• Yuval Reina

• Zahar Chikishev

Private Leaderboard Score: 0.04732

Private Leaderboard Place: 12

This archive holds the code and weights which were used to create and inference the 12th place solution in “RSNA Intracranial Hemorrhage Detection” competition.

The solution consists of the following components, run consecutively

• Prepare data and metadata

• Training features generating neural networks

• Training shallow neural networks based on the features and metadata

  • By Yuval

  • By Zahar

• Ensembling

• Serialized – folder containing files for serialized training and inferencing of base models and shallow pooled – res.

• Production – folder, kept as reference, holds the original notebooks used to train the models and the submissions

• Notebooks – folder, holds jupyter notebooks to prepare metadata, training and inferencing Zahar’s shallow networks, end ensembling the full solution. It should be run in order appearing in this document below.

CPU intel i9-9920, RAM 64G, GPU Tesla V100, GPU Titan RTX.

OS: Ubuntu 18.04 TLS

CUDA – 10.1

GCP virtual machine with n-8 cores and K-80 GPU

1. Download train and test data from Kaggle and update ./Serialized/defenitions.py with the locations of train and test data

2. If you want to use our trained models, download and inflate models (for models in Serialized) put everything in one models folder and update ./Serialize/defenitions.py

notebooks/DICOM_metadata_to_CSV.ipynb - traverses DICOM files and extracts metadata into a dataframe. Produces three dataframes, one for the train images and two for the stage 1&2 test images.

notebooks/Metadata.ipynb - gets the output of the previous notebook and post-processes the collected metadata. Prepares metadata features for training, will be used as an input to Zahar's shallow NNs. Specifically, outputs two dataframes saved in train_md.csv and test_md.csv with the metadata features.

The last section of the notebook also prepares weights for the training images. The weights are selected to simulate the distribution to that we encounter in the test images.

Production/Prepare.ipynbis used to prepare the train.csv and test.csv for the base mosels and yuval's Sallow NN

./Serialized/train_base_models.ipynb is used to train the base models using, You should change the 2nd cell, and enter part of the name of the GPU you use, and the name of the model to train (look at defenitions.py for a list of names).

# here you should set which model parameters you want to choose (see definitions.py) and what GPU to use params=parameters['se_resnet101_5'] # se_resnet101_5, se_resnext101_32x4d_3, se_resnext101_32x4d_5 device=device_by_name("Tesla") # RTX , cpu

Beware, running this notebook to completion for a single base network will take a day or two.

./Serialized/Post Full Head Models Train .ipynb is used to train this shallow networks. This notebook trains all the networks. You should change the 2nd to reflect the GPU you use.

notebooks/Training.ipynb - trains a shallow neural network based on the generated features and the metadata. All of the models are fine-tuned after a regular training step. The fine tuning is different in that it uses weighted random sampling, with weights defined by notebooks/Metadata.ipynb.

./Serialized/prepare_ensembling.ipynb is used for inferencing this shallow model and prepare the results for ensembling.

notebooks/Ensembling.ipynb - ensembles the results from all shallow NNs into final predictions and prepares the final submissions.

The two final submissions are obtained by running this notebook and the difference is the following:

Safe submission ensembles regular Zahar and Yuval's models.

Risky submission ensembles weighted Zahar's models and regular Yuval's models, while the ensembling uses by-sample weighted log-loss with the same weights as defined before.