{ "id": 193505, "title": "10th Place Solution with code", "url": "/competitions/rsna-str-pulmonary-embolism-detection/discussion/193505", "author_name": "OrKatz", "post_date": "2020-10-27T12:03:10.439000", "votes": 44, "comment_count": 21, "views": 0, "content": "

code - https://github.com/OrKatz7/RSNA-Pulmonary-Embolism-Detection
\nFull Pipeline
\n\"\"

\n

Overall Strategy

\n
    \n
  1. Train an image-level 2d CNN and save to hard drive features
    2. \n
  2. Train an exam-level 3d CNN and save to hard drive features
    3. \n
  3. input the 2d and 3d features into sequence model
    4. \n
\n

2D CNN Modeling

\n
    \n
  1. Data pre-processing - based on Ian Pan: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/182930
    2. \n
  2. augmentation - RandomBrightnessContrast, HorizontalFlip, ElasticTransform, GridDistortion, VerticalFlip, ShiftScaleRotate, RandomCrop
    3. \n
  3. cnn models - efficientnet-b3, efficientnet-b4, efficientnet-b5
    4. \n
\n

3D CNN Modeling
\nbased on @boliu0 pipline https://www.kaggle.com/boliu0/monai-3d-cnn-training

\n

Sequence Model
\nInput - Slice embeddings from multi 2d models + exam embeddings from 3d models
\nloss - rsna metric

", "messages": [ { "id": 1061877, "postDate": "2020-10-27T12:03:10.440Z", "content": "

code - https://github.com/OrKatz7/RSNA-Pulmonary-Embolism-Detection
\nFull Pipeline
\n\"\"

\n

Overall Strategy

\n
    \n
  1. Train an image-level 2d CNN and save to hard drive features
    2. \n
  2. Train an exam-level 3d CNN and save to hard drive features
    3. \n
  3. input the 2d and 3d features into sequence model
    4. \n
\n

2D CNN Modeling

\n
    \n
  1. Data pre-processing - based on Ian Pan: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/182930
    2. \n
  2. augmentation - RandomBrightnessContrast, HorizontalFlip, ElasticTransform, GridDistortion, VerticalFlip, ShiftScaleRotate, RandomCrop
    3. \n
  3. cnn models - efficientnet-b3, efficientnet-b4, efficientnet-b5
    4. \n
\n

3D CNN Modeling
\nbased on @boliu0 pipline https://www.kaggle.com/boliu0/monai-3d-cnn-training

\n

Sequence Model
\nInput - Slice embeddings from multi 2d models + exam embeddings from 3d models
\nloss - rsna metric

", "rawMarkdown": "**code** - https://github.com/OrKatz7/RSNA-Pulmonary-Embolism-Detection\n**Full Pipeline**\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1094066%2F2c42c343c91bc3fd09f089c43fd8c17d%2FRSNA.png?generation=1603799132907836&alt=media)\n\n**Overall Strategy**\n1. Train an image-level 2d CNN and save to hard drive features\n2. Train an exam-level 3d CNN and save to hard drive features\n3. input the 2d and 3d features into sequence model\n\n**2D CNN Modeling**\n1. Data pre-processing - based on Ian Pan: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/182930\n2. augmentation - RandomBrightnessContrast, HorizontalFlip, ElasticTransform, GridDistortion, VerticalFlip, ShiftScaleRotate, RandomCrop\n3. cnn models - efficientnet-b3, efficientnet-b4, efficientnet-b5\n\n**3D CNN Modeling**\nbased on @boliu0 pipline https://www.kaggle.com/boliu0/monai-3d-cnn-training\n\n**Sequence Model**\nInput - Slice embeddings from multi 2d models + exam embeddings from 3d models\nloss - rsna metric\n\n\n\n\n\n\n", "votes": 44 }, { "id": 1062004, "postDate": "2020-10-27T13:53:41.957Z", "content": "

Congrats for the gold medal! I'm glad to see my MONAI 3D model was used in a gold medal solution.

", "rawMarkdown": "Congrats for the gold medal! I'm glad to see my MONAI 3D model was used in a gold medal solution.", "votes": 5 }, { "id": 1061904, "postDate": "2020-10-27T12:28:35.827Z", "content": "

Congratulations! Arguably the best position you can have on this leaderboard ;).

", "rawMarkdown": "Congratulations! Arguably the best position you can have on this leaderboard ;).", "votes": 4 }, { "id": 1066340, "postDate": "2020-11-01T16:33:43.050Z", "content": "

Congratulations! Thank you for sharing your architecture, looks interesting.

