Creating data pipeline and training pipeline for left ventricular blood pool (myocardium) semantic segmentation from DICOM MRI images left.
Created for the purpose of the interview, focused on creating machine learning pipeline to train and evaluate the model. The purpose is not to make best performing / fast model.
Project is organized as a python package. So first you should install it, best with development mode.
python setup.py developThen install PyTorch, see install guide.
Download data and extract it to DATA_DIR according to your choice.
Run notebooks/tests.ipynb to verify correct function of parsing.py/polygon_to_mask() and create data for unit testing of this method.
Run scripts/prepare_data.py data pipline to preprocess data and split it to train/dev dataset.
python scripts/prepare_data.pyDATA_DIR/datasets/ folder was created with train/ and dev/ subfolders containing preprocessed pickled examples prepared for training pipeline.
Run /scripts/train.py to train the model.
python scripts/train.py --train_dir DATA_DIR/datasets/train/ --dev_dir DATA_DIR/datasets/dev/ --exp exp01which starts training of the net and it stores experiment data to experiments/exp01/ folder.
How did you verify that you are parsing the contours correctly?
- First I converted contours to mask by untested method for some random images
- Then I plotted corresponding dicom and mask images to verify that mask is correct for those random images notebooks/tests.ipynb
- I stored those verified masks for unit testing without manual verification
What changes did you make to the code, if any, in order to integrate it into our production code base?
- types in docstring
- changed return type in parse_dicom_file to np.ndarray
- changed some names function and variable names
- added some unit tests
Did you change anything from the pipelines built in Parts 1 to better streamline the pipeline built in Part 2? If so, what? If not, is there anything that you can imagine changing in the future?
- To avoid converting contours to mask in every epoch I created prepare_data.py which loads whole dataset, split it to train and dev dataset and store each example separately in pickle file into train/dev folder.
How do you/did you verify that the pipeline was working correctly?
- I created test_model_wrapper.py that is checking if:
- loading from dataset is correct
- trainset is shuffled for each epoch and devset not
- loss function == log(nr_classes) before starting to train
Given the pipeline you have built, can you see any deficiencies that you would change if you had more time? If not, can you think of any improvements/enhancements to the pipeline that you could build in?
- Add image enhancement in trainset (flip, rotation, adding noise, change contrast, etc.)
- Check gradients are not zero
- Check activations approximately equally distributed between -1 and 1 (good weight initialization)
- Prediction and evaluation (plot examples with the lowest/highest iou score)
- Learning resume
- Loading saved model
- Automatic hyperparameter tuning
- Better logging
- The aim of this project was to prepare data for training and create training pipeline. Training phase has not been tuned. More images would probably be neccessary.
After building the pipeline, please discuss any changes that you made to the pipeline you built in Phase 1, and why you made those changes.
- See merge request #2
Let’s assume that you want to create a system to outline the boundary of the blood pool (i-contours), and you already know the outer border of the heart muscle (o-contours). Compare the differences in pixel intensities inside the blood pool (inside the i-contour) to those inside the heart muscle (between the i-contours and o-contours); could you use a simple thresholding scheme to automatically create the i-contours, given the o-contours? Why or why not? Show figures that help justify your answer.
Do you think that any other heuristic (non-machine learning)-based approaches, besides simple thresholding, would work in this case? Explain.
- Check Conclusion in notebooks/Thresholding.ipynb