Paper link: https://arxiv.org/abs/1702.03833

please cite the paper if you find this project useful for your research

OS: linux

Hardware: NVIDIA GPU required

3rd-party software: CUDA, caffe, opencv, theano, lasagne, matlab, and other dependencies required by them. We use the bleeding edge edition of theano, lasagne and caffe.

HOW TO TRAIN

Firstly, specify the path to target folder in SETTINGS.json:

1. IN_TRAIN_DATA_PATH, path to train folder, please put in it raw folder
2. IN_VALIDATE_DATA_PATH, path to validate folder, please put it in raw folder
3. OUT_TRAIN_DATA_PATH, path to which the preprocess results of train folder are stored, please put it in clean folder
4. OUT_VALIDATE_DATA_PATH, path to which the preprocess results of validate folder are stored, please put it in clean floder
5. TRAIN_LABEL_PATH, path to train.csv, in which the number of rows should equal to the numpy of subfolders of IN_TRAIN_DATA_PATH
6. VAL_LABEL_PATH, path to validate.csv
7. FUSION_SNAPSHOT_PATH, path to model snapshot, please do not modify this

This algorithm includes two stages, the first one is detecting the heart area, which requires opencv, matlab and caffe. The parameters are already stored in folder /stage1, to re-train those parameters by your own, please refer to the README.md in stage1. The output of this stage is a stack of images at per frame, and stored in disk.

The second stage is computing volume out of a patch stack. It is an ensamble consists of 6 models. Each model is packaged by an adaptor class in directory fusion(fcn1, fcn2...). Our auto-learning procedure learning the weight of each model in ensamble. And it will automatically gives the submit result.

Then, run python train.py

How to predict

1. IN_TEST_DATA_PATH, path to test folder
2. OUT_TEST_DATA_PATH, path to which the preprocess results of test folder are stored
3. SUBMIT_PATH, path to save the submit csv file

Place the new test data in IN_TEST_DATA_PATH, also specify OUT_TEST_DATA_PATH and SUBMIT_PATH

Then, run python predict.py