LFA-Net: Lightweight Retinal Segmentation

• LFA-Net is a lightweight retinal vessel segmentation architecture that integrates multiscale convolution with selective region-aware and LiteFusion-Attention mechanisms.
• The design features a compact U-Net-like encoder–decoder with three encoder stages and refined skip connections to effectively capture thin vessel details.
• It achieves competitive Dice and Jaccard scores with only 0.11M parameters and 4.46 GFLOPs, ensuring strong performance in resource-constrained clinical environments.

Searching arXiv for the exact papers needed to ground the article and disambiguate the term “LFA-Net.” arXiv search query: "LFA-Net retinal vessel segmentation (Mehmood et al., 26 Sep 2025)" LFA-Net denotes “LFA-Net: A Lightweight Network with LiteFusion Attention for Retinal Vessel Segmentation”, a lightweight encoder–decoder model for pixelwise vessel extraction from retinal fundus images. It was introduced by Mehwish Mehmood, Ivor Spence, and Muhammad Fahim for the setting in which segmentation quality on thin vessels must be balanced against strict limits on parameters, memory, and compute. In the reported configuration, LFA-Net combines multiscale convolutional encoding, Region-Aware Attention (RAA) in selective skip pathways, and a custom bottleneck module termed LiteFusion-Attention, which the manuscript describes as incorporating residual learning connections, modulation-based attention, and Vision Mamba-inspired dynamics (Mehmood et al., 26 Sep 2025).

1. Definition, scope, and problem setting

LFA-Net is a retinal vessel segmentation architecture. Its task is to classify each pixel in a fundus image as vessel or background, with particular emphasis on small vessel segmentation and computational efficiency. The paper frames the problem around two persistent constraints: thin vessels are difficult to preserve under lightweight architectures, while heavier attention-based or transformer-style systems often exceed the computational budget of real-world clinical deployment (Mehmood et al., 26 Sep 2025).

The motivation is explicitly clinical. The manuscript associates retinal vessel morphology with early signs of diabetic retinopathy, hypertensive retinopathy, glaucoma, cardiovascular disease, and neurological disease. At the same time, it argues that deployment settings such as mobile ophthalmology units, point-of-care devices, and low-resource clinics require models that are not only accurate but also compact. The reported headline complexity figures are 0.11 million parameters, 0.42 MB model size, and 4.46 GFLOPs (Mehmood et al., 26 Sep 2025).

Architecturally, LFA-Net is described as a compact U-Net-like encoder–decoder rather than a standard pretrained backbone. The text states that the network uses three encoder stages, selective skip refinement, and a bottleneck attention block. The manuscript does not provide a full stage-wise channel table, so the exact per-level feature widths are unspecified (Mehmood et al., 26 Sep 2025).

2. Network topology and feature flow

The encoder is built from multiscale convolution blocks. At each stage, the input is processed in parallel by a $1\times1$ convolution, a $3\times3$ convolution, and a dilated $3\times3$ convolution, after which the outputs are concatenated and passed through normalization, pooling, and nonlinearity (Mehmood et al., 26 Sep 2025). The encoder equations are given as

$$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$

$C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$

where $X_k$ is the input to stage $k$, $C_{di}^{3\times3}$ denotes the dilated branch, and $\oplus$ denotes channel concatenation. The manuscript refers to $C_1$, $3\times3$0, and $3\times3$1, indicating three encoder stages before the bottleneck (Mehmood et al., 26 Sep 2025).

At the bottleneck, the deepest encoder representation $3\times3$2 is processed by LiteFusion-Attention and then by Region-Aware Attention, after which it is concatenated with the original feature map:

$3\times3$3

This formulation indicates that bottleneck processing is not purely replacement-based; it retains an explicit residual-style path through concatenation with $3\times3$4 (Mehmood et al., 26 Sep 2025).

The decoder uses transposed convolutions for upsampling and merges the upsampled features with RAA-refined skip features from the encoder. The stage equations are reported as

$3\times3$5

$3\times3$6

$3\times3$7

$3\times3$8

where $3\times3$9 is sigmoid and $3\times3$0 is the final vessel probability map (Mehmood et al., 26 Sep 2025). The manuscript’s notation is somewhat inconsistent in the sequencing of $3\times3$1 and $3\times3$2, but the intended structure is clear: a shallow encoder–decoder with selective skip refinement and a bottleneck attention block.

3. LiteFusion-Attention and Region-Aware Attention

The central novelty of LFA-Net is LiteFusion-Attention, a bottleneck module intended to combine local and global context with low overhead. The paper describes the module as using residual learning, modulation-based attention, token mixing, channel mixing, and Vision Mamba-inspired dynamics (Mehmood et al., 26 Sep 2025). A cautious reading is that the “Vision Mamba-inspired” characterization is conceptual rather than canonical, because the equations shown are not those of a standard state-space Mamba block.

