EA-LiteUNet: An Edge-Adaptive and Resource-Efficient U-Net for Boundary-Sensitive Dermoscopic Image Segmentation

2.1 Boundary-Oriented Segmentation Strategies

• Boundary strategies use auxiliary supervision, boundary-sensitive losses, and contour-guided features, but often leave early feature degradation and downsampling-induced aliasing unaddressed.Prior approaches include edge-detection branches and contour-prediction modules; high-frequency boundary distortions can propagate through the network and limit supervision-driven enhancement.Source: §2.1 · p. 3

2.2 Attention-Guided Feature Modulation in Lightweight U-Net Variants

• Attention modules emphasize informative features and regulate skip connections, but often lack boundary or prediction awareness and add computational overhead.The related work motivates lightweight, prediction-aware modulation to refine boundary-sensitive features.Source: §2.2 · pp. 3–4

2.3 Efficient and Adaptive Inference Mechanisms

• Efficiency approaches include lightweight U-Nets, efficient convolutions, compact designs, early exits, and dynamic depth control, yet rarely adapt inference to localized boundary complexity.Fixed pathways may waste computation on simple regions or provide insufficient contour refinement on complex margins.Source: §2.3 · p. 4

2.4 Summary and Motivation

• The identified gap is that few methods jointly combine boundary-aware representation learning, selective feature modulation, and resource-adaptive inference for boundary-sensitive segmentation.The paper proposes an edge-adaptive, computationally efficient architecture for practical resource constraints.Source: §2.4 · p. 4

3.1 Overall Architecture

• The six-stage encoder–decoder uses channel widths {8, 16, 24, 32, 48, 64}, stage-dependent operations, attention-guided skip connections, and adaptive decoder refinement.Shallow stages use edge-adaptive downsampling, while deeper stages retain max-pooling; the decoder adjusts refinement depth according to prediction stability and structural complexity.Source: §3.1 · pp. 4–6

3.2 Efficient Boundary-Aware Representation Module (EBARM)

• EBARM divides Layers 1–3 into a boundary-sensitive stage and Layers 4–6 into a semantic-context stage, using HybridConv and anti-aliasing downsampling early to preserve fine boundary structure.HybridConv combines depthwise separable convolutions with efficient channel attention; AA-Down uses BlurPool before subsampling to suppress high-frequency artifacts.Source: §3.2 · pp. 6–7

3.2.2 Deep Stage (Layers 4-6): Progressive Large-Kernel Context Modeling.

• The deep-stage PLK-DW module increases kernel size across Layers 4–6 to expand the receptive field as resolution decreases, capturing global context in Stage 6.Stage 4 transitions from local to semi-global features, while standard max-pooling is retained for semantic abstraction.Source: §3.2 · p. 6, §3.2.2 · p. 7

3.3 Attention-Guided Modulation Mechanism (AGMM)

• AGMM replaces passive skip transfer with LDDA and PIF-V2, combining attention-based feature refinement with prediction-guided regulation of encoder-to-decoder information flow.PIF-V2 uses prediction-derived cues to adaptively regulate feature transfer.Source: §3.3 · p. 7, Figure 4 · p. 8

3.3.1 Lightweight Dual-Domain Attention (LDDA)

• LDDA sequentially applies channel-wise and spatial attention to regulate encoder skip-feature responses relevant to boundary delineation.The module operates on encoder features at each skip-connection stage.Source: §3.3.1 · p. 8

(1) Channel-first attention: ECA (Efficient Channel Attention Formula)

• ECA channel weights are broadcast for channel-wise scaling, selectively enhancing boundary-relevant cues while suppressing redundant responses.The passage also introduces global average pooling in the attention formulation, but provides no formula values here.Source: (1) Channel-first attention: ECA… · p. 8

(2) Spatial-next attention

• A lightweight spatial mask is generated from the channel-refined feature, so spatial gating follows channel selection.Source: (2) Spatial-next attention · p. 8
• Element-wise multiplication uses the channel-refined feature rather than the raw input to suppress localized noise and enhance structurally salient regions.Source: (2) Spatial-next attention · p. 8

(3) Dual-domain coupling: channel-then-spatial

• Sequential channel reweighting and spatial gating let LDDA regulate which features matter and where, producing boundary-refined encoder representations with minimal overhead.Depthwise convolution and ECA keep the added computation minimal.Source: (3) Dual-domain coupling:… · p. 9

