README.md

metadata

title: TEDM
emoji: 🐨
colorFrom: purple
colorTo: yellow
sdk: gradio
sdk_version: 3.35.2
app_file: app.py
pinned: false
license: mit

Robust semi-supervised segmentation with timestep ensembling diffusion models

Results

Training data size	1 (1%)	3 (2%)	6 (3%)	12 (96%)	197 (100%)
	JSRT (labelled in-domain)
Baseline	84.4 $\pm$ 5.4	91.7 $\pm$ 3.7	93.3 $\pm$ 2.9	95.3 $\pm$ 2.3	97.3 $\pm$ 1.2
LEDM	90.8 $\pm$ 3.5	94.1 $\pm$ 1.6	95.5 $\pm$ 1.4	96.4 $\pm$ 1.4	97.0 $\pm$ 1.3
LEDMe	93.7 $\pm$ 2.6	95.5 $\pm$ 1.5	96.7 $\pm$ 1.5	97.0 $\pm$ 1.1	97.6 $\pm$ 1.2
Ours	93.1 $\pm$ 3.4	94.8 $\pm$ 1.4	95.8 $\pm$ 1.2	96.6 $\pm$ 1.1	97.3 $\pm$ 1.2
	NIH (unlabelled in-domain)
Baseline	68.5 $\pm$ 12.8	71.2 $\pm$ 15.1	71.4 $\pm$ 15.9	77.8 $\pm$ 14.0	81.5 $\pm$ 12.7
LEDM	63.3 $\pm$ 12.2	78.0 $\pm$ 10.1	81.2 $\pm$ 9.3	85.9 $\pm$ 7.4	88.9 $\pm$ 5.9
LEDMe	70.3 $\pm$ 11.4	78.3 $\pm$ 9.8	83.0 $\pm$ 8.6	84.4 $\pm$ 8.1	90.1 $\pm$ 5.3
Ours	80.3 $\pm$ 9.0	86.4 $\pm$ 6.2	89.2 $\pm$ 5.5	91.3 $\pm$ 4.1	92.9 $\pm$ 3.2
	Montgomery (out-of-domain)
Baseline	77.1 $\pm$ 12.0	83.0 $\pm$ 12.2	80.9 $\pm$ 14.7	83.8 $\pm$ 14.9	94.1 $\pm$ 6.6
LEDM	79.3 $\pm$ 8.1	85.9 $\pm$ 7.4	89.4 $\pm$ 6.7	92.3 $\pm$ 7.2	94.4 $\pm$ 7.2
LEDMe	80.7 $\pm$ 6.6	86.3 $\pm$ 6.5	89.5 $\pm$ 5.9	91.2 $\pm$ 5.6	95.3 $\pm$ 4.0
Ours	90.5 $\pm$ 5.3	91.4 $\pm$ 6.1	93.3 $\pm$ 6.0	94.6 $\pm$ 6.0	95.1 $\pm$ 6.9

Training

• training the backbone

python train.py --dataset CXR14 --data_dir <PATH TO CXR14 DATASET>

• our method

python train.py --experiment TEDM --data_dir <PATH TO JSRT DATASET> --n_labelled_images <TRAINING SET SIZE>

• LEDM method

python train.py --experiment LEDM --data_dir <PATH TO JSRT DATASET> --n_labelled_images <TRAINING SET SIZE>

• LEDMe method

python train.py --experiment LEDMe --data_dir <PATH TO JSRT DATASET> --n_labelled_images <TRAINING SET SIZE>

• baseline method

python train.py --experiment JSRT_baseline --data_dir <PATH TO JSRT DATASET> --n_labelled_images <TRAINING SET SIZE>

Testing

• update

  • DATADIR in paths dataloaders/JSRT.py, dataloaders/NIH.py and dataloaders/Montgomery.py
  • NIHPATH, NIHFILE, MONPATH and MONFILE in paths auxiliary/postprocessing/run_tests.py and auxiliary/postprocessing/testing_shared_weights.py

• for baseline and LEDM methods, run

python auxiliary/postprocessing/run_tests.py --experiment <PATH TO LOG FOLDER>

• for our method, run

python auxiliary/postprocessing/testing_shared_weights.py --experiment <PATH TO LOG FOLDER>

Figures and reporting

VS Code notebooks can be found in auxiliary/notebooks_and_reporting.