Sauron U-Net: Simple automated redundancy elimination in medical image segmentation via filter pruning

Juan Miguel Valverde, Artem Shatillo, Jussi Tohka

Abstract: Highlights•Sauron clusters and prunes neural networks feature maps during the training.•Sauron reduced the FLOPs of nnUNet inference by over 90% without sacrificing the segmentation quality.•Sauron eliminated more filters and achieved higher Dice coefficients than four state-of-the-art filter pruning methods.•After pruning with Sauron, the feature maps in the last layer were highly interpretable.