{ "id": 452992, "title": "514th Place Solution for the RSNA 2023 Abdominal Trauma Detection Competition", "url": "/competitions/rsna-2023-abdominal-trauma-detection/discussion/452992", "author_name": "Hiroshi Sakiyama", "post_date": "2023-11-04T10:44:20.671000", "votes": 2, "comment_count": 0, "views": 0, "content": "

First of all, I would like to thank Radiological Society of North America and Kaggle for organizing and running this competition. I would also like to thank all the competitors who shared their views and notebooks.
\nCongratulations to all the winners! Although my ranking was not good, I would like to share with you what I did and the results. I worked on the prediction using a small number of images to conserve computational resources. I made predictions only for the liver, pancreas and kidneys of a part of patients. I set my goals small, because dealing with a lot of data was likely to cause me to give up in the process.

\n

Context

\n\n

Overview of the Approach

\n

Twenty input images were selected per series_id. From the selected images, the nearest neighbors model selected images that showed the target organs, excluding images that only showed lungs, legs, etc. The remaining images were used to predict damage to the spleen, liver, and kidneys, and so on.

\n

Details of the submission

\n

First, using the notebook [1] method by Parham Mostame, 20 images (256*256 PNG images) were selected per series_id. In this process, unusual DICOM files were standardized according to the post [2] by Hui Ming Lin.
\n\"\"
\nBecause some of the selected images contained mainly non-target-related items, such as lungs and legs, only the necessary images were further selected. For this, 206 sets of data with segmentation information were used to predict images containing the liver, images containing the spleen, images containing the liver and spleen, and so on. The model was from the \"Digit Recognizer\" Competition notebook [3, 4] and used nearest neighbors learning. It runs fast, can be used without a GPU [5], and seems to have worked somewhat well.
\nThen, about one-third of the data was used to make predictions for each organ. The aforementioned nearest neighbors model was used, but was planed to be changed later to a better model. For the liver, it appeared to work somewhat well, and for the spleen, some of the results seemed to work well. The kidneys were not so good, and the bowels were not good at all. This time, I stopped here and only made final predictions for the liver, spleen, and kidneys for limited patients. Where no prediction was made, the average value was used.
\nIn the private score results, the score was better when only the liver and spleen were predicted, without including the kidney prediction. (Note: LS means liver and spleen; LSK means liver, spleen and kidneys.)
\n\"\"
\nNext time, I would like to learn more and work with a better model appropriately.

\n

Sources

\n

[1] Parham Mostame, Construct 3D arrays from DCM/NII (+ 3 view angles).
\n[2] Hui Ming Lin, Standardizing Unusual Dicoms.
\n[3] Hendrik, Top Score using Nearest Neighbours.
\n[4] Shadab Hussain, Digit Recognition using Nearest Neighbors.
\n[5] Hiroshi Sakiyama, Top Score using Nearest Neighbours without GPU.

", "messages": [ { "id": 2512298, "postDate": "2023-11-04T10:44:20.670Z", "content": "

First of all, I would like to thank Radiological Society of North America and Kaggle for organizing and running this competition. I would also like to thank all the competitors who shared their views and notebooks.
\nCongratulations to all the winners! Although my ranking was not good, I would like to share with you what I did and the results. I worked on the prediction using a small number of images to conserve computational resources. I made predictions only for the liver, pancreas and kidneys of a part of patients. I set my goals small, because dealing with a lot of data was likely to cause me to give up in the process.

\n

Context

\n\n

Overview of the Approach

\n

Twenty input images were selected per series_id. From the selected images, the nearest neighbors model selected images that showed the target organs, excluding images that only showed lungs, legs, etc. The remaining images were used to predict damage to the spleen, liver, and kidneys, and so on.

\n

Details of the submission

\n

First, using the notebook [1] method by Parham Mostame, 20 images (256*256 PNG images) were selected per series_id. In this process, unusual DICOM files were standardized according to the post [2] by Hui Ming Lin.
\n\"\"
\nBecause some of the selected images contained mainly non-target-related items, such as lungs and legs, only the necessary images were further selected. For this, 206 sets of data with segmentation information were used to predict images containing the liver, images containing the spleen, images containing the liver and spleen, and so on. The model was from the \"Digit Recognizer\" Competition notebook [3, 4] and used nearest neighbors learning. It runs fast, can be used without a GPU [5], and seems to have worked somewhat well.
\nThen, about one-third of the data was used to make predictions for each organ. The aforementioned nearest neighbors model was used, but was planed to be changed later to a better model. For the liver, it appeared to work somewhat well, and for the spleen, some of the results seemed to work well. The kidneys were not so good, and the bowels were not good at all. This time, I stopped here and only made final predictions for the liver, spleen, and kidneys for limited patients. Where no prediction was made, the average value was used.
\nIn the private score results, the score was better when only the liver and spleen were predicted, without including the kidney prediction. (Note: LS means liver and spleen; LSK means liver, spleen and kidneys.)
\n\"\"
\nNext time, I would like to learn more and work with a better model appropriately.

