MedDiffSeg (Under Construction)

IMPORTANT NOTE

This is a temporary repository. This project is currently under active development and construction.

Purpose

This repository is being created with the goal of comprehensively upgrading and improving upon the original MedSegDiff version.

The original repository can be found here: MedSegDiff

Comparison (Work in Progress)

The table below illustrates the (planned) differences between the original version and this new, upgraded version.

MedSegDiff (Original Version)	MedDiffSeg (New Version)
ImprintLab/MedSegDiff	New Project (MedDiffSeg)

• For training, run:

!python scripts/segmentation_train.py \
  --data_dir directory_path \
  --out_dir directory_path \
  --image_size 256 \
  --num_channels 128 \
  --class_cond False \
  --num_res_blocks 2 \
  --num_heads 1 \
  --learn_sigma True \
  --use_scale_shift_norm False \
  --attention_resolutions 16 \
  --diffusion_steps 1000 \
  --noise_schedule linear \
  --rescale_learned_sigmas False \
  --rescale_timesteps False \
  --lr 1e-4 \
  --batch_size 16 \
  --log_interval 50 \
  --save_interval 1000 \
  --lr_anneal_steps 100000

• For testing, run:

python scripts/segmentation_sample.py \
  --data_name BRATS3D \
  --data_dir /home/admin1/Projects/MedDiffSeg_GPU/data/training_test \
  --out_dir /home/admin1/Projects/MedDiffSeg_GPU/output \
  --model_path /home/admin1/Projects/MedDiffSeg_GPU/models/emasavedmodel_0.9999_100000.pt \
  --image_size 256 \
  --num_channels 128 \
  --class_cond False \
  --num_res_blocks 2 \
  --num_heads 1 \
  --learn_sigma True \
  --use_scale_shift_norm False \
  --attention_resolutions 16 \
  --diffusion_steps 1000 \
  --noise_schedule linear \
  --rescale_learned_sigmas False \
  --rescale_timesteps False \
  --num_ensemble 5 \
  --batch_size 1 \
  --gpu_dev 0 \
  --debug True \
  --version medsegdiff-v2 \
  --in_ch 5

---

Kaggle data link: Brats 2021

Please check back later for detailed updates. Thank you for your patience!