{ "id": 359562, "title": "6th place solution", "url": "/competitions/mayo-clinic-strip-ai/discussion/359562", "author_name": "Kodai", "post_date": "2022-10-12T15:49:12.527000", "votes": 5, "comment_count": 1, "views": 0, "content": "

Hi, this competition is my first challenge at kaggle. I really enjoyed this competition and thought hard for about three months. I don't know why I could obtain a gold medal, but I'm sure it's thanks to everyone who have shared the notebooks or suggestions.
\nHere I describe my solution in brief.

\n

Submission kernel:
\nhttps://www.kaggle.com/code/kodaihatayama/mayo02-submission

\n

Preprocessing

\n

I created tiles with the following steps. I didn't resize the image in all preprocessing steps.

\n\n

For using pyvips, I refered to the following kernel. Thank you, @analokamus !
\nhttps://www.kaggle.com/code/analokamus/how-to-use-pyvips-offline

\n

Validation and Undersampling

\n

I used the hold-out method with undersampling.

\n\n

Model

\n

I created the ensemble model of the following 2 methods.

\n\n

The difference of private LB between CNN only model and the ensemble model was a bit (approximately 0.676 → 0.668).
\nAs a side note, I created a model per institution (center id). (This approach may not be effective against test dataset because the institutions of test may be different from them of train.)

\n

For CNN, I refered to the following kernel. Thank you, @vbookshelf !
\nhttps://www.kaggle.com/code/vbookshelf/cnn-how-to-use-160-000-images-without-crashing

\n

What didn't work for me

\n\n

Thanks for reading.

", "messages": [ { "id": 1984324, "postDate": "2022-10-12T15:49:12.527Z", "content": "

Hi, this competition is my first challenge at kaggle. I really enjoyed this competition and thought hard for about three months. I don't know why I could obtain a gold medal, but I'm sure it's thanks to everyone who have shared the notebooks or suggestions.
\nHere I describe my solution in brief.

\n

Submission kernel:
\nhttps://www.kaggle.com/code/kodaihatayama/mayo02-submission

\n

Preprocessing

\n

I created tiles with the following steps. I didn't resize the image in all preprocessing steps.

\n\n

For using pyvips, I refered to the following kernel. Thank you, @analokamus !
\nhttps://www.kaggle.com/code/analokamus/how-to-use-pyvips-offline

\n

Validation and Undersampling

\n

I used the hold-out method with undersampling.

\n\n

Model

\n

I created the ensemble model of the following 2 methods.

\n\n

The difference of private LB between CNN only model and the ensemble model was a bit (approximately 0.676 → 0.668).
\nAs a side note, I created a model per institution (center id). (This approach may not be effective against test dataset because the institutions of test may be different from them of train.)

\n

For CNN, I refered to the following kernel. Thank you, @vbookshelf !
\nhttps://www.kaggle.com/code/vbookshelf/cnn-how-to-use-160-000-images-without-crashing

\n

What didn't work for me

\n\n

Thanks for reading.

", "rawMarkdown": "Hi, this competition is my first challenge at kaggle. I really enjoyed this competition and thought hard for about three months. I don't know why I could obtain a gold medal, but I'm sure it's thanks to everyone who have shared the notebooks or suggestions. \nHere I describe my solution in brief.\n\nSubmission kernel:\nhttps://www.kaggle.com/code/kodaihatayama/mayo02-submission\n\n### Preprocessing\nI created tiles with the following steps. **I didn't resize the image** in all preprocessing steps.\n- Cut 6 rectangles of width 512 at equal intervals from every image \n(The purpose of this step is to create tiles quickly and avoid memory crashes without resizing.)\n- Make tiles of size 512x512 from the above rectangles and pick up the top 8 dark tiles per image\n- Split each tile to instances of size 32x32 for training\n\nFor using pyvips, I refered to the following kernel. Thank you, @analokamus !\nhttps://www.kaggle.com/code/analokamus/how-to-use-pyvips-offline\n\n### Validation and Undersampling\nI used the hold-out method with undersampling. \n- Split patient id to train and validation (80:20) which was stratified by class (CE, LAA)\n- Perform undersampling to be CE:LAA = 1:1 on the train dataset only\n\n### Model\nI created the ensemble model of the following 2 methods.\n - CNN \n - lightGBM\n\nThe difference of private LB between CNN only model and the ensemble model was a bit (approximately 0.676 → 0.668).\nAs a side note, I created a model per institution (center id). (This approach may not be effective against test dataset because the institutions of test may be different from them of train.)\n\nFor CNN, I refered to the following kernel. Thank you, @vbookshelf !\nhttps://www.kaggle.com/code/vbookshelf/cnn-how-to-use-160-000-images-without-crashing\n\n### What didn't work for me\n- grayscale\n- histogram equalization\n\nThanks for reading.", "votes": 5 }, { "id": 1985062, "postDate": "2022-10-13T04:38:34.837Z", "content": "

Congratulations @kodaihatayama. Solo gold in your first competition is a fantastic result!

", "rawMarkdown": "Congratulations @kodaihatayama. Solo gold in your first competition is a fantastic result!", "votes": 1 } ], "comments": [ { "id": 1985062, "author_name": "vbookshelf", "author_url": "", "post_date": "2022-10-13T04:38:34.837000", "content": "

Congratulations @kodaihatayama. Solo gold in your first competition is a fantastic result!

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1984324": "Hi, this competition is my first challenge at kaggle. I really enjoyed this competition and thought hard for about three months. I don't know why I could obtain a gold medal, but I'm sure it's thanks to everyone who have shared the notebooks or suggestions. \nHere I describe my solution in brief.\n\nSubmission kernel:\nhttps://www.kaggle.com/code/kodaihatayama/mayo02-submission\n\n### Preprocessing\nI created tiles with the following steps. **I didn't resize the image** in all preprocessing steps.\n- Cut 6 rectangles of width 512 at equal intervals from every image \n(The purpose of this step is to create tiles quickly and avoid memory crashes without resizing.)\n- Make tiles of size 512x512 from the above rectangles and pick up the top 8 dark tiles per image\n- Split each tile to instances of size 32x32 for training\n\nFor using pyvips, I refered to the following kernel. Thank you, @analokamus !\nhttps://www.kaggle.com/code/analokamus/how-to-use-pyvips-offline\n\n### Validation and Undersampling\nI used the hold-out method with undersampling. \n- Split patient id to train and validation (80:20) which was stratified by class (CE, LAA)\n- Perform undersampling to be CE:LAA = 1:1 on the train dataset only\n\n### Model\nI created the ensemble model of the following 2 methods.\n - CNN \n - lightGBM\n\nThe difference of private LB between CNN only model and the ensemble model was a bit (approximately 0.676 → 0.668).\nAs a side note, I created a model per institution (center id). (This approach may not be effective against test dataset because the institutions of test may be different from them of train.)\n\nFor CNN, I refered to the following kernel. Thank you, @vbookshelf !\nhttps://www.kaggle.com/code/vbookshelf/cnn-how-to-use-160-000-images-without-crashing\n\n### What didn't work for me\n- grayscale\n- histogram equalization\n\nThanks for reading.", "1985062": "Congratulations @kodaihatayama. Solo gold in your first competition is a fantastic result!" } }