{ "id": 453631, "title": "73rd Place Solution for the RSNA 2023 Abdominal Trauma Detection Competition", "url": "/competitions/rsna-2023-abdominal-trauma-detection/discussion/453631", "author_name": "Mike", "post_date": "2023-11-07T04:57:20.171000", "votes": 2, "comment_count": 0, "views": 0, "content": "

Context

\n\n

Overview of the Approach

\n

The approach I used was based off of previous RSNA competitions, and is similar to some of the approaches used in this competition, but much simpler.

\n

Essentially, I used a 2D model to extract features from each image slice, followed by a classifier.

\n

Details of the Submission

\n

I used a slightly modified version of https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion to preprocess images.

\n

Following the solutions from previous competitions, I used a 2D model (efficientnet) for feature extraction on each slice. The stack of features was then input into a CNN or RNN (I didn't see much difference in score between the two), again, just like the previous competition.

\n

This alone was able to just beat the mean baseline.

\n

I later implemented a U-Net for organ segmentation and used it to extract the slices containing the organs of interest. Performing this step before 2D slice model further raised my score to where I am currently on the LB.

\n

I also wanted to try extracting each organ and creating a model per organ, but ended up not having enough time to do so.

\n

One thing that didn't work well for me was image augmentations. I'm not sure why this is, as it seemed to work well for other teams.

\n

Sources

\n

Some useful notebooks and discussion topics
\nhttps://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion
\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053
\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/441557

", "messages": [ { "id": 2515641, "postDate": "2023-11-07T04:57:20.170Z", "content": "

Context

\n\n

Overview of the Approach

\n

The approach I used was based off of previous RSNA competitions, and is similar to some of the approaches used in this competition, but much simpler.

\n

Essentially, I used a 2D model to extract features from each image slice, followed by a classifier.

\n

Details of the Submission

\n

I used a slightly modified version of https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion to preprocess images.

\n

Following the solutions from previous competitions, I used a 2D model (efficientnet) for feature extraction on each slice. The stack of features was then input into a CNN or RNN (I didn't see much difference in score between the two), again, just like the previous competition.

\n

This alone was able to just beat the mean baseline.

\n

I later implemented a U-Net for organ segmentation and used it to extract the slices containing the organs of interest. Performing this step before 2D slice model further raised my score to where I am currently on the LB.

\n

I also wanted to try extracting each organ and creating a model per organ, but ended up not having enough time to do so.

\n

One thing that didn't work well for me was image augmentations. I'm not sure why this is, as it seemed to work well for other teams.

\n

Sources

\n

Some useful notebooks and discussion topics
\nhttps://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion
\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053
\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/441557

", "rawMarkdown": "# Context\n- Business context: [https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/overview](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/overview)\n- Data context: [https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/data](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/data)\n\n# Overview of the Approach\nThe approach I used was based off of previous RSNA competitions, and is similar to some of the approaches used in this competition, but much simpler.\n\nEssentially, I used a 2D model to extract features from each image slice, followed by a classifier.\n\n# Details of the Submission\nI used a slightly modified version of [https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion](https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion) to preprocess images.\n\nFollowing the solutions from previous competitions, I used a 2D model (efficientnet) for feature extraction on each slice. The stack of features was then input into a CNN or RNN (I didn't see much difference in score between the two), again, just like the previous competition.\n\nThis alone was able to just beat the mean baseline.\n\nI later implemented a U-Net for organ segmentation and used it to extract the slices containing the organs of interest. Performing this step before 2D slice model further raised my score to where I am currently on the LB.\n\nI also wanted to try extracting each organ and creating a model per organ, but ended up not having enough time to do so.\n\nOne thing that didn't work well for me was image augmentations. I'm not sure why this is, as it seemed to work well for other teams.\n\n# Sources\nSome useful notebooks and discussion topics\n[https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion](https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion)\n[https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053)\n[https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/441557](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/441557)", "votes": 2 } ], "comments": [], "raw_markdown_by_id": { "2515641": "# Context\n- Business context: [https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/overview](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/overview)\n- Data context: [https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/data](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/data)\n\n# Overview of the Approach\nThe approach I used was based off of previous RSNA competitions, and is similar to some of the approaches used in this competition, but much simpler.\n\nEssentially, I used a 2D model to extract features from each image slice, followed by a classifier.\n\n# Details of the Submission\nI used a slightly modified version of [https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion](https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion) to preprocess images.\n\nFollowing the solutions from previous competitions, I used a 2D model (efficientnet) for feature extraction on each slice. The stack of features was then input into a CNN or RNN (I didn't see much difference in score between the two), again, just like the previous competition.\n\nThis alone was able to just beat the mean baseline.\n\nI later implemented a U-Net for organ segmentation and used it to extract the slices containing the organs of interest. Performing this step before 2D slice model further raised my score to where I am currently on the LB.\n\nI also wanted to try extracting each organ and creating a model per organ, but ended up not having enough time to do so.\n\nOne thing that didn't work well for me was image augmentations. I'm not sure why this is, as it seemed to work well for other teams.\n\n# Sources\nSome useful notebooks and discussion topics\n[https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion](https://www.kaggle.com/code/theoviel/get-started-quicker-dicom-png-conversion)\n[https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053)\n[https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/441557](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/441557)" } }