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

LightMIS. Ultra-Lightweight Medical Image Segmentation Without a Stage-Wise Decoder. By Andrei Arhire and two others, from Alexandru Ioan Cuza University of Iasi and the University of Wurzburg. Published on the twenty-fourth of September. Its code repository is public and real: about twenty source files.

The task is two-dimensional binary medical image segmentation: marking, pixel by pixel, which part of an image is the structure of interest and which is not. The paper's target is doing this with networks that are ultra-lightweight, light enough to execute on the device itself, without giving up accuracy against larger models. Its title names the component it gives up to get there: a learned stage-wise decoder.

LightMIS is a scalable family of convolutional networks built around that omission. Instead of decoding stage by stage, it brings all of the encoder's outputs to the same resolution, combines them once, and spends its effort refining that single fused representation.

Concretely, the encoder has five levels. Scale-Aligned Projection blocks align the outputs of all five to a common resolution, and they are aggregated in one step. An Adaptive Fusion Cascade then refines the fused result. The cascade combines Adaptive Kernel Fusion with a newly proposed Progressive Receptive Fusion module, which uses temporary channel expansion, complementary depthwise receptive fields, and progressive information transfer across branches.

Three sizes are evaluated: LightMIS-T, LightMIS-S, and full LightMIS. Testing uses five-fold cross-validation under a common nnU-Net protocol, on six datasets: DRIVE, Kvasir-SEG, DSB18, BUSI, ISIC-2017, and ISIC-2018. Full LightMIS has zero point one three one million parameters and needs zero point five seven five GFLOPs for a two hundred and fifty-six pixel square image.

Averaged across modalities, it reaches a Dice score of eighty-six point seven one percent and an IoU of seventy-eight point nine nine percent. Mobile U-ViT obtains eighty-six point seven five Dice and seventy-nine point zero seven IoU, so the gaps are four hundredths and eight hundredths of a percentage point.

Against Mobile U-ViT, nnWNet, and nnU-Net, LightMIS cuts parameter count by between ninety point five eight and ninety-nine point six one percent, and GFLOPs by between eighty-two point five four and ninety-six point one four percent. On an Arm Mali-G52 MC2 GPU, every variant runs with full GPU delegation. Median delegated latency ranges from fifty-three point three one milliseconds for LightMIS-T to one hundred and thirty-eight point three one milliseconds for full LightMIS.

The authors frame the result as a favorable trade between accuracy and complexity, with on-device execution shown to be feasible for the tasks they evaluated. Dropping the stage-wise decoder costs a few hundredths of a point of Dice, and removes most of the parameters and computation.