{
  "id": 452255,
  "title": "63rd Place Solution for the RSNA 2023 Abdominal Trauma Detection",
  "url": "/competitions/rsna-2023-abdominal-trauma-detection/discussion/452255",
  "author_name": "sho1_24",
  "post_date": "2023-11-01T13:40:54.265000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I created my solution following the advice of the <a href=\"https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053#2425570\" target=\"_blank\">discussion</a>. My solution is not the solution of the top winners, but I hope it will be of some help. When participating in this competition, I referred to various public notebooks and data. I would like to express my appreciation to everyone who was willing to post.</p>\n<h2>Overall solution</h2>\n<ul>\n<li><p>Trauma detection for all organs is performed using a simple single-stage architecture. Then, detection for the three organs is performed using a multi-stage architecture, and finally ensembled.</p></li>\n<li><p>single-stage architecture</p>\n<ul>\n<li>using <a href=\"https://arxiv.org/ftp/arxiv/papers/2002/2002.04752.pdf\" target=\"_blank\">CT-Net</a></li></ul></li>\n<li><p>multi-stage architecture</p>\n<ul>\n<li>for kidney+liver+spleen</li>\n<li>stage1 : sub-volume prediction using EfficientNet-v1b0</li>\n<li>stage2 : prediction using CT-Net</li></ul></li>\n</ul>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3593902%2Fd6f113ee7831e5e087d7eeb7148640ca%2F1.PNG?generation=1698845265814490&amp;alt=media\" alt=\"\"></p>\n<p>I was doing trial and error right up until the deadline, so I ended up with a half-finished architecture configuration.<br>\nIn the case of only single-stage architecture, the score was LB0.61 without tuning.</p>\n<h2>notebook</h2>\n<ul>\n<li><a href=\"https://www.kaggle.com/code/sho124/rsna-infer-notebook\" target=\"_blank\">inference notebook</a></li>\n</ul>\n<h2>Impression</h2>\n<ul>\n<li>I had no experience implementing recognition technology using 3D data. So, this competition was very challenging and educational for me. </li>\n<li>My best solution was created during the two-day extension period. I think this two-day extension was controversial. But it was a precious two days where I was able to get the best score for me. As a result I was able to get my first medal.</li>\n</ul>",
  "messages": [
    {
      "id": 2508121,
      "postDate": "2023-11-01T13:40:54.267Z",
      "content": "<p>I created my solution following the advice of the <a href=\"https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053#2425570\" target=\"_blank\">discussion</a>. My solution is not the solution of the top winners, but I hope it will be of some help. When participating in this competition, I referred to various public notebooks and data. I would like to express my appreciation to everyone who was willing to post.</p>\n<h2>Overall solution</h2>\n<ul>\n<li><p>Trauma detection for all organs is performed using a simple single-stage architecture. Then, detection for the three organs is performed using a multi-stage architecture, and finally ensembled.</p></li>\n<li><p>single-stage architecture</p>\n<ul>\n<li>using <a href=\"https://arxiv.org/ftp/arxiv/papers/2002/2002.04752.pdf\" target=\"_blank\">CT-Net</a></li></ul></li>\n<li><p>multi-stage architecture</p>\n<ul>\n<li>for kidney+liver+spleen</li>\n<li>stage1 : sub-volume prediction using EfficientNet-v1b0</li>\n<li>stage2 : prediction using CT-Net</li></ul></li>\n</ul>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3593902%2Fd6f113ee7831e5e087d7eeb7148640ca%2F1.PNG?generation=1698845265814490&amp;alt=media\" alt=\"\"></p>\n<p>I was doing trial and error right up until the deadline, so I ended up with a half-finished architecture configuration.<br>\nIn the case of only single-stage architecture, the score was LB0.61 without tuning.</p>\n<h2>notebook</h2>\n<ul>\n<li><a href=\"https://www.kaggle.com/code/sho124/rsna-infer-notebook\" target=\"_blank\">inference notebook</a></li>\n</ul>\n<h2>Impression</h2>\n<ul>\n<li>I had no experience implementing recognition technology using 3D data. So, this competition was very challenging and educational for me. </li>\n<li>My best solution was created during the two-day extension period. I think this two-day extension was controversial. But it was a precious two days where I was able to get the best score for me. As a result I was able to get my first medal.</li>\n</ul>",
      "rawMarkdown": "I created my solution following the advice of the [discussion](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053#2425570). My solution is not the solution of the top winners, but I hope it will be of some help. When participating in this competition, I referred to various public notebooks and data. I would like to express my appreciation to everyone who was willing to post.\n\n## Overall solution  \n- Trauma detection for all organs is performed using a simple single-stage architecture. Then, detection for the three organs is performed using a multi-stage architecture, and finally ensembled.\n\n- single-stage architecture\n - using [CT-Net](https://arxiv.org/ftp/arxiv/papers/2002/2002.04752.pdf)\n- multi-stage architecture\n - for kidney+liver+spleen\n - stage1 : sub-volume prediction using EfficientNet-v1b0\n - stage2 : prediction using CT-Net\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3593902%2Fd6f113ee7831e5e087d7eeb7148640ca%2F1.PNG?generation=1698845265814490&alt=media)\n\nI was doing trial and error right up until the deadline, so I ended up with a half-finished architecture configuration.\nIn the case of only single-stage architecture, the score was LB0.61 without tuning.\n\n## notebook\n- [inference notebook](https://www.kaggle.com/code/sho124/rsna-infer-notebook)\n\n## Impression\n- I had no experience implementing recognition technology using 3D data. So, this competition was very challenging and educational for me. \n- My best solution was created during the two-day extension period. I think this two-day extension was controversial. But it was a precious two days where I was able to get the best score for me. As a result I was able to get my first medal.",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2508121": "I created my solution following the advice of the [discussion](https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053#2425570). My solution is not the solution of the top winners, but I hope it will be of some help. When participating in this competition, I referred to various public notebooks and data. I would like to express my appreciation to everyone who was willing to post.\n\n## Overall solution  \n- Trauma detection for all organs is performed using a simple single-stage architecture. Then, detection for the three organs is performed using a multi-stage architecture, and finally ensembled.\n\n- single-stage architecture\n - using [CT-Net](https://arxiv.org/ftp/arxiv/papers/2002/2002.04752.pdf)\n- multi-stage architecture\n - for kidney+liver+spleen\n - stage1 : sub-volume prediction using EfficientNet-v1b0\n - stage2 : prediction using CT-Net\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3593902%2Fd6f113ee7831e5e087d7eeb7148640ca%2F1.PNG?generation=1698845265814490&alt=media)\n\nI was doing trial and error right up until the deadline, so I ended up with a half-finished architecture configuration.\nIn the case of only single-stage architecture, the score was LB0.61 without tuning.\n\n## notebook\n- [inference notebook](https://www.kaggle.com/code/sho124/rsna-infer-notebook)\n\n## Impression\n- I had no experience implementing recognition technology using 3D data. So, this competition was very challenging and educational for me. \n- My best solution was created during the two-day extension period. I think this two-day extension was controversial. But it was a precious two days where I was able to get the best score for me. As a result I was able to get my first medal."
  }
}