Tree-NET cuts medical image segmentation FLOPs 4-13x, boosts Dice scores

Tree-NET, a new framework from researchers Orhan Demirci and Bulent Yilmaz, tackles one of medical imaging's persistent challenges: balancing segmentation accuracy with computational efficiency. The system employs dual bottleneck supervision, a novel approach that compresses both input images and segmentation labels into dense, low-dimensional latent representations using autoencoder-style Encoder-Net and Decoder-Net components. A Bridge-Net then performs segmentation on these compressed features. This design reduces FLOPs by a factor of 4 to 13 and decreases memory usage, while maintaining or improving accuracy compared to standard U-NET, U-NET++, and Polyp-PVT backbones.

Tested on skin lesion (ISIC 2018) and polyp (CVC-ClinicDB) segmentation, Tree-NET shows concrete gains: with an untrained U-NET++ backbone, Dice score improved from 0.829 to 0.862 and IoU from 0.736 to 0.787. On CVC-ClinicDB with a Polyp-PVT backbone, it hit 0.946 Dice and 0.901 IoU, matching or exceeding baselines. Because Tree-NET works entirely on compressed features, it can be dropped into existing segmentation models without modifying their internal architecture or increasing model size. The paper, accepted by Neural Computing and Applications, is available on arXiv (2501.02140).