3.3.2 Prediction Information Fusion (PIF-V2)

• PIF-V2 uses intermediate decoder predictions to guide skip modulation, reinforcing boundary-relevant responses while keeping computation lightweight.Region- and boundary-related attention maps, together with learnable fusion weights and adaptive gating, provide context-aware feature modulation.Source: §3.3.2 · p. 9

3.3.3 Attention-Guided Skip Modulation

• Combining LDDA with PIF-V2 turns skip connections into actively regulated pathways that refine encoder features and integrate prediction-aware cues for boundary-sensitive fusion.The two modules add only marginal computational overhead.Source: §3.3.3 · p. 9

3.4 Boundary-Aware Resource-Adaptive Inference Mechanism

• The resource-adaptive inference mechanism adjusts decoder depth to boundary uncertainty and structural complexity, using confidence-based early termination for stable cases and deeper refinement for ambiguous ones.Its two complementary components are boundary-driven dynamic depth modulation and confidence-guided early termination.Source: §3.4 · pp. 9–10, Figure 5 · p. 10

3.4.1 Confidence-Guided Early-Exit Prediction Heads (EEPH)

• EEPH attaches lightweight prediction heads to intermediate decoder stages and uses pixel-wise entropy to estimate confidence, terminating decoding when accumulated entropy reduction indicates further refinement is unlikely to improve boundary localization.Lower entropy indicates more stable predictions, particularly along boundaries; the confidence threshold is jointly optimized with the network.Source: §3.4.1 · p. 10

3.4.2 Boundary-driven Dynamic Depth Controller (BD-DDC)

• BD-DDC combines prediction uncertainty, boundary-neighborhood confidence, and boundary-attention sparsity into a learned complexity score that determines whether decoding terminates or activates a deeper stage.Boundary confidence focuses on pixels near the predicted contour, while concentrated attention indicates a distinctive boundary and diffuse attention suggests ambiguity.Source: §3.4.2 · p. 11

3.4.3 Hardware-Aware Computational Profiling (HAP)

• HAP profiles per-stage latency and activation memory on target hardware, then caps sample-adaptive decoding depth to the largest depth meeting both device budgets.It can only reduce the controller-selected depth; when hardware is not the binding constraint, the learned early-exit and depth-activation behavior remains unchanged.Source: §3.4.3 · p. 12

4.1 Datasets

• The ISIC2018 dataset contains 2,594 images, while PH2 contains 200 images and is used to assess robustness to acquisition-device and lesion-appearance variation.Source: §4.1 · p. 12
• ISIC2017 contributes 2,000 images to enable fairer comparisons with earlier studies, including U-Net++ and Attention U-Net.Source: §4.1 · p. 12
• Each dataset is divided into training, validation, and test sets in an 8:1:1 ratio.Source: §4.1 · p. 12

4.2 Evaluation Metrics

• Dice similarity coefficient, IoU, sensitivity, and specificity jointly characterize lesion overlap accuracy, detection capability, and background discrimination.Source: §4.2 · p. 12
• HD95 measures maximum prediction-to-ground-truth boundary discrepancy using the 95th percentile of directed distances, reducing sensitivity to small outliers and noise; lower values indicate closer boundary matches.Source: §4.2 · p. 13
• Efficiency is measured by trainable parameter count and GFLOPs, the floating-point operations for one forward pass at 256 × 256 input resolution and batch size 1.Source: §4.2 · p. 13
• Parameters and GFLOPs explicitly represent the trade-off between segmentation performance and computational efficiency.Source: §4.2 · p. 13

• EA-Lite-UNet was implemented in PyTorch and trained on an NVIDIA GeForce RTX 4060 GPU; input images were resized to 256×256 and normalized.Source: §4.3 · p. 13
• Training augmentation used random rotation and flipping, brightness and contrast jitter, and cutout-based regularization to simulate diverse clinical scenarios.Source: §4.3 · p. 13
• Training used batch size 8, AdamW with an initial learning rate of 1×10-3, and a CosineAnnealingLR schedule to decay the learning rate.Source: §4.3 · p. 13
• The segmentation objective combined binary cross-entropy and Dice loss with boundary-aware supervision, while cost deviation loss trained the controller to enforce computational budgets.The total loss weights boundary supervision and FLOPs regulation using λ_b and λ_c, respectively.Source: §4.3 · p. 13
• Controller target budgets were {0.5,0.75,1.0}×CU-Net; it adjusted early-exit and depth-activation thresholds, with results reported across budgets for the accuracy–efficiency trade-off.Source: §4.3 · p. 13
• Hyperparameters were tuned using the ISIC2018 validation set and reused for the remaining settings unless otherwise stated.Source: §4.3 · p. 13

