PaperlayerView exact revision on arXiv

Source-linked AI summary

LightMIS: Ultra-Lightweight Medical Image Segmentation Without a Stage-Wise Decoder

Andrei Arhire, Mihaela-Elena Breabăn, Radu Timofte

arXiv:2609.28327v1cs.CV

TL;DR

Compact medical image segmentation must balance local detail and broader context, while many high-performing models remain computationally demanding. LightMIS directly aligns and aggregates a five-level encoder hierarchy without a learned stage-wise decoder. Across six datasets, it achieves modality-macro Dice and IoU close to Mobile U-ViT, with substantially lower reported model complexity.

  • Problem

    High-performing medical image segmentation models can be computationally demanding, while accurate segmentation requires both fine details and broader spatial context.

    Source: §1 · p. 1

  • Method

    LightMIS aligns five encoder levels to a common resolution, aggregates them directly, and refines the result with AFC, including the proposed PRF module.

    Source: §1 · p. 2, §6 · p. 17

  • Results

    86.71% modality-macro Dice and 78.99% modality-macro IoU for LightMIS, versus Mobile U-ViT's 86.75% and 79.07%, respectively.

    Source: §6 · p. 17

  • Takeaways & Limitations

    On an Arm Mali-G52 MC2 GPU, all three LightMIS variants achieved complete GPU delegation, demonstrating on-device execution feasibility for the evaluated input size.

    Source: §6 · p. 17

    The study is restricted to public 2D binary datasets and does not establish performance on multiclass or volumetric segmentation.

    Source: §5 · p. 16

Abstract

from arXiv · show

We present LightMIS, a scalable family of ultra-lightweight convolutional networks for 2D binary medical image segmentation without a learned stage-wise decoder. LightMIS aligns the outputs of a five-level encoder to a common resolution using Scale-Aligned Projection blocks, aggregates them once, and refines the fused representation with an Adaptive Fusion Cascade. The cascade combines Adaptive Kernel Fusion with the proposed Progressive Receptive Fusion module, which uses temporary channel expansion, complementary depthwise receptive fields, and progressive cross-branch information transfer. We evaluate LightMIS-T, LightMIS-S, and LightMIS using five-fold cross-validation under a common nnU-Net v2.3.1 protocol on DRIVE, Kvasir-SEG, DSB18, BUSI, ISIC-2017, and ISIC-2018. Full LightMIS contains 0.131 M parameters and requires 0.575 GFLOPs for a $3\times256\times256$ input, achieving modality-macro Dice and IoU scores of 86.71% and 78.99%, respectively. Mobile U-ViT obtains 86.75% Dice and 79.07% IoU, so the observed differences are 0.04 and 0.08 percentage points. Relative to Mobile U-ViT, nnWNet, and nnU-Net, LightMIS reduces parameter count by 90.58$-$99.61% and GFLOPs by 82.54$-$96.14%. On an Arm Mali-G52 MC2 GPU, all LightMIS variants achieve full GPU delegation, with median delegated latency ranging from 53.31 ms for LightMIS-T to 138.31 ms for LightMIS. These results demonstrate a favorable accuracy$-$complexity trade-off and on-device execution feasibility for the evaluated tasks. The code is publicly available at https://github.com/AndreiiArhire/LightMIS.

1. Introduction

LightMIS targets compact medical-image segmentation without a learned stage-wise decoder, directly aligning and aggregating the encoder hierarchy. The full model ranks second in modality-macro Dice and IoU among evaluated models, with a compact parameter and compute budget.

Motivation

  • Compact models matter because deployment may be challenging in point-of-care systems, mobile platforms, and resource-constrained clinical environments.The introduction also frames segmentation as needing both fine details and broader spatial context.Source: §1 · p. 1

Main contribution and evaluation

  • Full LightMIS achieved the second-highest modality-macro Dice and IoU among evaluated models, with 0.131 M parameters and 0.575 GFLOPs.The evaluation covered six datasets under a unified protocol.Source: §1 · pp. 2–3

Architecture

  • LightMIS replaces stage-wise decoding with direct alignment and aggregation of the complete encoder hierarchy, followed by compact refinement and prediction.The paper describes conventional U-shaped networks as relying on repeated upsampling and cross-scale feature fusion.Source: §1 · p. 2

