Demonstration how to get kits19 intersection over union (IOU) above 80 % with 2D UNet using transfer learning!
Raw image | Mask (Ground Truth) | Prediction |
---|---|---|
- Read challange paper: kits19.grand-challenge.org
- Follow neheller's github to get data a libraries: github
- If you have downloaded
data
folder includingimaging.nii.gz
andsegmentation.nii.gz
in each folder (first 209 only), you can start using my scripts- Run
sanitycheck.py
to createpng
folder with visualisation of raw data - Run
convertniitonpz.py
to createnpz
folder with preprocessed images compressed in to*.npz
files for fast data loading - Run
train.py
to createmodels
folder with timestamps folders containing TensorBoard, .csv and models! - Run
predict.py
to createmodels/{timestamp}/predictions
with predicted heatmaps in grayscale
- Run
Pull Request welcome
If you need help with your project, my work helps you or you have any ideas how to improve my code, let me know about it!
π¦π» www.MiroslavKabat.com
βοΈ hello@miroslavkabat.com
I spent in this project more like π 16 man-hours with β 18 months of getting knowledge and πΈ 6000 $ for workstation, I appreciate any donations. PayPal Donations Thank you!