I wanted to publish one of the solutions that we used to tackle the competition task.
\nThe pipeline loads 20 equally distant slices for each MRI type and stacks them together in 80 channels in total (4x20). It performs independent data augmentations on each MRI type together with randomly removing some of them on the fly. It also randomly chooses an index of the first slice after which all others are selected with the same distances between each other.
Our validation AUC was 0.706. Although it’s pointless to refer to any of the scores because of obvious reasons, I thought that some people might find this code useful: Multimodal 2d CNN with independent augmentations - training
\nThe framework Monai is also available here. The most recent version is 0.7.0.
", "messages": [ { "id": "1561461", "postDate": "10/27/2021 14:57:51", "content": "I wanted to publish one of the solutions that we used to tackle the competition task.
\nThe pipeline loads 20 equally distant slices for each MRI type and stacks them together in 80 channels in total (4x20). It performs independent data augmentations on each MRI type together with randomly removing some of them on the fly. It also randomly chooses an index of the first slice after which all others are selected with the same distances between each other.
Our validation AUC was 0.706. Although it’s pointless to refer to any of the scores because of obvious reasons, I thought that some people might find this code useful: Multimodal 2d CNN with independent augmentations - training
\nThe framework Monai is also available here. The most recent version is 0.7.0.
", "rawMarkdown": "I wanted to publish one of the solutions that we used to tackle the competition task.\nThe pipeline loads 20 equally distant slices for each MRI type and stacks them together in 80 channels in total (4x20). It performs independent data augmentations on each MRI type together with randomly removing some of them on the fly. It also randomly chooses an index of the first slice after which all others are selected with the same distances between each other.\n\nOur validation AUC was 0.706. Although it’s pointless to refer to any of the scores because of obvious reasons, I thought that some people might find this code useful: [Multimodal 2d CNN with independent augmentations - training](https://www.kaggle.com/mikecho/rsna-miccai-monai-multimodal-2d-cnn-training)\n\nThe framework Monai is also available [here](https://www.kaggle.com/mikecho/monai-v060-deep-learning-in-healthcare-imaging). The most recent version is 0.7.0.", "votes": null }, { "id": "1565791", "postDate": "10/31/2021 02:57:01", "content": "Really appreciate your work, thank you.
\nI learned a lot from your posts during the competition, I implemented all your ideas from the very first notebook, and Deep-Auc was a great resource.
\nThank you again.
I'm very happy to hear that! Thanks for this kind feedback!
", "rawMarkdown": "I'm very happy to hear that! Thanks for this kind feedback!", "votes": null } ], "comments": [ { "id": 1565791, "author_name": "vultureteam", "author_url": "", "post_date": "10/31/2021 02:57:01", "content": "Really appreciate your work, thank you.
\nI learned a lot from your posts during the competition, I implemented all your ideas from the very first notebook, and Deep-Auc was a great resource.
\nThank you again.
I'm very happy to hear that! Thanks for this kind feedback!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1561461": "I wanted to publish one of the solutions that we used to tackle the competition task.\nThe pipeline loads 20 equally distant slices for each MRI type and stacks them together in 80 channels in total (4x20). It performs independent data augmentations on each MRI type together with randomly removing some of them on the fly. It also randomly chooses an index of the first slice after which all others are selected with the same distances between each other.\n\nOur validation AUC was 0.706. Although it’s pointless to refer to any of the scores because of obvious reasons, I thought that some people might find this code useful: [Multimodal 2d CNN with independent augmentations - training](https://www.kaggle.com/mikecho/rsna-miccai-monai-multimodal-2d-cnn-training)\n\nThe framework Monai is also available [here](https://www.kaggle.com/mikecho/monai-v060-deep-learning-in-healthcare-imaging). The most recent version is 0.7.0.", "1565791": "Really appreciate your work, thank you. \nI learned a lot from your posts during the competition, I implemented all your ideas from the very first notebook, and Deep-Auc was a great resource.\nThank you again.", "1566047": "I'm very happy to hear that! Thanks for this kind feedback!" }, "source": "meta" }