The first stage of LiteFusion projects and refines the bottleneck feature:

$3\times3$3

A global context attention branch then computes channel-like weights:

$3\times3$4

In parallel, a spatial filtering branch produces

$3\times3$5

and the two branches are combined as

$3\times3$6

where $3\times3$7 denotes element-wise multiplication (Mehmood et al., 26 Sep 2025). This implements a lightweight global-to-local modulation step.

A second modulation stage uses the difference between max-pooled and average-pooled descriptors:

$3\times3$8

$3\times3$9

$$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$0

$$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$1

with $$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$2 and $$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$3 (Mehmood et al., 26 Sep 2025). The manuscript interprets this as a focal-style enhancement step. Subsequent equations define fusion, token mixing, and channel mixing, but the notation is partially corrupted in the text. The reported forms are

$$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$4

$$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$5

$$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$6

$$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$7

$$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$8

$$C_{ms_k} = C^{1\times1}(X_k) \oplus C^{3\times3}(X_k) \oplus C_{di}^{3\times3}(X_k)$$9

The paper presents these operations as the efficient mechanism by which local and global context are fused without full transformer self-attention (Mehmood et al., 26 Sep 2025).

The second attention mechanism is Region-Aware Attention. It is used in selective skip pathways and also appears in the bottleneck formulation. Given an input tensor $C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$0, the module begins with local refinement

$C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$1

followed by multi-scale pooled summaries

$C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$2

which are combined as

$C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$3

The module then defines a channel-averaged semantic summary and attention map

$C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$4

and modulates the input via

$C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$5

The paper states that $C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$6 is the number of feature channels required for each pixel to have differentiating areas, and $C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$7 is the number of classes (Mehmood et al., 26 Sep 2025). In operational terms, RAA is intended to improve the spatial precision of the skip information before decoder fusion.

4. Training protocol, data, and supervision

LFA-Net is evaluated on the three standard retinal vessel datasets DRIVE, STARE, and CHASE_DB (Mehmood et al., 26 Sep 2025). The reported dataset details are: DRIVE with 40 total images and a 20/20 train/test split at original resolution 565 × 584; STARE with 20 images and a 16/4 split at 700 × 605; and CHASE_DB with 28 images and a 20/8 split at 1024 × 1024. All images are resized to 512 × 512 (Mehmood et al., 26 Sep 2025).

The manuscript also states that 80% and 20% of the images were used for model training and validation from each dataset, which is slightly inconsistent with the tabulated train/test counts. The most conservative interpretation is that the paper reports official or practical train/test splits while implying an internal validation partition within the training data, but this is not fully clarified (Mehmood et al., 26 Sep 2025).

Data augmentation is explicitly limited to rotation by 20 degrees and contrast modifications, implemented using CLoDSA and IMGAUG. The augmented image counts are reported as 1080 for DRIVE, 1024 for STARE, and 1080 for CHASE_DB (Mehmood et al., 26 Sep 2025).

The implementation uses TensorFlow and Keras on an NVIDIA RTX A4000 GPU with 16 GB GDDR6 VRAM. Optimization uses Adam with learning rate 0.002 and batch size 8 (Mehmood et al., 26 Sep 2025). The manuscript does not specify the number of epochs, any learning-rate schedule, weight decay, or a detailed preprocessing pipeline beyond resizing and augmentation.

The loss is weighted Dice loss, written as

$C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$8

where $C_k = \mathit{LeakyReLU}(M_p(BN(C_{ms_k})))$9 is the predicted segmentation, $X_k$0 is the ground truth, $X_k$1 is the class weight, $X_k$2 is the number of classes, $X_k$3 is the number of pixels, and $X_k$4 is a smoothing constant (Mehmood et al., 26 Sep 2025). The accompanying prose in the paper appears to reverse the roles of $X_k$5 and $X_k$6, but the equation itself makes the intended interpretation evident.

5. Reported efficiency, benchmark results, and ablations

The efficiency profile is one of the paper’s defining claims. In the comparison table, LFA-Net is reported with 0.11 M parameters, 4.46 G FLOPs, and 0.42 MB size, versus 0.87 M / 47.6 G / 3.50 MB for FS-UNet, 7.76 M / 96.68 G / 29.60 MB for U-Net, and 9.25 M / 371.68 G / 35.33 MB for Attention U-Net (Mehmood et al., 26 Sep 2025). This positions the model as substantially smaller than both standard and previously lightweight baselines.

