al., 3 Sep 2025).

Visual analyses in the paper report high agreement with ground truth in shape and contour alignment, and attribute improved boundary fidelity to the BAA and SCSE-enhanced decoder (Yu et al., 3 Sep 2025). Failure cases nonetheless remain at extremely ambiguous borders and in the presence of artifacts such as hair and rulers.

4. Objective functions, optimization, and training configuration

MedLiteNet is trained using a weighted combination of Dice loss and binary cross entropy. The paper gives Dice loss as

$3\times3$7

where $3\times3$8 is the predicted probability, $3\times3$9 is the ground-truth label, $1\times1$0 is a smoothing term, and $1\times1$1 is the number of pixels (Yu et al., 3 Sep 2025). Dice is motivated as directly optimizing overlap while counteracting foreground–background class imbalance. Binary cross entropy is given as

$1\times1$2

The total loss is

$1\times1$3

with $1\times1$4 in the reported experiments (Yu et al., 3 Sep 2025). The paper explicitly states that it did not include a separate boundary-aware loss, and that this mixture empirically outperformed single-loss or more complex schemes.

The dataset is ISIC 2018 segmentation, described as having approximately $1\times1$5 training images with dermatologist-annotated lesion masks and resolutions ranging from about $1\times1$6 to larger formats (Yu et al., 3 Sep 2025). The official train/test split is used. Inputs are cropped or resized to $1\times1$7 for training, while $1\times1$8 is referenced for latency measurements (Yu et al., 3 Sep 2025).

Pre-processing uses ImageNet statistics for normalization. Augmentations include random horizontal and vertical flips, rotations, elastic deformation, scaling, translation, brightness–contrast adjustments, CLAHE, gamma modifications, and Gaussian noise and blur injections (Yu et al., 3 Sep 2025). Training is implemented in PyTorch on a single NVIDIA RTX A6000 GPU with batch size $1\times1$9, AdamW, an initial learning rate of $t=6$0, and cosine annealing to near zero (Yu et al., 3 Sep 2025). Stability and generalization measures include gradient clipping with max-norm $t=6$1, automatic mixed precision using FP16/FP32 AMP, EMA with decay $t=6$2, and gradient accumulation with updates every two iterations (Yu et al., 3 Sep 2025). The text mentions $t=6$3 epochs, whereas the hyperparameter table lists $t=6$4; both are reported in the paper (Yu et al., 3 Sep 2025). The paper also highlights progressive size-increment training and FP16 inference in its contribution summary.

At inference time, the model uses a sigmoid output with threshold $t=6$5, and optionally six-fold test-time augmentation for uncertainty estimation (Yu et al., 3 Sep 2025). A plausible implication is that the reported deployment behavior spans both strict real-time single-model settings and slower but potentially more robust ensemble or TTA configurations.

5. Computational profile and reported empirical performance

The efficiency profile of MedLiteNet is central to its positioning. The paper reports approximately $t=6$6M parameters in the encoder and approximately $t=6$7–$t=6$8M parameters for the full model (Yu et al., 3 Sep 2025). Raw inference latency is reported as about $t=6$9 ms per $F \in \mathbb{R}^{H \times W \times C},$0 image on an NVIDIA RTX A6000, increasing to about $F \in \mathbb{R}^{H \times W \times C},$1 ms with six-fold test-time augmentation (Yu et al., 3 Sep 2025). The paper emphasizes that full MHSA is confined to the bottleneck so that the quadratic token–token cost remains manageable, and it states that depth-wise separable MBConv reduces both parameters and multiply–adds nearly by an order of magnitude at $F \in \mathbb{R}^{H \times W \times C},$2 and $F \in \mathbb{R}^{H \times W \times C},$3 (Yu et al., 3 Sep 2025).

The reported evaluation metrics are Dice and IoU. The paper notes that sensitivity and specificity can be misleading because of class imbalance. It provides the metric relationship

$F \in \mathbb{R}^{H \times W \times C},$4

and

$F \in \mathbb{R}^{H \times W \times C},$5

with $F \in \mathbb{R}^{H \times W \times C},$6 and $F \in \mathbb{R}^{H \times W \times C},$7 denoting the predicted and ground-truth sets (Yu et al., 3 Sep 2025).

The quantitative results reported in the paper are not presented as a single value pair, but as several closely related summaries tied to experimental setting. The abstract reports, for a single model on ISIC 2018, Dice $F \in \mathbb{R}^{H \times W \times C},$8 and IoU $F \in \mathbb{R}^{H \times W \times C},$9 with fewer than $N=H\times W$0M parameters (Yu et al., 3 Sep 2025). The abstract also reports an ensemble of three variants with Dice $N=H\times W$1, IoU $N=H\times W$2, and total parameters below $N=H\times W$3M (Yu et al., 3 Sep 2025). In the body text, the “final integrated test” is described as achieving a validation-set Dice of $N=H\times W$4 and IoU $N=H\times W$5, together with a test-set Dice of $N=H\times W$6 and IoU $N=H\times W$7 (Yu et al., 3 Sep 2025). The paper attributes these differences to splits, validation/test reporting, and ensemble or TTA usage.