4.4.1 Results on the ISIC2018 Dataset

• 92.08% Dice score, 12.89-pixel HD95, and 94.67% sensitivity were best on ISIC2018, outperforming existing segmentation models.The model also handles lesions of varying sizes and complex backgrounds, producing masks closer to ground-truth annotations.Source: §4.4.1 · p. 14, Table 1 · p. 14

4.4.2 Results on the PH2 Dataset

• 94.17% Dice score, 12.01-pixel HD95, and 95.07% sensitivity were best on PH2; versus LB-UNet, Dice rose 0.18%, sensitivity 0.26%, and HD95 fell 4.93 pixels.Against MAFUNet, HD95 decreased from 18.16 to 12.01 pixels.Source: §4.4.2 · p. 15

4.5.1.(a) Cumulative ablation study on ISIC-2018

• Cumulative improvements validate complementary, non-redundant contributions of the proposed modules; replacing conventional downsampling with AA-Down substantially reduces HD95, while incorporating PLK-DW further increases the Dice score.Source: 4.5.1.(a) Cumulative ablation study on… · p. 18

4.5.1.(b) Training Dynamics Analysis

• Model-7’s Dice trajectory has higher values, reduced variance, and earlier convergence than baseline, while its HD95 falls faster early and reaches lower final convergence.Baseline trajectories show larger oscillations and slower stabilization; the full configuration also has reduced HD95 oscillations.Source: 4.5.1.(b) Training Dynamics Analysis · pp. 18–19, Figure 9 · p. 18

4.5.2 Orthogonal Ablation of LDDA and PIF-V2

• PIF-V2 improves over the EBARM-only Model4 even without LDDA (+0.81 Dice, −2.20px HD95), ruling out under-trained LDDA as the explanation for its gain.The orthogonal test compares neither, LDDA-only, PIF-V2-only, and joint configurations; the joint Model6 achieves the best Dice/HD95 and lowest run-to-run variance among them.Source: §4.5.2 · pp. 19–20

4.5.3.(a) EBARM Internal Components

• Removing HybridConv causes the largest degradation on four of five metrics, whereas removing AA-Down incurs the largest HD95 cost (+2.48px) despite a Dice cost of only −0.26.PLK-DW has the smallest impact across metrics, including ΔHD95 = +1.90px; these results identify HybridConv as broadly critical and AA-Down as boundary-specific.Source: Dice (%) mIoU (%) HD95(px) SPE(%) SEN(%) · p. 21

4.5.3.(b) AGMM Internal Components

• Removing either AGMM component degrades Dice and HD95: LDDA has the larger overall effect (ΔDice = −0.63, ΔHD95 = +3.56), while PIF-V2 is especially important for HD95 (+2.98).Within LDDA, channel attention drives most of the gain.Source: 4.5.3.(b) AGMM Internal Components · p. 22
• Fixing AGMM’s learnable τ and γ to a static sigmoid costs 0.36 pp in Dice and 1.89 px in HD95, showing that adaptive gating sharpness matters.The stated design rationale is that channel reweighting sharpens the feature map before the subsequent spatial mask is learned.Source: 4.5.3.(b) AGMM Internal Components · p. 22
• In AGMM, the combined channel-and-spatial ablation gives ΔHD95 = +2.70 versus +3.56 for LDDA as a whole, indicating synergistic rather than additive attention effects.One attention-component removal costs more than the other in both Dice (−0.30 vs. −0.23) and HD95 (+1.67 vs. +1.03), and boundary protection is attributed to specialized, non-redundant mechanisms across EBARM and AGMM.Source: 4.5.3.(b) AGMM Internal Components · p. 22