", "rawMarkdown": "Congratulations! Thank you for sharing your architecture, looks interesting.", "votes": 1 }, { "id": 1063622, "postDate": "2020-10-29T06:09:04.937Z", "content": "

Congrats on your solo gold! Nice solution, surprising that you achieved accuracy with 224*224 images. Will full resolution inputs move your score up?

", "rawMarkdown": "Congrats on your solo gold! Nice solution, surprising that you achieved accuracy with 224*224 images. Will full resolution inputs move your score up?", "votes": 1, "replies": [ { "id": 1063843, "postDate": "2020-10-29T12:01:22.943Z", "content": "

According to the other solutions it is very possible. I did not try to run on a larger image size.

", "rawMarkdown": "According to the other solutions it is very possible. I did not try to run on a larger image size.", "votes": 1 } ] }, { "id": 1063027, "postDate": "2020-10-28T12:50:28.443Z", "content": "

Impressive @orkatz2 . Congrats on winning the prize money. Great work. Ensembling two diverse models crossed 9 hrs for myself, will look at your inference code to look how it has been fitted within the 9hrs run time. Congrats once again :)

", "rawMarkdown": "Impressive @orkatz2 . Congrats on winning the prize money. Great work. Ensembling two diverse models crossed 9 hrs for myself, will look at your inference code to look how it has been fitted within the 9hrs run time. Congrats once again :)", "votes": 1, "replies": [ { "id": 1063032, "postDate": "2020-10-28T12:56:22.693Z", "content": "

https://github.com/OrKatz7/RSNA-Pulmonary-Embolism-Detection/blob/main/submission.ipynb

", "rawMarkdown": "https://github.com/OrKatz7/RSNA-Pulmonary-Embolism-Detection/blob/main/submission.ipynb", "votes": 3 }, { "id": 1063061, "postDate": "2020-10-28T13:23:17.720Z", "content": "

Thanks for sharing @orkatz2

", "rawMarkdown": "Thanks for sharing @orkatz2 " } ] }, { "id": 1062713, "postDate": "2020-10-28T06:01:27.293Z", "content": "

Congratulations on one moving up one more step and becoming a prize winner :) @orkatz2

", "rawMarkdown": "Congratulations on one moving up one more step and becoming a prize winner :) @orkatz2 ", "votes": 1 }, { "id": 1062236, "postDate": "2020-10-27T17:15:53.193Z", "content": "

Congratz on the nice finish !

\n

I'm assuming all your models were single fold ones, since otherwise it would've taken ages to run, right ?

", "rawMarkdown": "Congratz on the nice finish !\n\nI'm assuming all your models were single fold ones, since otherwise it would've taken ages to run, right ?", "votes": 1, "replies": [ { "id": 1062249, "postDate": "2020-10-27T17:20:19.040Z", "content": "

B5,b4,b3 - 2 folds
\n3d densenet121 - 3 folds X 3
\nLstm - 2 folds
\nNote that my input size is not 512X512. My models were trained on 224X224 so is faster.

", "rawMarkdown": "B5,b4,b3 - 2 folds\n3d densenet121 - 3 folds X 3\nLstm - 2 folds\nNote that my input size is not 512X512. My models were trained on 224X224 so is faster.", "votes": 2 }, { "id": 1062307, "postDate": "2020-10-27T18:12:09.373Z", "content": "

That's a quite big amount of models, impressive. I'll have to check your code to see how you made that fit in the 9 hours.

", "rawMarkdown": "That's a quite big amount of models, impressive. I'll have to check your code to see how you made that fit in the 9 hours." }, { "id": 1063035, "postDate": "2020-10-28T12:58:14.230Z", "content": "

In fact it takes 7.5 hours.
\nsubmission code

", "rawMarkdown": "In fact it takes 7.5 hours.\n[submission code](https://github.com/OrKatz7/RSNA-Pulmonary-Embolism-Detection/blob/main/submission.ipynb)", "votes": 1 } ] }, { "id": 1062093, "postDate": "2020-10-27T15:08:37.970Z", "content": "

So we have had pretty much the same approach. Except you have used a sequence model to stack and we have used a decision tree :D . Good Job!

", "rawMarkdown": "So we have had pretty much the same approach. Except you have used a sequence model to stack and we have used a decision tree :D . Good Job!", "votes": 1 }, { "id": 1062068, "postDate": "2020-10-27T14:46:07.420Z", "content": "

@orkatz2 Congratulations ….great work!!

", "rawMarkdown": "@orkatz2 Congratulations ....great work!!", "votes": 1 }, { "id": 1061972, "postDate": "2020-10-27T13:26:21.633Z", "content": "

Great work @orkatz2. Congratulations !

