LightMIS: an ultra-lightweight medical image segmentation model for edge devices

Researchers released LightMIS, a 2D binary medical segmentation model with 131K parameters that aggregates multi-scale encoder features without a learned stage-wise decoder. Under the same nnU-Net protocol it matches nnWNet's Dice/IoU while using 98.14% fewer parameters and 94.58% fewer GFLOPs; on an Arm Mali-G52 GPU it ran about 10.5x faster with 100% GPU operator coverage. Code and checkpoints are public.