Theia-4869 / U-RISC

EM cytomembrane segmentation with neural network ensemble and morphological processing.

EM cytomembrane segmentation with neural network ensemble and morphological processing

Environment

Hardware

• 4 NVIDIA 3090Ti GPUs (24GB memory each)
• 32 CPUs

Packages

pip install -r requirements.txt

Data

Processed data and pretrained ResNet (50 and 152) can be downloaded from the links below. Put it in top-level folder.

Baidu Disk password: 8eqb

PKU Disk

Model

Simple Track

DFF, backbone ResNet-50

Complex Track

CASENet, backbone ResNet-152

Training

Simple Track

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset simple --model DFF --backbone resnet50 --batch-size 8 --lr 0.001 --epochs 200 --crop-size 960 --kernel-size 5 --edge-weight 0.4

Simple Track (Visualization)

CUDA_VISIBLE_DEVICES=0,1,2,3 python train_vis.py --dataset simple --model DFF --backbone resnet50 --batch-size 8 --lr 0.001 --epochs 200 --crop-size 960 --kernel-size 5 --edge-weight 0.4

Complex Track

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset complex --model CASENet --backbone resnet152 --batch-size 4 --lr 0.001 --epochs 100 --crop-size 1280 --kernel-size 9 --edge-weight 0.4

Complex Track (Visualization)

CUDA_VISIBLE_DEVICES=0,1,2,3 python train_vis.py --dataset complex --model CASENet --backbone resnet152 --batch-size 4 --lr 0.001 --epochs 100 --crop-size 1280 --kernel-size 9 --edge-weight 0.4

Validaing and Morphological Processing

Simple Track

CUDA_VISIBLE_DEVICES=0 python val.py --dataset simple --model DFF --backbone resnet50

Complex Track

CUDA_VISIBLE_DEVICES=0 python val.py --dataset complex --model CASENet --backbone resnet152

Morphological Processing

python val_mor.py

Testing and Ensembling

Simple Track

CUDA_VISIBLE_DEVICES=0 python test.py --dataset simple --model DFF --backbone resnet50

Complex Track

CUDA_VISIBLE_DEVICES=0 python test.py --dataset complex --model CASENet --backbone resnet152