LightMIS: Ultra-Lightweight Medical Image Segmentation Without a Stage-Wise Decoder
LightMIS segments medical images with 0.131M parameters and near Mobile U-ViT Dice while cutting compute sharply.
LightMIS is a family of ultra-lightweight convolutional networks for 2D binary medical image segmentation that omits a learned stage-wise decoder. Full LightMIS has 0.131 million parameters and needs 0.575 GFLOPs for a 3x256x256 input, reaching modality-macro Dice of 86.71% and IoU of 78.99% under an nnU-Net v2.3.1 protocol. That is within 0.04 Dice points of Mobile U-ViT while using 90.58-99.61% fewer parameters than Mobile U-ViT, nnWNet, and nnU-Net. On an Arm Mali-G52, median delegated latency ranges from 53.31 ms for LightMIS-T to 138.31 ms for full LightMIS.
- No learned stage-wise decoder; uses projection blocks and an Adaptive Fusion Cascade.
- Full LightMIS has 0.131M parameters, 0.575 GFLOPs, and 86.71% Dice.
- Parameters fall 90.58-99.61% versus Mobile U-ViT, nnWNet, and nnU-Net.
- Tested on DRIVE, Kvasir-SEG, DSB18, BUSI, and ISIC datasets.
- Products
- nnU-Net
- AI models
- LightMISMobile U-ViT
From the original article
1 min read at huggingface.co
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.
By Andrei Arhire, Mihaela-Elena Breabăn, Radu Timofte · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.