\n

Sources

\n

[1] Parham Mostame, Construct 3D arrays from DCM/NII (+ 3 view angles).
\n[2] Hui Ming Lin, Standardizing Unusual Dicoms.
\n[3] Hendrik, Top Score using Nearest Neighbours.
\n[4] Shadab Hussain, Digit Recognition using Nearest Neighbors.
\n[5] Hiroshi Sakiyama, Top Score using Nearest Neighbours without GPU.

", "rawMarkdown": "First of all, I would like to thank Radiological Society of North America and Kaggle for organizing and running this competition. I would also like to thank all the competitors who shared their views and notebooks.\nCongratulations to all the winners! Although my ranking was not good, I would like to share with you what I did and the results. I worked on the prediction using a small number of images to conserve computational resources. I made predictions only for the liver, pancreas and kidneys of a part of patients. I set my goals small, because dealing with a lot of data was likely to cause me to give up in the process.\n\n## **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**\nTwenty input images were selected per series_id. From the selected images, the nearest neighbors model selected images that showed the target organs, excluding images that only showed lungs, legs, etc. The remaining images were used to predict damage to the spleen, liver, and kidneys, and so on.\n\n## **Details of the submission**\nFirst, using the [notebook [1]](https://www.kaggle.com/code/parhammostame/construct-3d-arrays-from-dcm-nii-3-view-angles?scriptVersionId=138964121) method by [Parham Mostame](https://www.kaggle.com/parhammostame), 20 images (256*256 PNG images) were selected per series_id. In this process, unusual DICOM files were standardized according to the [post [2]](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427217) by [Hui Ming Lin](https://www.kaggle.com/huiminglin). \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5800072%2F1afcba921294d620d637fa28cb1c681c%2Fcor.png?generation=1699094504090575&alt=media)\nBecause some of the selected images contained mainly non-target-related items, such as lungs and legs, only the necessary images were further selected. For this, 206 sets of data with segmentation information were used to predict images containing the liver, images containing the spleen, images containing the liver and spleen, and so on. The model was from the [\"Digit Recognizer\" Competition] (https://www.kaggle.com/competitions/digit-recognizer) notebook [[3](https://www.kaggle.com/code/t0m0ff3l/top-score-using-nearest-neighbours?scriptVersionId=12826678), [4](https://www.kaggle.com/code/shadabhussain/digit-recognition-using-nearest-neighbors)] and used nearest neighbors learning. It runs fast, can be used without a GPU [[5](https://www.kaggle.com/code/hiroshisakiyama/top-score-using-nearest-neighbours-without-gpu?scriptVersionId=146578718)], and seems to have worked somewhat well. \nThen, about one-third of the data was used to make predictions for each organ. The aforementioned nearest neighbors model was used, but was planed to be changed later to a better model. For the liver, it appeared to work somewhat well, and for the spleen, some of the results seemed to work well. The kidneys were not so good, and the bowels were not good at all. This time, I stopped here and only made final predictions for the liver, spleen, and kidneys for limited patients. Where no prediction was made, the average value was used.\nIn the private score results, the score was better when only the liver and spleen were predicted, without including the kidney prediction. (Note: LS means liver and spleen; LSK means liver, spleen and kidneys.)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5800072%2F061e1fe83f5391b7e5763477f98fa576%2FRSNAscore.png?generation=1699094621201857&alt=media)\nNext time, I would like to learn more and work with a better model appropriately.\n\n## **Sources**\n[1] [Parham Mostame](https://www.kaggle.com/parhammostame), [Construct 3D arrays from DCM/NII (+ 3 view angles)](https://www.kaggle.com/code/parhammostame/construct-3d-arrays-from-dcm-nii-3-view-angles?scriptVersionId=138964121).\n[2] [Hui Ming Lin](https://www.kaggle.com/huiminglin), [Standardizing Unusual Dicoms](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427217).\n[3] [Hendrik](https://www.kaggle.com/t0m0ff3l), [Top Score using Nearest Neighbours](https://www.kaggle.com/code/t0m0ff3l/top-score-using-nearest-neighbours?scriptVersionId=12826678).\n[4] [Shadab Hussain](https://www.kaggle.com/shadabhussain), [Digit Recognition using Nearest Neighbors](https://www.kaggle.com/code/shadabhussain/digit-recognition-using-nearest-neighbors).