", "rawMarkdown": "Great work @orkatz2. Congratulations !", "votes": 1 }, { "id": 1061914, "postDate": "2020-10-27T12:34:23.953Z", "content": "

Good job! Congrats on results and thanks for sharing details solution and code @orkatz2

", "rawMarkdown": "Good job! Congrats on results and thanks for sharing details solution and code @orkatz2 ", "votes": 1 }, { "id": 1061903, "postDate": "2020-10-27T12:26:44.257Z", "content": "

Congrats @orkatz2 and
\nThanks for good explanation!

", "rawMarkdown": "Congrats @orkatz2 and \nThanks for good explanation!", "votes": 1 }, { "id": 1331341, "postDate": "2021-06-01T12:08:10.930Z", "content": "

What does your output label represent

", "rawMarkdown": "What does your output label represent\n" }, { "id": 1152670, "postDate": "2021-01-14T10:57:58.443Z", "content": "

Hey @orkatz2 , late to join in, Congratulations firstly, your submission link for GitHub is broken.
\nCan you kindly provide it again if possible

", "rawMarkdown": "Hey @orkatz2 , late to join in, Congratulations firstly, your submission link for GitHub is broken.\nCan you kindly provide it again if possible", "isDeleted": true }, { "id": 1062378, "postDate": "2020-10-27T19:19:54.840Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1062004, "author_name": "Bo", "author_url": "", "post_date": "2020-10-27T13:53:41.957000", "content": "

Congrats for the gold medal! I'm glad to see my MONAI 3D model was used in a gold medal solution.

", "votes": 5, "replies": [] }, { "id": 1061904, "author_name": "Gilles Vandewiele", "author_url": "", "post_date": "2020-10-27T12:28:35.827000", "content": "

Congratulations! Arguably the best position you can have on this leaderboard ;).

", "votes": 4, "replies": [] }, { "id": 1066340, "author_name": "Alin Cijov", "author_url": "", "post_date": "2020-11-01T16:33:43.050000", "content": "

Congratulations! Thank you for sharing your architecture, looks interesting.

", "votes": 1, "replies": [] }, { "id": 1063622, "author_name": "arutema47", "author_url": "", "post_date": "2020-10-29T06:09:04.937000", "content": "

Congrats on your solo gold! Nice solution, surprising that you achieved accuracy with 224*224 images. Will full resolution inputs move your score up?

", "votes": 1, "replies": [ { "id": 1063843, "author_name": "OrKatz", "author_url": "", "post_date": "2020-10-29T12:01:22.943000", "content": "

According to the other solutions it is very possible. I did not try to run on a larger image size.

", "votes": 1, "replies": [] } ] }, { "id": 1063027, "author_name": "Manoj Prabhakar", "author_url": "", "post_date": "2020-10-28T12:50:28.443000", "content": "

Impressive @orkatz2 . Congrats on winning the prize money. Great work. Ensembling two diverse models crossed 9 hrs for myself, will look at your inference code to look how it has been fitted within the 9hrs run time. Congrats once again :)

", "votes": 1, "replies": [ { "id": 1063032, "author_name": "OrKatz", "author_url": "", "post_date": "2020-10-28T12:56:22.693000", "content": "

https://github.com/OrKatz7/RSNA-Pulmonary-Embolism-Detection/blob/main/submission.ipynb

", "votes": 3, "replies": [] }, { "id": 1063061, "author_name": "Manoj Prabhakar", "author_url": "", "post_date": "2020-10-28T13:23:17.720000", "content": "

Thanks for sharing @orkatz2

", "votes": 0, "replies": [] } ] }, { "id": 1062713, "author_name": "Eisa", "author_url": "", "post_date": "2020-10-28T06:01:27.293000", "content": "

Congratulations on one moving up one more step and becoming a prize winner :) @orkatz2

", "votes": 1, "replies": [] }, { "id": 1062236, "author_name": "Theo Viel", "author_url": "", "post_date": "2020-10-27T17:15:53.193000", "content": "

Congratz on the nice finish !

\n

I'm assuming all your models were single fold ones, since otherwise it would've taken ages to run, right ?

", "votes": 1, "replies": [ { "id": 1062249, "author_name": "OrKatz", "author_url": "", "post_date": "2020-10-27T17:20:19.040000", "content": "

B5,b4,b3 - 2 folds
\n3d densenet121 - 3 folds X 3
\nLstm - 2 folds
\nNote that my input size is not 512X512. My models were trained on 224X224 so is faster.

", "votes": 2, "replies": [] }, { "id": 1062307, "author_name": "Theo Viel", "author_url": "", "post_date": "2020-10-27T18:12:09.373000", "content": "

That's a quite big amount of models, impressive. I'll have to check your code to see how you made that fit in the 9 hours.

