LightMIS: Ultra-Lightweight Medical Image Segmentation Without a Stage-Wise Decoder
Andrei Arhire, Mihaela-Elena Breabăn, Radu Timofte
cs.CV
2026-09-24
LightMIS drops U-Net's stage-wise decoder for one-shot aggregation, hitting 86.71% macro Dice at 0.131M params, 0.04 behind Mobile U-ViT, 138 ms on Mali-G52.
Medical segmentation still defaults to U-Net: an encoder downsamples, a decoder upsamples, skip links stitch the two. Lightweight follow-ups slim the convolutions, attention, or MLPs. The skeleton is usually still a symmetric decoder. On point-of-care devices, phones, and tight clinical budgets, that decoder is a parameter and memory tax.
LightMIS asks whether a five-level encoder still needs that stage-wise path if its hierarchy is aligned once and aggregated once.
"Without a stage-wise decoder" is defined tightly: no learned cascade of resolution-recovery stages, and no level-by-level skip fusion. Bilinear alignment and a final logit resize to the input stay.
A five-level encoder, each level an Adaptive Fusion Cascade (AFC). First Adaptive Kernel Fusion (AKF, adapted from AULUNet: a channel gate mixes 3×3 and dilated 3×3 depthwise convolutions), then Progressive Receptive Fusion (PRF), then residual AKF. Between levels, 2×2 average pooling and a 1×1 projection set the next width.
PRF temporarily expands channels to 2C and splits into three branches: 3×3, dilated 3×3, and 5×5. The channel-wise mean of one branch is added to the next before convolution. A learned scalar residual-adds the fused result. Narrow nets get extra capacity from a temporary widen and complementary receptive fields, without thickening the whole encoder.
Scale-Aligned Projection (SAP) maps all five outputs to a shared width F, bilinearly aligns them to the second-level resolution, concatenates, 1×1-compresses back to F, and runs one more AFC before a pointwise head.
Three widths, same topology. Tiny uses channels (3,6,10,16,24) and F=10: 0.017M parameters, 0.099 GFLOPs. Small sits in between. Full LightMIS uses (10,20,32,48,80) and F=32: 0.131M parameters, 0.575 GFLOPs on 3×256×256 inputs.
Everything is dropped into nnU-Net v2.3.1's 2D config. Six datasets, five-fold, 200 epochs from scratch, deep supervision off, no post-processing. Baselines are restricted to official public implementations under the same plans and splits.
Modality-macro averages ISIC-2017 and ISIC-2018 into one dermoscopy score, then weights dermoscopy, endoscopy, fundus, ultrasound, and microscopy equally:
| Model | Params | GFLOPs | Macro Dice | Macro IoU | Mean Dice rank |
| LightMIS | 0.131M | 0.575 | 86.71 | 78.99 | 3.42 |
| Mobile U-ViT | 1.390M | 3.294 | 86.75 | 79.07 | 4.00 |
| nnWNet | 7.044M | 10.618 | 86.51 | 78.84 | 4.50 |
| nnU-Net | 33.472M | 14.882 | 86.58 | 78.74 | 4.58 |
Against the last three, parameters drop 90.58% to 99.61% and GFLOPs drop 82.54% to 96.14%. Macro Dice sits 0.04 points behind Mobile U-ViT.
On DRIVE, Dice 82.15 and IoU 69.82 lead the six datasets, and paired tests against those three baselines clear both the confidence-interval and FDR bars. Kvasir-SEG is 89.87 versus nnU-Net 90.02. BUSI is 80.16 versus Mobile U-ViT 80.90. DSB18 has the best boundaries: HD95 4.59, ASSD 1.53. Statistical superiority is confirmed only on DRIVE. Other datasets should not be read as ties.
On a Motorola Moto G24 Power (Arm Mali-G52 MC2), all three LightMIS variants reach 100% GPU delegation. Median delegated latency runs from 53.31 ms (Tiny) to 138.31 ms (full). Mobile U-ViT is 657.20 ms, nnWNet 1454.44 ms on the same device. AULUNet is 18.53 ms. UNeXt, at 0.577 GFLOPs, is 42.14 ms; full LightMIS at 0.575 GFLOPs is more than 3× slower on device. GFLOPs do not predict phone latency.
Ablations: five encoder levels win; a sixth roughly doubles parameters and loses overlap. Parameter-matched heads: deepest-feature-only collapses DRIVE Dice to 52.62; SAP plus a final AFC reaches 82.97 macro, a slim U-Net decoder 82.38.
This is a measurement of how far "drop the stage-wise decoder, align once" goes under a shared nnU-Net protocol. 0.131M parameters land within 0.04 Dice of a 1.39M hybrid, and win on thin retinal vessels. The useful setting is on-device 2D binary segmentation. It is not a new state of the art across the board. It is an accuracy-versus-size operating point.
The authors list them. Public 2D binary datasets only. A shared nnU-Net recipe can hurt models that expect pretraining or their own optimizer. DRIVE and ISIC results are pooled-label cross-validation, not official challenge tests. All datasets except DRIVE are resized to 256×256 without preserving aspect ratio. Splits are image-level, so the same subject can leak across folds. Macro gaps are tiny and most paired intervals cross zero. The phone study is one Mali-G52 and measures delegated inference, not a full app. ResAKF slightly hurts ISIC and DSB18; it is kept for the better macro score. No prospective clinical evaluation.