Results

  • LightMIS achieves the highest mean Dice and IoU on DRIVE and the lowest mean ASSD and HD95 on DSB18.The paper also reports 100% GPU operator coverage on a smartphone with an Arm Mali-G52 MC2 GPU.Source: §1 · p. 2

Proposed modules

  • PRF enriches narrow feature representations through channel expansion, multi-kernel depthwise convolutions, and progressive cross-branch information transfer.AFC combines PRF with an Adaptive Kernel Fusion block adapted from AULUNet and residual adaptive-kernel refinement.Source: §1 · p. 2

Configurations and evaluation

  • LightMIS-T, LightMIS-S, and LightMIS offer different trade-offs between segmentation accuracy and computational efficiency.The three configurations are evaluated across overlap and boundary accuracy, complexity, memory, and desktop and smartphone latency.Source: §1 · p. 3

2. Related Work

Related work spans U-shaped convolutional and Transformer-based models, lightweight segmentation networks, and decoder-free alternatives. LightMIS is positioned as a direct-aggregation architecture evaluated across multiple datasets and imaging settings.

Convolutional and Transformer-based models

  • Convolutional U-Net variants alter skip pathways, attention, multi-scale representation, or contextual modeling while preserving the overall U-shaped organization.The section also describes Transformer and state-space approaches for long-range dependencies.Source: §2 · p. 3

Lightweight segmentation networks

  • Many compact segmentation approaches reduce complexity with efficient operators or feature-processing strategies while retaining U-shaped encoder–decoder structures.Examples include UNeXt, MALUNet, DSNET, LEU-Net, and WIDENet.Source: §2 · pp. 3–4

Decoder-free segmentation

  • Existing decoder-free approaches include multihead feature fusion, attention-based global feature recovery, and patch-merge attention with multi-path aggregation.Their evaluation remains largely focused on retinal vessel segmentation.Source: §2 · p. 4

LightMIS positioning

  • LightMIS aligns all five encoder outputs for direct aggregation, then refines the hierarchy with AFC; its PRF progressively transfers spatial responses between depthwise branches.The same architecture is evaluated without task-specific modifications across six datasets and five imaging settings.Source: §2 · p. 4

3. Methodology

LightMIS uses Scale-Aligned Projection blocks to align a five-level encoder hierarchy for direct aggregation, then refines the fused representation. Its Progressive Receptive Fusion module expands narrow features temporarily and transfers information progressively across receptive-field paths.

3.1. LightMIS Architecture

  • Each of five encoder levels applies AFC, whose first component is AKF adapted from AULUNet with its final ReLU replaced by GELU.AKF output passes through PRF and then residual AKF refinement.Source: §3.1 · p. 4
  • SAP projects each encoder output to shared width F and aligns it to E1's spatial dimensions before aggregation.The aligned representations are concatenated into a 5F-channel tensor, projected back to F channels, and refined by a final AFC.Source: §3.1 · p. 4
  • The three configurations contain 0.017 M, 0.038 M, and 0.131 M parameters and require 0.099, 0.195, and 0.575 GFLOPs, respectively.They retain the same architecture while varying encoder channels and aggregation width.Source: §3.1 · pp. 5–6

3.2. Progressive Receptive Fusion

  • PRF temporarily expands narrow features and divides them among depthwise 3 × 3, dilated 3 × 3, and 5 × 5 spatial paths.The module is intended to enrich compact representations without increasing overall network width.Source: §3.2 · p. 6
  • PRF adds each processed branch's channel-averaged spatial response to the next branch before its convolution, enabling progressive cross-branch interaction.Channel-wise averaging yields a single spatial map broadcast across the next branch's channels, without requiring matching channel dimensions.Source: §3.2 · p. 6
  • PRF concatenates its three branch outputs, projects them to the original channel width, and combines the fused representation with input U through learnable γP.γP is initialized to zero separately for each PRF instance.Source: §3.2 · p. 6

4. Experiments and Results

The experiments evaluate LightMIS across six datasets and five imaging modalities, examining overlap scores and model complexity alongside on-device execution, statistical significance, and architectural choices. Ablations assess the complete AFC configuration and PRF, with parameter-matched tests used to contextualize feature-fusion and processing choices.

