{
  "id": 611858,
  "title": "55st Place Solution",
  "url": "/competitions/rsna-intracranial-aneurysm-detection/discussion/611858",
  "author_name": "Less",
  "post_date": "2025-10-15T07:03:49.045000",
  "votes": 14,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Thanks to Kaggle and RSNA for hosting this competition.</p>\n<p>This is my second time participating in a competition organized by RSNA. I won a bronze medal in 2024 and reach silver this year. I will actively participate in future competitions.</p>\n<p>During the competition, we tried many methods. Based on the publicly available inference code, we built the training model code, but after many attempts, we didn't make significant progress. It wasn't until <a href=\"https://www.kaggle.com/kospintr\" target=\"_blank\">@kospintr</a> shared the training code, which included a lot of data augmentation techniques, that we retrained the model based on this code, changed the Backbone to EfficientNet V2 B3, set the learning rate to 0.0001, and used 10 - fold cross - validation to retrain 10 models. The CV (around 0.92) and LB (around 0.71) improved significantly.<br>\nThen, based on this, I adopted the TTA method shared by <a href=\"https://www.kaggle.com/sacuscreed\" target=\"_blank\">@sacuscreed</a>,  and the LB score improved by about 0.005. The code is as follows:</p>\n<p><code>final_pred = (predict_ensemble(volume) + predict_ensemble(np.flip(volume,-2))[[1,0,3,2,5,4,6,8,7,10,9,11,12,13]])/2</code></p>\n<p>We also built a YOLO - based solution as the second option, and my teammates came up with many effective suggestions. Thanks to everyone's contributions.</p>",
  "messages": [
    {
      "id": 3302150,
      "postDate": "2025-10-15T07:03:49.047Z",
      "content": "<p>Thanks to Kaggle and RSNA for hosting this competition.</p>\n<p>This is my second time participating in a competition organized by RSNA. I won a bronze medal in 2024 and reach silver this year. I will actively participate in future competitions.</p>\n<p>During the competition, we tried many methods. Based on the publicly available inference code, we built the training model code, but after many attempts, we didn't make significant progress. It wasn't until <a href=\"https://www.kaggle.com/kospintr\" target=\"_blank\">@kospintr</a> shared the training code, which included a lot of data augmentation techniques, that we retrained the model based on this code, changed the Backbone to EfficientNet V2 B3, set the learning rate to 0.0001, and used 10 - fold cross - validation to retrain 10 models. The CV (around 0.92) and LB (around 0.71) improved significantly.<br>\nThen, based on this, I adopted the TTA method shared by <a href=\"https://www.kaggle.com/sacuscreed\" target=\"_blank\">@sacuscreed</a>,  and the LB score improved by about 0.005. The code is as follows:</p>\n<p><code>final_pred = (predict_ensemble(volume) + predict_ensemble(np.flip(volume,-2))[[1,0,3,2,5,4,6,8,7,10,9,11,12,13]])/2</code></p>\n<p>We also built a YOLO - based solution as the second option, and my teammates came up with many effective suggestions. Thanks to everyone's contributions.</p>",
      "rawMarkdown": "Thanks to Kaggle and RSNA for hosting this competition.\n\nThis is my second time participating in a competition organized by RSNA. I won a bronze medal in 2024 and reach silver this year. I will actively participate in future competitions.\n\nDuring the competition, we tried many methods. Based on the publicly available inference code, we built the training model code, but after many attempts, we didn't make significant progress. It wasn't until @kospintr shared the training code, which included a lot of data augmentation techniques, that we retrained the model based on this code, changed the Backbone to EfficientNet V2 B3, set the learning rate to 0.0001, and used 10 - fold cross - validation to retrain 10 models. The CV (around 0.92) and LB (around 0.71) improved significantly.\nThen, based on this, I adopted the TTA method shared by @sacuscreed,  and the LB score improved by about 0.005. The code is as follows:\n\n`final_pred = (predict_ensemble(volume) + predict_ensemble(np.flip(volume,-2))[[1,0,3,2,5,4,6,8,7,10,9,11,12,13]])/2`\n\nWe also built a YOLO - based solution as the second option, and my teammates came up with many effective suggestions. Thanks to everyone's contributions.",
      "votes": 14
    },
    {
      "id": 3302160,
      "postDate": "2025-10-15T08:04:54.257Z",
      "content": "<p>Congratulations and thanks for the mention. I'm glad that flip plus left-right remaping worked too for you. I've read from many participants even from hosts to avoid it completely, but in my head maked sense.</p>\n<p>Personally may be I should focus more in already contrasted solutions instead tryng to reinvent the wheel. I've got stuck at .65 with aneurysm 3D segmentation.</p>",
      "rawMarkdown": "Congratulations and thanks for the mention. I'm glad that flip plus left-right remaping worked too for you. I've read from many participants even from hosts to avoid it completely, but in my head maked sense.\n\nPersonally may be I should focus more in already contrasted solutions instead tryng to reinvent the wheel. I've got stuck at .65 with aneurysm 3D segmentation.",
      "votes": 1,
      "replies": [
        {
          "id": 3302168,
          "postDate": "2025-10-15T08:26:31.603Z",
          "content": "<p>Thank you again.</p>",
          "rawMarkdown": "Thank you again."
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 3302160,
      "author_name": "Ángel Jacinto Sánchez Ruiz",
      "author_url": "",
      "post_date": "2025-10-15T08:04:54.257000",
      "content": "<p>Congratulations and thanks for the mention. I'm glad that flip plus left-right remaping worked too for you. I've read from many participants even from hosts to avoid it completely, but in my head maked sense.</p>\n<p>Personally may be I should focus more in already contrasted solutions instead tryng to reinvent the wheel. I've got stuck at .65 with aneurysm 3D segmentation.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 3302168,
          "author_name": "Less",
          "author_url": "",
          "post_date": "2025-10-15T08:26:31.603000",
          "content": "<p>Thank you again.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3302150": "Thanks to Kaggle and RSNA for hosting this competition.\n\nThis is my second time participating in a competition organized by RSNA. I won a bronze medal in 2024 and reach silver this year. I will actively participate in future competitions.\n\nDuring the competition, we tried many methods. Based on the publicly available inference code, we built the training model code, but after many attempts, we didn't make significant progress. It wasn't until @kospintr shared the training code, which included a lot of data augmentation techniques, that we retrained the model based on this code, changed the Backbone to EfficientNet V2 B3, set the learning rate to 0.0001, and used 10 - fold cross - validation to retrain 10 models. The CV (around 0.92) and LB (around 0.71) improved significantly.\nThen, based on this, I adopted the TTA method shared by @sacuscreed,  and the LB score improved by about 0.005. The code is as follows:\n\n`final_pred = (predict_ensemble(volume) + predict_ensemble(np.flip(volume,-2))[[1,0,3,2,5,4,6,8,7,10,9,11,12,13]])/2`\n\nWe also built a YOLO - based solution as the second option, and my teammates came up with many effective suggestions. Thanks to everyone's contributions.",
    "3302160": "Congratulations and thanks for the mention. I'm glad that flip plus left-right remaping worked too for you. I've read from many participants even from hosts to avoid it completely, but in my head maked sense.\n\nPersonally may be I should focus more in already contrasted solutions instead tryng to reinvent the wheel. I've got stuck at .65 with aneurysm 3D segmentation."
  }
}