This repository is the official implementation of our participation in FLARE competition. Experiments on the MICCAI FLARE23 challenge leaderboard validate promising performance achieving high segmentation accuracy with average Dice similarity coefficients of 89.84 % and 50.26 % for multi-organ and tumor regions respectively. Additionally, efficient inference is exhibited with an average runtime of 18 seconds per 512 × 512 × 215 3D volume with less than 2G GPU memory consumption. See our paper in reference section for more details.
Our approach is based on the classic two-phase (location-segmentation) cascaded processing stream wherein a lightweight CNN in phase one employing partial convolution and a novel hybrid CNN-Transformer model with synergistic amalgamation of scale-aware modulator and self-attention in phase two are proposed.
- Ubuntu 20.04.5 LTS
- Python 3.8
- torch 2.0.1
- torchvision 0.15.2
- CUDA 11.8
- monai 1.2.0
Training files can be found in folder "flare2023_train". Change custom transforms stream in "monai_datamodule.py". Remember to specify phase number before training. Run in terminal:
cd flare2023_train
python train.py
Inferencing files can be found in folder "flare2023_inference". we provide our best model weights for both phases.
Run in terminal:
cd flare2023_inference
python inference.py