4.1. Datasets and Experimental Setup

  • The benchmark evaluates six publicly available 2D binary medical image segmentation datasets across five imaging settings, using five-fold image-level cross-validation and a shared nnU-Net v2.3.1 pipeline with models trained from scratch.The datasets span fundus, endoscopy, microscopy, ultrasound, and dermoscopy.Source: §4.1 · pp. 6–7

4.2. Comparison Under a Unified Training Protocol

  • 82.15% Dice and 69.82% IoU lead on DRIVE, while LightMIS records competitive results on the other datasets, including 89.32%/82.60% on ISIC-2017 and 89.83%/83.39% on ISIC-2018.On BUSI, Mobile U-ViT leads with 80.90% Dice and 73.27% IoU; on DSB18, LightMIS is 0.29 Dice and 0.40 IoU percentage points below the respective highest scores but has the lowest HD95 and ASSD values of 4.59 and 1.53. On Kvasir-SEG, nnU-Net’s 90.02% Dice is 0.15 points above LightMIS, which has the lowest ASSD at 5.44.Source: Table 3 · p. 8, Table 4 · p. 9, Table 5 · p. 9
  • 86.71% modality-macro Dice and 78.99% IoU place LightMIS just below Mobile U-ViT’s 86.75% Dice and 79.07% IoU, with substantially fewer parameters and GFLOPs.The observed differences are 0.04 and 0.08 percentage points; LightMIS uses 90.58% fewer parameters and 82.54% fewer GFLOPs than Mobile U-ViT, and 99.61% fewer parameters and 96.14% fewer GFLOPs than nnU-Net. The study describes these as close observed averages, not evidence of statistical superiority.Source: Table 6 · p. 10, §4.2 · p. 11
  • Model-specific trainer settings do not yield a consistent advantage over the controlled protocol, and the unified 200-epoch comparison may not be fair to all architectures.Recommended settings improve MK-UNet on DRIVE and BUSI but degrade AULUNet most on BUSI; the passage also notes possible effects of pretraining dependence and training duration across model families.Source: Table 7 · p. 10, §4.2 · p. 12

4.3. On-Device Benchmarking

  • At batch size one on the Mali-G52, LightMIS-T, LightMIS-S, and LightMIS averaged 53.39 ± 0.58, 76.32 ± 0.58, and 138.43 ± 1.04 ms; full LightMIS beat Mobile U-ViT (654.58 ± 23.45 ms) and nnWNet (1453.26 ± 15.56 ms).Full LightMIS had 100% GPU coverage, while LightMIS-T required no CPU fallback.Source: §4.3 · pp. 13–14, Table 8 · p. 13

4.4. Statistical Analysis

  • LightMIS shows statistically confirmed superiority over nnU-Net, nnWNet, and Mobile U-ViT only on DRIVE under the paired confidence-interval and adjusted-significance criteria.The analysis uses paired out-of-fold predictions, 10,000 paired bootstrap resamples, and Benjamini–Hochberg-adjusted sign-flip tests; nonsignificant results are not treated as evidence of equivalence.Source: §4.4 · p. 14, Table 10 · p. 14

4.5. Ablation Studies

  • Five encoder levels offer the best tested accuracy–complexity trade-off; six levels approximately double parameters while reducing overlap scores despite slight gains in selected Kvasir-SEG boundary metrics.PRF ablations find receptive-field diversity improves all reported BUSI and Kvasir-SEG metrics versus uniform 3 × 3 depthwise processing; progressive branch interaction notably improves BUSI overlap, while channel expansion contributes more clearly on Kvasir-SEG.Source: §4.5 · p. 14, Table 11 · p. 15
  • Jointly widening encoder channels and aggregation width produces the strongest overlap results on Kvasir-SEG, DRIVE, and DSB18, while BUSI performance appears to saturate beyond LightMIS-S.The scaling analysis separately varies aggregation width and encoder width before motivating their joint scaling.Source: §4.5 · p. 14, Table 12 · p. 15
  • Replacing the final ResAKF block with AKF raises modality-macro Dice from 85.74% to 85.92%, though small reductions occur on both ISIC datasets and DSB18.The largest gains are on Kvasir-SEG, DRIVE, and BUSI; ResAKF is retained for its aggregate trade-off, not a universal per-dataset benefit.Source: Table 13 · p. 15, §4.5 · p. 15
  • The complete AKF–PRF–ResAKF sequence achieves the highest BUSI Dice and IoU among evaluated configurations, but differing parameter counts prevent isolating composition effects.This motivates parameter-matched comparisons at the encoder and prediction-head levels.Source: Table 14 · p. 15, §4.5 · p. 15
  • The proposed prediction head achieves the highest Dice across DRIVE, BUSI, and Kvasir-SEG, with its largest margin on BUSI.The parameter-matched comparison uses the same LightMIS-T encoder and training protocol. Early encoder features matter for thin retinal structures on DRIVE; simple aligned sum or concatenation does not outperform deepest-feature prediction on BUSI or Kvasir-SEG.Source: §4.5 · pp. 15–16, Table 15 · p. 16
  • Among parameter-matched PRF replacements, PRF has the highest macro Dice and leads on BUSI and DRIVE, while a residual depthwise 3×3 block is marginally better on Kvasir-SEG.Alternatives include a parallel multi-kernel block, CMRF, AKF, and PRF without progressive cross-branch transfer; the complete configuration has the strongest task-aggregated trade-off, with dataset-dependent component effects.Source: Table 16 · p. 16, §4.5 · p. 16

