i-pan/kaggle-rsna-cspine

6th place solution source code for RSNA 2022 Cervical Spine Fracture Detection challenge

RSNA 2022 Cervical Spine Fracture Detection

Solution Summary

Detailed writeup: https://docs.google.com/document/d/1GcHeUDks2dnECKmJn97nF2djFRTVVnGpORFpUH0nMuU/edit?usp=sharing Kaggle post: https://www.kaggle.com/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/362651

Hardware

Ubuntu 20.04

8-core Intel processor

64 GB RAM

2 NVIDIA RTX 3090 GPUs with 24 GB VRAM

Environment

bash src/environment.sh
conda activate skp

Download Data

mkdir data
cd data
kaggle competitions download -c rsna-2022-cervical-spine-fracture-detection

Unzip the files.

Initial Setup

From src/etl:

python 00_extract_metadata.py
python 01_convert_to_png.py
python 02_convert_nifti_to_numpy.py
python 03_generate_whole_seg_192x192x192_numpy.py
python 04_create_cv_splits.py
python 05_create_cv_splits_for_whole_cspine_segmentation.py

Train Segmentation Models

From src:

bash ./train_segmentation_models.sh 

Generate Pseudo-segmentations

From src/etl:

python 06_pseudosegmentations_for_studies.py
python 07_add_pseudosegmentations_to_training.py

Retrain Segmentation Models

From src:

bash ./retrain_segmentation_models.sh

Crop Vertebra

From src/etl:

python 08_vertebra_locations_for_each_level.py
python 09_create_3d_chunk_for_each_vertebra.py

Train 3D CNN Vertebra-level Classification Models

From src:

bash ./train_x3d_classifiers.sh

Generate Slice-level Pseudo-labels

From src/etl:

python 10_get_cas.py
python 11_get_cas_pseudolabels.py
python 12_get_cropped_pngs.py

Train TD CNN Classification Models

From src:

bash ./train_tdcnn_classifiers.py

Extract Features

From src/etl:

python 13_extract_chunk_features_x3d.py
python 14_extract_chunk_features_tdcnn.py
python 15_fuse_features.py

Train Final Sequence Models

From src:

bash ./train_sequence_models.py

Inference

See Kaggle notebook: https://www.kaggle.com/code/vaillant/rsna-c-spine-submission?scriptVersionId=108890272