The learning pipeline for the echochardiography datasets is based in the following elements:
Data-selection and management; Model training and tuning; Model validation (performance evaluation and clinical evaluation); AI-based device modification, and (perhaps) AI-based production model (Fig. 1).
Fig 1. Total product lifecycle (TPLC) approach on AI/ML workflow from Good Machine Learning Practices
Prepare your mp4 files and annotations files as suggested here. It is suggested that files follow this organisation and localisation. NOTE. extra video files can be renamed with a different extension to avoid taking them into account.
Click to expand and see files organisation and location!
$ cd ~/vitaluskcl/datasets/echocardiography/videos-echo-annotated-ALL # Alternatively: ~/datasets/echocardiography-vital/videos-echo-annotated-06-subjects
$ tree -s
.
├── [ 4096] 01NVb-003-050
│ ├── [ 4096] T1
│ │ ├── [ 986] 01NVb-003-050-1-4CV.json
│ │ └── [ 1803334463] 01NVb-003-050-1 echo.mp4
│ ├── [ 4096] T2
│ │ ├── [ 988] 01NVb-003-050-2-4CV.json
│ │ └── [ 1752445210] 01NVb-003-050-2 echo.mp4
│ └── [ 4096] T3
│ ├── [ 987] 01NVb-003-050-3-4CV.json
│ └── [ 1062609410] 01NVb-003-050-3 echo.mp4
├── [ 4096] 01NVb-003-051
│ ├── [ 4096] T1
│ │ ├── [ 986] 01NVb-003-051-1-4CV.json
│ │ └── [ 826247505] 01NVb-003-051-1 echo.mp4
│ ├── [ 4096] T2
│ │ ├── [ 988] 01NVb-003-051-2-4CV.json
│ │ └── [ 1234164657] 01NVb-003-051-2 echo.mp4
│ └── [ 4096] T3
│ ├── [ 906] 01NVb-003-051-3-4CV.json
│ └── [ 1198707159] 01NVb-003-051-3 echo.mp4
See more details on data curation here
2. Learning pipeline notebook 📓
The jupyter nobebook learning_pipeline_notebook.ipynb involves pre-processing, segment sampling, model and hyperparameter tunning pipeline (Fig. 1).
Fig 1. Deep learning pipeline of the AI-empowered echocardiography.
-
config_files/users_paths_files See README
-
Open a terminal, load your conda environment and run the script.
cd $HOME/repositories/vital-ultrasound/ai-assisted-echocardiography-for-low-resource-countries/scripts/learning-pipeline
export PYTHONPATH=$HOME/repositories/vital-ultrasound/ai-assisted-echocardiography-for-low-resource-countries/ #set PYTHONPATH environment variable
conda activate rt-ai-echo-VE
jupyter notebook # to open *.ipynb in your web-browser
- Copying results to a different path
cd $HOME/repositories/echocardiography/scripts/learning-pipeline
cp -r results/ ~/Desktop/results_today
- Temporal files
EchoClassesDataset()creates a temporal path at$HOME/datasets/vital-us/echocardiography/temporal-files/echovideodatafiles_FRAMESPERCLIP{$K}_PIXELSIZE_{$NW}W{$NH}Hwhere K are thenumber_of_frames_per_segment_in_a_clipand{$NW}and{$NH}are pixel size of the ultrasound image.
Example:
mx19@sie133-lap:~/datasets/vital-us/echocardiography/temporal-files/echovideodatafiles_FRAMESPERCLIP500_PIXELSIZE_200W200H$ ll
total 1.9G
13370667 drwxrwxr-x 2 mx19 mx19 12K May 16 13:50 .
12599341 drwxrwxr-x 3 mx19 mx19 4.0K May 16 13:48 ..
13373904 -rw-rw-r-- 1 mx19 mx19 24M May 16 13:49 videoID_00_040-1_label_00.pth
13374254 -rw-rw-r-- 1 mx19 mx19 23M May 16 13:49 videoID_00_043-2_label_00.pth
13373934 -rw-rw-r-- 1 mx19 mx19 23M May 16 13:49 videoID_01_040-2_label_00.pth
13374255 -rw-rw-r-- 1 mx19 mx19 23M May 16 13:49 videoID_01_041-3_label_00.pth
13373958 -rw-rw-r-- 1 mx19 mx19 23M May 16 13:49 videoID_02_040-1_label_00.pth
3. Heuristics for learning pipeline 📓
- Open a terminal, load your conda environment and run the script.
cd $HOME/repositories/echocardiography/scripts/learning-pipeline
export PYTHONPATH=$HOME/repositories/echocardiography/ #set PYTHONPATH environment variable
conda activate rt-ai-echo-VE
jupyter notebook # to open *.ipynb in your web-browser
4. Evaluation of learning pipeline 📓
- Open a terminal, load your conda environment and run the script.
cd $HOME/repositories/echocardiography/scripts/learning-pipeline
export PYTHONPATH=$HOME/repositories/echocardiography/ #set PYTHONPATH environment variable
conda activate rt-ai-echo-VE
jupyter notebook # to open *.ipynb in your web-browser