UC5 "Deep Image Annotation"

Project DEEPHEALTH

Repository for the Use Case 5 "Deep Image Annotation" containing:

• models based on the EDDL/ECVL libraries;
• models based on the PyTorch-Lightning library.

Previous models, used during the development, and all of the PyTorch models can be found in the release "v1_dev" available to download.

FOLDER STRUCTURE

• conda_envs: YAML files for creating the environments based on the EDDl/ECVL (cuDNN version) and PyTorch-Lightning
• demo, demo_src: code to be used in demos
• src: source code
  • src/eddl: EDDL-based models
  • src/PyTorch-Lightning: PyTorch (Lightning) models
  • src/preproc: code for pre-processing the IU-CHEST and the MIMIC-CXR datasets

HOW TO USE

• First, preprocess the datasets.

  • For the IU-CHEST dataset, use the Makefile.mk in src/preproc
  • For the MIMIC-CXR dataset, use the Jupyter Notebook src/preproc/NB_mimic_ds.ipynb

• Set up the experiments using the Jupyter notebooks src/eddl/0_experiments.ipynb and src/eddl/0_experiments_mimic.ipynb for, respectively, the IU-CHEST and the MIMIC-CXR datasets. Some experiments are provided in the two notebooks.

• For the experiments already defined in the jupyter notebooks at the item above, use the two makefiles src/eddl/Makefile.mk and src/eddl/Makefile_MIMIC.mk for, respectively, the IU-CHEST and the MIMIC-CXR datasets.

Once configured an experiment, the correct sequence of the scripts is: 1_train_cnn.py, 2_train_rnn.py, 3_test_rnn.py.

CONTACT

Franco Alberto Cardillo, ILC-CNR, francoalberto.cardillo@ilc.cnr.it