5. Limitations

The evaluation is limited by its datasets, validation design, image preprocessing, and mobile benchmark scope. The paired analysis also uses fixed out-of-fold predictions and does not capture variation from repeated training.

5. Limitations

  • The study covers public 2D binary segmentation and does not establish performance on multiclass or volumetric tasks.The common nnU-Net protocol may also be suboptimal for baselines designed for pretraining, different optimizers, deep supervision, or model-specific schedules.Source: §5 · p. 16
  • For DRIVE, ISIC-2017, and ISIC-2018, pooling official labelled partitions means results are pooled-label cross-validation results, not official challenge-test results.Image-level folds lack subject grouping or category stratification, so independence across folds cannot be guaranteed when images share a subject, lesion, or acquisition sequence.Source: §5 · p. 16
  • Except for DRIVE, images were resized to 256×256 without preserving aspect ratio, which may alter anatomical or lesion geometry.The effect of aspect-ratio-preserving resizing remains unevaluated.Source: §5 · p. 16
  • The modality-macro average combines ISIC-2017 and ISIC-2018 into one dermoscopy score before equally weighting the five imaging modalities.Leading-model differences are small and should be interpreted using paired confidence intervals; the fixed out-of-fold analyses do not capture variation from repeated training.Source: §5 · p. 16
  • The mobile benchmark covers one smartphone GPU and software stack, measuring delegated model execution rather than complete application latency.It does not establish performance across other edge processors, sustained thermal conditions, energy budgets, or clinical workflows; no prospective clinical evaluation or external institutional validation was performed.Source: §5 · p. 16

6. Conclusion

LightMIS directly aggregates aligned features from five encoder levels and refines them with an Adaptive Fusion Cascade, without a learned stage-wise decoder. Under a common nnU-Net protocol, it achieves numerically close modality-macro scores to Mobile U-ViT with low parameter and computational costs, and all variants achieve complete GPU delegation on the tested smartphone GPU.

6. Conclusion

  • LightMIS directly aggregates all five aligned encoder levels without a learned stage-wise decoder, then refines the representation with an Adaptive Fusion Cascade.The cascade combines Adaptive Kernel Fusion, Progressive Receptive Fusion, and residual adaptive-kernel refinement.Source: §6 · p. 17
  • 86.71% modality-macro Dice and 78.99% modality-macro IoU were achieved by LightMIS with 0.131 M parameters and 0.575 GFLOPs; Mobile U-ViT scored 86.75% and 79.07%, respectively.The six-dataset evaluation used a common nnU-Net protocol; the differences are described as numerically close, not evidence of equivalence or statistical superiority.Source: §6 · p. 17
  • All three LightMIS variants achieved complete GPU delegation on an Arm Mali-G52 MC2 GPU, with median delegated latency from 53.31 ms for LightMIS-T to 138.31 ms for LightMIS.The measurements support on-device execution feasibility for the evaluated input size, not end-to-end clinical deployment.Source: §6 · p. 17
  • Future work will address patient-grouped and external validation, multiclass and volumetric segmentation, quantized execution on additional edge processors, and PRF and AFC on other dense prediction tasks.Source: §6 · p. 17

Related papers

Loading 2609.28327v1…