How to win¶

Based on the winning solutions in FLARE21, we summarize the following strategies to obtain better performance in this challenge.

Strategies to improve the segmentation performance¶

• Preprocessing: intensity normalization, resampling...
• Extensive data augmentations;
• Coarse-to-fine (two-stage or cascaded) framework;
• Postprocessing: connected component analysis;

Strategies to improve the computational efficiency¶

• Whole-volume based input;
• Lightweight network modules: residual block with bottleneck, separated convolutions, and pyramid pooling;
• Accelerate inference by ONNX Runtime, TensorRT...

Important: You can get a perfect GPU-Time score if limiting the maximum  GPU memory consumption within 2GB.

The code and paper of FLARE21 winners have been publicly available at 

https://flare.grand-challenge.org/Awards/

Some useful tutorials¶

• Using Docker to Share Your Software
• nnU-Net : The no-new-UNet for automatic segmentation
• Introducing the generally nuanced DL framework Segmentation Challenge