", "votes": 0, "replies": [] }, { "id": 1063035, "author_name": "OrKatz", "author_url": "", "post_date": "2020-10-28T12:58:14.230000", "content": "

In fact it takes 7.5 hours.
\nsubmission code

", "votes": 1, "replies": [] } ] }, { "id": 1062093, "author_name": "Jan Bre", "author_url": "", "post_date": "2020-10-27T15:08:37.970000", "content": "

So we have had pretty much the same approach. Except you have used a sequence model to stack and we have used a decision tree :D . Good Job!

", "votes": 1, "replies": [] }, { "id": 1062068, "author_name": "Gryffindor", "author_url": "", "post_date": "2020-10-27T14:46:07.420000", "content": "

@orkatz2 Congratulations ….great work!!

", "votes": 1, "replies": [] }, { "id": 1061972, "author_name": "Vee", "author_url": "", "post_date": "2020-10-27T13:26:21.633000", "content": "

Great work @orkatz2. Congratulations !

", "votes": 1, "replies": [] }, { "id": 1061914, "author_name": "KhanhVD", "author_url": "", "post_date": "2020-10-27T12:34:23.953000", "content": "

Good job! Congrats on results and thanks for sharing details solution and code @orkatz2

", "votes": 1, "replies": [] }, { "id": 1061903, "author_name": "Heroseo", "author_url": "", "post_date": "2020-10-27T12:26:44.257000", "content": "

Congrats @orkatz2 and
\nThanks for good explanation!

", "votes": 1, "replies": [] }, { "id": 1331341, "author_name": "ambitiousF", "author_url": "", "post_date": "2021-06-01T12:08:10.930000", "content": "

What does your output label represent

", "votes": 0, "replies": [] }, { "id": 1152670, "author_name": "", "author_url": "", "post_date": "2021-01-14T10:57:58.443000", "content": "

Hey @orkatz2 , late to join in, Congratulations firstly, your submission link for GitHub is broken.
\nCan you kindly provide it again if possible

", "votes": 0, "replies": [] }, { "id": 1062378, "author_name": "", "author_url": "", "post_date": "2020-10-27T19:19:54.840000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1061877": "**code** - https://github.com/OrKatz7/RSNA-Pulmonary-Embolism-Detection\n**Full Pipeline**\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1094066%2F2c42c343c91bc3fd09f089c43fd8c17d%2FRSNA.png?generation=1603799132907836&alt=media)\n\n**Overall Strategy**\n1. Train an image-level 2d CNN and save to hard drive features\n2. Train an exam-level 3d CNN and save to hard drive features\n3. input the 2d and 3d features into sequence model\n\n**2D CNN Modeling**\n1. Data pre-processing - based on Ian Pan: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/182930\n2. augmentation - RandomBrightnessContrast, HorizontalFlip, ElasticTransform, GridDistortion, VerticalFlip, ShiftScaleRotate, RandomCrop\n3. cnn models - efficientnet-b3, efficientnet-b4, efficientnet-b5\n\n**3D CNN Modeling**\nbased on @boliu0 pipline https://www.kaggle.com/boliu0/monai-3d-cnn-training\n\n**Sequence Model**\nInput - Slice embeddings from multi 2d models + exam embeddings from 3d models\nloss - rsna metric\n\n\n\n\n\n\n", "1062004": "Congrats for the gold medal! I'm glad to see my MONAI 3D model was used in a gold medal solution.", "1061904": "Congratulations! Arguably the best position you can have on this leaderboard ;).", "1066340": "Congratulations! Thank you for sharing your architecture, looks interesting.", "1063622": "Congrats on your solo gold! Nice solution, surprising that you achieved accuracy with 224*224 images. Will full resolution inputs move your score up?", "1063027": "Impressive @orkatz2 . Congrats on winning the prize money. Great work. Ensembling two diverse models crossed 9 hrs for myself, will look at your inference code to look how it has been fitted within the 9hrs run time. Congrats once again :)", "1062713": "Congratulations on one moving up one more step and becoming a prize winner :) @orkatz2 ", "1062236": "Congratz on the nice finish !\n\nI'm assuming all your models were single fold ones, since otherwise it would've taken ages to run, right ?", "1062093": "So we have had pretty much the same approach. Except you have used a sequence model to stack and we have used a decision tree :D . Good Job!", "1062068": "@orkatz2 Congratulations ....great work!!", "1061972": "Great work @orkatz2. Congratulations !", "1061914": "Good job! Congrats on results and thanks for sharing details solution and code @orkatz2 ", "1061903": "Congrats @orkatz2 and \nThanks for good explanation!", "1331341": "What does your output label represent\n", "1152670": "Hey @orkatz2 , late to join in, Congratulations firstly, your submission link for GitHub is broken.\nCan you kindly provide it again if possible", "1062378": "" } }