4.5.3.(c) BA-RAIM Internal Components

• Removing BD-DDC causes the largest accuracy loss (ΔDice = −0.40, ΔHD95 = +1.98 px, +15.4%) yet gives the fastest inference and lowest FLOPs (0.55G).The authors attribute its accuracy role to boundary-complexity-driven depth activation that safeguards against premature exit on ambiguous boundaries.Source: 4.5.3.(c) BA-RAIM Internal Components · p. 23
• Removing EEPH leaves accuracy essentially unchanged (ΔDice = +0.07, ΔHD95 = −0.33 px) but increases FLOPs from 0.62G to 1.02G and latency by +69.2% GPU and +195.8% CPU.This identifies EEPH as an efficiency-oriented gate; without it, some early termination remains through BD-DDC’s evaluation.Source: 4.5.3.(c) BA-RAIM Internal Components · p. 23
• Removing HAP produces an intermediate change (ΔDice = −0.23, ΔHD95 = +0.78 px; FLOPs +9.7%; latency +2 ms GPU/+4 ms CPU), consistent with its hardware-conditioned depth-cap role.The cap binds only occasionally on a fixed platform rather than serving as a per-sample decision criterion.Source: 4.5.3.(c) BA-RAIM Internal Components · p. 23
• BA-RAIM’s accuracy-protection ranking is BD-DDC > HAP > EEPH, while its compute-savings ranking is EEPH > HAP > BD-DDC, showing complementary roles.BD-DDC is indispensable for preserving BA-RAIM’s boundary accuracy, whereas early exit and depth modulation contribute different compute savings.Source: 4.5.3.(c) BA-RAIM Internal Components · p. 23

Discussion

• Across all three datasets, EA-LiteUNet jointly improves HD95, Dice, and Sensitivity, unlike efficient baselines that trade boundary fidelity or regional accuracy for efficiency.The discussion attributes this joint improvement to explicit regulation of boundary-related signals rather than relying on region-based supervision alone.Source: §5 · pp. 23–24
• EBARM, LDDA–PIF-V2, and BA-RAIM respectively preserve encoded detail, gate skip features, and add decoding depth for ambiguous boundaries; progressive inclusion lowers HD95 and stabilizes boundary-metric convergence.The paper frames these components as representation-, feature-, and inference-level boundary-frequency regulation, while noting that this interpretation is motivated by ablation consistency rather than formal frequency-domain derivation.Source: §5 · p. 24
• EA-LiteUNet achieves its improvements with 0.29M parameters and 1.17 GFLOPs, while resource-adaptive inference can trade accuracy for lower latency under prescribed FLOPs budgets.The discussion presents this efficiency as relevant to resource-constrained clinical or point-of-care devices.Source: §5 · p. 24
• Evaluation is limited to dermoscopic datasets, validation-set-tuned entropy thresholds, and one hardware platform, with no formal significance testing or expert clinical evaluation.These limits leave cross-modality generalization, transfer of adaptive thresholds, hardware-dependent efficiency, and clinical credibility unresolved.Source: §5 · pp. 24–25
• Future work will test cross-modality and multi-organ generalization, learn adaptive-inference criteria, benchmark more edge and mobile hardware, and add statistical and expert clinical evaluation.The authors also propose formal frequency-domain analysis of boundary-signal propagation.Source: §5 · p. 25

6. Conclusion

• The encoder mitigates aliasing, skip connections enhance contour-relevant features, and the decoder adjusts refinement depth by prediction confidence to improve boundary fidelity and region consistency at low computational complexity.These mechanisms regulate boundary-sensitive signals throughout the segmentation pipeline.Source: §6 · p. 25
• On ISIC2018, EA-LiteUNet achieves a Dice score of 92.08% and HD95 of 12.89 pixels with 0.29M parameters and 1.17 GFLOPs, competitive or superior to recent lightweight and attention-based baselines.Experiments cover ISIC2017, ISIC2018, and PH2; ablation analysis confirms each proposed module contributes complementarily and non-redundantly.Source: §6 · p. 25
• Future work will address cross-modality generalization, learnable adaptive-inference thresholds, and clinical validation to further establish practical utility.The conclusion also suggests the signal-processing perspective may inform efficient, boundary-aware medical image architectures more broadly.Source: §6 · p. 25

Funding

• The authors received no funding for this work.Source: Funding · p. 28