\n[5] [Hiroshi Sakiyama](https://www.kaggle.com/hiroshisakiyama), [Top Score using Nearest Neighbours without GPU](https://www.kaggle.com/code/hiroshisakiyama/top-score-using-nearest-neighbours-without-gpu?scriptVersionId=146578718).\n", "votes": 2 } ], "comments": [], "raw_markdown_by_id": { "2512298": "First of all, I would like to thank Radiological Society of North America and Kaggle for organizing and running this competition. I would also like to thank all the competitors who shared their views and notebooks.\nCongratulations to all the winners! Although my ranking was not good, I would like to share with you what I did and the results. I worked on the prediction using a small number of images to conserve computational resources. I made predictions only for the liver, pancreas and kidneys of a part of patients. I set my goals small, because dealing with a lot of data was likely to cause me to give up in the process.\n\n## **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**\nTwenty input images were selected per series_id. From the selected images, the nearest neighbors model selected images that showed the target organs, excluding images that only showed lungs, legs, etc. The remaining images were used to predict damage to the spleen, liver, and kidneys, and so on.\n\n## **Details of the submission**\nFirst, using the [notebook [1]](https://www.kaggle.com/code/parhammostame/construct-3d-arrays-from-dcm-nii-3-view-angles?scriptVersionId=138964121) method by [Parham Mostame](https://www.kaggle.com/parhammostame), 20 images (256*256 PNG images) were selected per series_id. In this process, unusual DICOM files were standardized according to the [post [2]](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427217) by [Hui Ming Lin](https://www.kaggle.com/huiminglin). \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5800072%2F1afcba921294d620d637fa28cb1c681c%2Fcor.png?generation=1699094504090575&alt=media)\nBecause some of the selected images contained mainly non-target-related items, such as lungs and legs, only the necessary images were further selected. For this, 206 sets of data with segmentation information were used to predict images containing the liver, images containing the spleen, images containing the liver and spleen, and so on. The model was from the [\"Digit Recognizer\" Competition] (https://www.kaggle.com/competitions/digit-recognizer) notebook [[3](https://www.kaggle.com/code/t0m0ff3l/top-score-using-nearest-neighbours?scriptVersionId=12826678), [4](https://www.kaggle.com/code/shadabhussain/digit-recognition-using-nearest-neighbors)] and used nearest neighbors learning. It runs fast, can be used without a GPU [[5](https://www.kaggle.com/code/hiroshisakiyama/top-score-using-nearest-neighbours-without-gpu?scriptVersionId=146578718)], and seems to have worked somewhat well. \nThen, about one-third of the data was used to make predictions for each organ. The aforementioned nearest neighbors model was used, but was planed to be changed later to a better model. For the liver, it appeared to work somewhat well, and for the spleen, some of the results seemed to work well. The kidneys were not so good, and the bowels were not good at all. This time, I stopped here and only made final predictions for the liver, spleen, and kidneys for limited patients. Where no prediction was made, the average value was used.\nIn the private score results, the score was better when only the liver and spleen were predicted, without including the kidney prediction. (Note: LS means liver and spleen; LSK means liver, spleen and kidneys.)\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5800072%2F061e1fe83f5391b7e5763477f98fa576%2FRSNAscore.png?generation=1699094621201857&alt=media)\nNext time, I would like to learn more and work with a better model appropriately.\n\n## **Sources**\n[1] [Parham Mostame](https://www.kaggle.com/parhammostame), [Construct 3D arrays from DCM/NII (+ 3 view angles)](https://www.kaggle.com/code/parhammostame/construct-3d-arrays-from-dcm-nii-3-view-angles?scriptVersionId=138964121).\n[2] [Hui Ming Lin](https://www.kaggle.com/huiminglin), [Standardizing Unusual Dicoms](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/427217).\n[3] [Hendrik](https://www.kaggle.com/t0m0ff3l), [Top Score using Nearest Neighbours](https://www.kaggle.com/code/t0m0ff3l/top-score-using-nearest-neighbours?scriptVersionId=12826678).\n[4] [Shadab Hussain](https://www.kaggle.com/shadabhussain), [Digit Recognition using Nearest Neighbors](https://www.kaggle.com/code/shadabhussain/digit-recognition-using-nearest-neighbors).\n[5] [Hiroshi Sakiyama](https://www.kaggle.com/hiroshisakiyama), [Top Score using Nearest Neighbours without GPU](https://www.kaggle.com/code/hiroshisakiyama/top-score-using-nearest-neighbours-without-gpu?scriptVersionId=146578718).\n" } }