RSNA Intracranial Hemorrhage Detection

• Competition name: RSNA Intracranial Hemorrhage Detection https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection
• Team name: Keep Digging Gold
• Private Leaderboard score: 0.04561
• Private Leaderboard place: 5th

Entrypoints for the 5th-place solution

The solution is presented in the Keep-Digging-Gold-Solution-Doc.pdf and Keep-Digging-Gold-Solution-Slides.pdf. A video is also available at https://www.youtube.com/watch?v=1zLBxwTAcAs&t=16s.

Train models

Follow the instructions in the following 3 repositories to train 9 models (as indicated in the Keep-Digging-Gold-Solution-Doc.pdf), generate predictions for the test set and the out-of-fold-train set

• [1] Bac Nguyen’s source code https://github.com/ngxbac/Kaggle-RSNA
• [2] Nguyen Tai Tri Duc 's pipeline to train InceptionV3 + Deepsupervision https://github.com/triducnguyentang/RSNA
• [3] Anjum 's pipeline https://github.com/Anjum48/rsna-ich

Post-processing

See the section at the end of this file.

Stacking

Details and code are presented at https://www.kaggle.com/mathormad/5th-place-solution-stacking-pipeline.

Postprocessing

Required inputs:

• df_dicom_metadata_train.csv and df_dicom_metadata_test.csv. These are created using dicom-metadata-to-csv-train-and-test.ipynb. The stage 1 labels can be appended to the train data on release, and the test data can be replaced with the stage 2 metadata.
• patients_stacking_splits.csv which contains the 5 fold cross-validation scheme for the post-processsing model. This is created by stacking-split.ipynb. When the stage 1 labels are released, the stage 1 images can be appended to this file as a 6th fold.
• A CSV file containing the out-of-fold predictions from a model
• A CSV file containing the test predictions (i.e. a submission file)

Run post-processing-v7-with-wo-meta.ipynb with the following options

• META_NUMBER_IMAGES_IN_USE = True
• NUMBER_FOLDS = 6 # 6 for if we include the stage 1 set