{
  "id": 70338,
  "title": "Solution share?",
  "url": "/competitions/rsna-pneumonia-detection-challenge/discussion/70338",
  "author_name": "",
  "post_date": "2018-11-02T09:40:51.791265700Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I did poorly with Mask-RCNN, but I'm curious to know what techniques the winners used.</p>",
  "messages": [
    {
      "id": "414193",
      "postDate": "11/02/2018 09:40:51",
      "content": "<p>I did poorly with Mask-RCNN, but I'm curious to know what techniques the winners used.</p>",
      "rawMarkdown": "I did poorly with Mask-RCNN, but I'm curious to know what techniques the winners used.",
      "votes": null
    },
    {
      "id": "414392",
      "postDate": "11/02/2018 17:08:06",
      "content": "<p>Also very curious to see the approach that the top guys were taken. </p>\n\n<p>For my part, didn't score too highly but I used the following:</p>\n\n<p><strong>Object Detection:</strong>\nRetinaNet (ResNet152 backbone, positive instances only, prediction confidence &gt; 0.32) <br>\n<strong>Classification:</strong> \nXception/MobileNet (binary target, pre-trained imagenet, fine-tuned, avg of 3 models) </p>\n\n<p>I found that combining the object detection results with a pneumonia-specific classifier gave better results than simply relying on the former, regardless of the algorithm (tried YOLO v3 &amp; Mask-RCNN too). Training RetinaNet with just the positive cases also meant faster experimentation when it came to changing parameters etc.</p>\n\n<p>Two Xception models were very similar but trained with different input sizes and I made some small modifications to the penultimate pre-head layers of one of the models. MobileNet I threw in at the end simply because it trained fast and seemed to improve local validation when combined in ensemble.</p>\n\n<p>I didn't do any kind of pre-processing and trained on 90% of the original training data. </p>",
      "rawMarkdown": "Also very curious to see the approach that the top guys were taken. \n\nFor my part, didn't score too highly but I used the following:\n\n__Object Detection:__\nRetinaNet (ResNet152 backbone, positive instances only, prediction confidence &gt; 0.32)  \n__Classification:__ \nXception/MobileNet (binary target, pre-trained imagenet, fine-tuned, avg of 3 models) \n\nI found that combining the object detection results with a pneumonia-specific classifier gave better results than simply relying on the former, regardless of the algorithm (tried YOLO v3 &amp; Mask-RCNN too). Training RetinaNet with just the positive cases also meant faster experimentation when it came to changing parameters etc.\n\nTwo Xception models were very similar but trained with different input sizes and I made some small modifications to the penultimate pre-head layers of one of the models. MobileNet I threw in at the end simply because it trained fast and seemed to improve local validation when combined in ensemble.\n\nI didn't do any kind of pre-processing and trained on 90% of the original training data.",
      "votes": null
    },
    {
      "id": "414394",
      "postDate": "11/02/2018 17:10:13",
      "content": "<p>We will be posting our solution overview this weekend. This week has been quite busy for the both of us so we haven't had much time. Sorry for the wait!</p>\n\n<p>edit: posted here <a href=\"https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421\">https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421</a></p>",
      "rawMarkdown": "We will be posting our solution overview this weekend. This week has been quite busy for the both of us so we haven't had much time. Sorry for the wait!\n\nedit: posted here https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 414392,
      "author_name": "taindow",
      "author_url": "",
      "post_date": "11/02/2018 17:08:06",
      "content": "<p>Also very curious to see the approach that the top guys were taken. </p>\n\n<p>For my part, didn't score too highly but I used the following:</p>\n\n<p><strong>Object Detection:</strong>\nRetinaNet (ResNet152 backbone, positive instances only, prediction confidence &gt; 0.32) <br>\n<strong>Classification:</strong> \nXception/MobileNet (binary target, pre-trained imagenet, fine-tuned, avg of 3 models) </p>\n\n<p>I found that combining the object detection results with a pneumonia-specific classifier gave better results than simply relying on the former, regardless of the algorithm (tried YOLO v3 &amp; Mask-RCNN too). Training RetinaNet with just the positive cases also meant faster experimentation when it came to changing parameters etc.</p>\n\n<p>Two Xception models were very similar but trained with different input sizes and I made some small modifications to the penultimate pre-head layers of one of the models. MobileNet I threw in at the end simply because it trained fast and seemed to improve local validation when combined in ensemble.</p>\n\n<p>I didn't do any kind of pre-processing and trained on 90% of the original training data. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 414394,
      "author_name": "vaillant",
      "author_url": "",
      "post_date": "11/02/2018 17:10:13",
      "content": "<p>We will be posting our solution overview this weekend. This week has been quite busy for the both of us so we haven't had much time. Sorry for the wait!</p>\n\n<p>edit: posted here <a href=\"https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421\">https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "414193": "I did poorly with Mask-RCNN, but I'm curious to know what techniques the winners used.",
    "414392": "Also very curious to see the approach that the top guys were taken. \n\nFor my part, didn't score too highly but I used the following:\n\n__Object Detection:__\nRetinaNet (ResNet152 backbone, positive instances only, prediction confidence &gt; 0.32)  \n__Classification:__ \nXception/MobileNet (binary target, pre-trained imagenet, fine-tuned, avg of 3 models) \n\nI found that combining the object detection results with a pneumonia-specific classifier gave better results than simply relying on the former, regardless of the algorithm (tried YOLO v3 &amp; Mask-RCNN too). Training RetinaNet with just the positive cases also meant faster experimentation when it came to changing parameters etc.\n\nTwo Xception models were very similar but trained with different input sizes and I made some small modifications to the penultimate pre-head layers of one of the models. MobileNet I threw in at the end simply because it trained fast and seemed to improve local validation when combined in ensemble.\n\nI didn't do any kind of pre-processing and trained on 90% of the original training data.",
    "414394": "We will be posting our solution overview this weekend. This week has been quite busy for the both of us so we haven't had much time. Sorry for the wait!\n\nedit: posted here https://www.kaggle.com/c/rsna-pneumonia-detection-challenge/discussion/70421"
  },
  "source": "meta"
}