The manuscript contains a discrepancy between the abstract and Table 3 for the main overlap metrics. The reported values are as follows:

Dataset	Abstract Dice / Jaccard	Table 3 Dice / Jaccard
DRIVE	83.28 / 72.85	83.18 / 71.24
STARE	87.44 / 79.31	87.16 / 77.29
CHASE_DB	84.50 / 74.70	84.05 / 72.52

On DRIVE, Table 3 reports Dice 83.18, Jaccard 71.24, Sensitivity 80.56, and Specificity 98.09, which the text describes as the best Dice and Jaccard among the listed methods (Mehmood et al., 26 Sep 2025). On STARE, Table 3 reports Dice 87.16, Jaccard 77.29, Sensitivity 88.13, and Specificity 98.30, with a substantial margin over the listed baselines in Dice, Jaccard, and Sensitivity (Mehmood et al., 26 Sep 2025). On CHASE_DB, Table 3 reports Dice 84.05, Jaccard 72.52, Sensitivity 82.25, and Specificity 98.15, again with the best Dice and Jaccard among the compared methods (Mehmood et al., 26 Sep 2025).

The ablation study isolates the contributions of multiscale encoding, skip connections, RAA, LiteFusion, and attention placement. The baseline Lightweight UNet no-skip connection (LU-NS) achieves Dice 80.30 and J 64.11, while Multiscale LU skip connection (MLU) improves to Dice 81.32 and J 66.99 (Mehmood et al., 26 Sep 2025). Adding only RAA in skip connections gives Dice 81.38 / J 68.69, and adding only LiteFusion at the bottleneck gives Dice 81.61 / J 69.11 (Mehmood et al., 26 Sep 2025). Combining them yields stronger gains, and the best reported configuration is

MLU + $X_k$7 in 1,2-Skip + $X_k$8-Bottleneck + $X_k$9-Bottleneck

with 0.11 M parameters, Dice 83.18, J 71.24, Acc 96.09, Sen 80.56, and Sp 98.09 (Mehmood et al., 26 Sep 2025). This directly supports the paper’s claim that attention should be used selectively, with the most effective skip placement occurring at the first and second skip connections rather than uniformly across all possible skip levels.

Qualitatively, the paper states that LFA-Net produces fewer false positives and false negatives, more continuous vascular trees, and better recovery of thin branches than the compared methods (Mehmood et al., 26 Sep 2025). The manuscript does not, however, provide a dedicated thin-vessel subset metric or a formal robustness analysis.

6. Naming ambiguities, related acronyms, and reporting limitations

The term “LFA-Net” is unusually prone to acronym confusion. In the current literature snapshot, several nearby names can be mistaken for it, but they are distinct models.

Name	Expansion	Relation to “LFA-Net”
LFA in RandLA-Net	Local Feature Aggregation	Module inside RandLA-Net, not a standalone network
LFRA-Net	Lightweight Focal and Region-Aware Attention Network	Different retinal vessel model with a related naming pattern
LFINet	Laplacian Frequency Interaction Network	Different model for rural road extraction
LF-Net	Learning Local Features from Images	Distinct local-feature pipeline
LFNet	Lattice Fusion Network	Distinct denoising architecture

The most important disambiguation is with RandLA-Net, where LFA denotes Local Feature Aggregation, a module used to compensate for the information loss of random sampling in large-scale 3D point cloud segmentation; the paper does not name the overall network “LFA-Net” (Hu et al., 2019). Confusion is also plausible with LFRA-Net, another lightweight retinal vessel segmentation model that uses focal modulation attention at the bottleneck and region-aware attention in skip connections, but whose formal name is LFRA-Net, not LFA-Net (Mehmood et al., 15 Sep 2025). Additional acronym-neighbor confusion can arise with LFINet for rural thematic road extraction (Chen et al., 4 May 2026), LF-Net for learned local image features (Ono et al., 2018), and LFNet for lattice fusion denoising (Hosseini, 2020).

Within its own manuscript, LFA-Net also has several reporting limitations. The paper does not specify the number of epochs, the learning-rate schedule, weight decay, precise feature dimensions, or measured runtime such as latency or FPS (Mehmood et al., 26 Sep 2025). It further does not isolate the internal contribution of each LiteFusion subcomponent, such as residual paths, modulation, token mixing, or channel mixing, in standalone ablations (Mehmood et al., 26 Sep 2025). The “Vision Mamba-inspired” description therefore remains architectural and qualitative rather than being tied to a canonical state-space formulation or to a separate component-level empirical analysis.

In summary, LFA-Net is best understood as a very compact retinal vessel segmentation network whose defining contribution is the use of LiteFusion-Attention at the bottleneck together with Region-Aware Attention in selected skips. Its significance lies less in introducing a wholly new segmentation paradigm than in demonstrating that carefully placed lightweight attention can deliver strong Dice and Jaccard performance under a computational budget of 0.11 M parameters, 0.42 MB, and 4.46 GFLOPs on DRIVE, STARE, and CHASE_DB (Mehmood et al., 26 Sep 2025).