The following summary reproduces the representative hybrid-model comparison stated in the paper:

Model	Parameters	Dice / IoU
TransUNet (2021)	105M	88.5% / 83.7%
FAT-Net (2022)	28M	89.0% / 80.2%
BACANet (2024)	7.56M	92.1% / 85.4%
MedLiteNet (2025)	3.2M	90.5% / 83.0%

Within this comparison, MedLiteNet is presented as substantially smaller than Transformer-heavy baselines while maintaining competitive segmentation accuracy (Yu et al., 3 Sep 2025). The paper further states that it has more than $N=H\times W$8 fewer parameters than common ViT backbones, although exact FLOPs and memory footprint are not reported.

6. Interpretation, limitations, and deployment relevance

The paper qualitatively attributes performance gains to the local–global fusion block, the boundary-aware attention module, and ASPP (Yu et al., 3 Sep 2025). It specifically notes that ASPP consistently improves Dice and IoU, especially for very small or very large lesions, although exact ablation deltas and component-wise compute costs are not tabulated (Yu et al., 3 Sep 2025). Qualitative analyses indicate strong performance on irregular borders, low-contrast regions, and multi-scale lesions. Reported failure modes include minor under-segmentation or over-segmentation at fuzzy boundaries and occasional false positives on hair or ruler marks (Yu et al., 3 Sep 2025).

The limitations section identifies three recurring issues: boundary precision can remain suboptimal in highly ambiguous or complex morphologies; sensitivity in very low-contrast regions is imperfect, particularly in illuminated or shadowed areas; and robustness to non-lesion artifacts such as hair and rulers is incomplete (Yu et al., 3 Sep 2025). Proposed future directions include preprocessing for artifact removal, hybrid edge operators combined with deep features, targeted augmentations for low-contrast scenarios, and attention variants tailored for subtle contrast and thin structures (Yu et al., 3 Sep 2025).

In deployment terms, the paper explicitly describes MedLiteNet as well-suited to resource-constrained devices and point-of-care settings because of its approximately $N=H\times W$9M parameters and approximately $d$0 ms latency at $d$1 resolution (Yu et al., 3 Sep 2025). FP16 inference and ensemble or TTA options are presented as permitting flexible accuracy–latency trade-offs. The architecture is stated to be compatible with quantization, pruning, and distillation, although no such experiments are reported (Yu et al., 3 Sep 2025). For strict real-time use, the paper recommends favoring single-model FP16 inference without TTA and considering moderate input resolutions such as $d$2–$d$3, depending on device capability and required precision (Yu et al., 3 Sep 2025).

Reproducibility is partially supported by the provision of implementation details, including framework, optimizer, learning-rate schedule, AMP, EMA, gradient clipping, and gradient accumulation (Yu et al., 3 Sep 2025). However, the paper does not provide code repository links, pretrained weights, or licensing information. This suggests that independent reproduction is feasible at the training-pipeline level, but not turnkey.

7. Position within lightweight medical image segmentation research

MedLiteNet occupies a specific point in the design space of medical image segmentation: a compact hybrid that retains U-Net-like encoder–decoder organization, adopts MobileNetV2-style MBConv for efficient local representation, and introduces Transformer processing only at the bottleneck to control the cost of self-attention (Yu et al., 3 Sep 2025). In contrast to heavier CNN–Transformer hybrids, its principal claim is not absolute state-of-the-art accuracy, but an accuracy–efficiency trade-off tailored to small-sample dermoscopic segmentation (Yu et al., 3 Sep 2025).

The paper’s own comparison makes this positioning explicit. TransUNet and other Transformer-heavy baselines are much larger; BACANet achieves higher absolute scores, but with a larger parameter count; MedLiteNet remains markedly smaller while sustaining Dice around $d$4–$d$5 and IoU around $d$6 on ISIC 2018 (Yu et al., 3 Sep 2025). The paper therefore frames the model as suitable when clinical-grade lesion masks are needed under tight compute and memory budgets, particularly in mobile, edge, or real-time dermatology tools and in settings where very large Transformers may overfit or be impractical (Yu et al., 3 Sep 2025).

A common misconception in reading hybrid architectures is to assume that any incorporation of Transformer blocks implies a globally attention-dominant design. MedLiteNet does not fit that pattern. Its Transformer is deliberately small and bottleneck-localized, and its boundary mechanism is not a modified global self-attention kernel but a gating formulation that reweights features using a boundary response map (Yu et al., 3 Sep 2025). Another possible misconception is that the model introduces a boundary-aware loss; the paper states that it does not, relying instead on a weighted mixture of Dice loss and BCE (Yu et al., 3 Sep 2025). These distinctions are important for interpreting both its empirical profile and its intended use cases within lightweight medical image segmentation.