Weak label in CT image segamentation

Train models using using weak supervision such as image-level labels, producing pixel-level labels for CT scans.

Dataset

Folder hierarchy of dataset.

--Biomedical_data
  --train.py
  --predict.py
  --train256 
    --liver 
    --noliver 
  --test256 
    --images_pngs 
    --masks_pngs 

Perform transfer learning on CT scans dataset

python train.py --image_dir ./train256/
                --output_checkpoint ./output_model/checkpoint.ckpt
                --batch_size 16
                --epochs 10

Generate prediction for test256/images_pngs

python predict.py --image_dir ./test256/images_pngs/
                  --checkpoint_path ./output_model/checkpoint.ckpt
                  --output_dir ./output_prediction
                  --sensitivity 130

Result visualization

Reference

Learning Deep Features for Discriminative Localization http://cnnlocalization.csail.mit.edu/Zhou_Learning_Deep_Features_CVPR_2016_paper.pdf

Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization https://arxiv.org/pdf/1610.02391.pdf