{
  "id": 214734,
  "title": "Self-supervised pretraining for this competition",
  "url": "/competitions/hpa-single-cell-image-classification/discussion/214734",
  "author_name": "Alex Lu",
  "post_date": "2021-01-27T14:15:33.826000",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I just wanted to drop a link to some relevant work that we've done on self-supervised pretraining for single cells in fluorescent microscopy images:<br>\n<a href=\"https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007348\" target=\"_blank\">https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007348</a><br>\n<a href=\"https://github.com/alexxijielu/paired_cell_inpainting\" target=\"_blank\">https://github.com/alexxijielu/paired_cell_inpainting</a></p>\n<p>My github repository was popular for the previous full-image classification challenge, and I hope it will help people this time around too - there are some scripts to download jpeg images from the HPA, and to segment these images into single cell crops. </p>\n<p>Although I don't have time to join the competition (and I'm not much of an engineer anyway), I'm very curious on if the self-supervised learning method we've developed will be a viable pre-training strategy for this competition. We've never investigated it as part of a supervised end-to-end neural network: we've only purposed it for unsupervised learning. We do know that our \"paired cell inpainting\" method produces features that are superior to other pretrained networks (e.g. ImageNet) in discriminating protein localization in single-cells without fine-tuning, and that they can be purposed to pick up on differences in single cells in the same image with clustering and distance-based statistics (check out the paper for details on this).</p>",
  "messages": [
    {
      "id": 1172713,
      "postDate": "2021-01-27T14:15:33.827Z",
      "content": "<p>I just wanted to drop a link to some relevant work that we've done on self-supervised pretraining for single cells in fluorescent microscopy images:<br>\n<a href=\"https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007348\" target=\"_blank\">https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007348</a><br>\n<a href=\"https://github.com/alexxijielu/paired_cell_inpainting\" target=\"_blank\">https://github.com/alexxijielu/paired_cell_inpainting</a></p>\n<p>My github repository was popular for the previous full-image classification challenge, and I hope it will help people this time around too - there are some scripts to download jpeg images from the HPA, and to segment these images into single cell crops. </p>\n<p>Although I don't have time to join the competition (and I'm not much of an engineer anyway), I'm very curious on if the self-supervised learning method we've developed will be a viable pre-training strategy for this competition. We've never investigated it as part of a supervised end-to-end neural network: we've only purposed it for unsupervised learning. We do know that our \"paired cell inpainting\" method produces features that are superior to other pretrained networks (e.g. ImageNet) in discriminating protein localization in single-cells without fine-tuning, and that they can be purposed to pick up on differences in single cells in the same image with clustering and distance-based statistics (check out the paper for details on this).</p>",
      "rawMarkdown": "I just wanted to drop a link to some relevant work that we've done on self-supervised pretraining for single cells in fluorescent microscopy images:\nhttps://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007348\nhttps://github.com/alexxijielu/paired_cell_inpainting\n\nMy github repository was popular for the previous full-image classification challenge, and I hope it will help people this time around too - there are some scripts to download jpeg images from the HPA, and to segment these images into single cell crops. \n\nAlthough I don't have time to join the competition (and I'm not much of an engineer anyway), I'm very curious on if the self-supervised learning method we've developed will be a viable pre-training strategy for this competition. We've never investigated it as part of a supervised end-to-end neural network: we've only purposed it for unsupervised learning. We do know that our \"paired cell inpainting\" method produces features that are superior to other pretrained networks (e.g. ImageNet) in discriminating protein localization in single-cells without fine-tuning, and that they can be purposed to pick up on differences in single cells in the same image with clustering and distance-based statistics (check out the paper for details on this).",
      "votes": 6
    },
    {
      "id": 3527604,
      "postDate": "2026-09-24T07:57:52.430Z",
      "content": "<p>Autoencoders, contrastive learning, BYOL, DINO… different ways to make models learn from images without labels.\n<a href=\"https://mlguidance.blogspot.com/2026/09/what-are-models-of-self-supervised.html\" target=\"_blank\">https://mlguidance.blogspot.com/2026/09/what-are-models-of-self-supervised.html</a></p>",
      "rawMarkdown": "Autoencoders, contrastive learning, BYOL, DINO… different ways to make models learn from images without labels.\nhttps://mlguidance.blogspot.com/2026/09/what-are-models-of-self-supervised.html"
    },
    {
      "id": 1173428,
      "postDate": "2021-01-27T21:40:12.137Z",
      "content": "<p>Upvotes. Will take a look. Thanks for being this up!</p>",
      "rawMarkdown": "Upvotes. Will take a look. Thanks for being this up!"
    }
  ],
  "comments": [
    {
      "id": 3527604,
      "author_name": "Kaustubh Verma",
      "author_url": "",
      "post_date": "2026-09-24T07:57:52.430000",
      "content": "<p>Autoencoders, contrastive learning, BYOL, DINO… different ways to make models learn from images without labels.\n<a href=\"https://mlguidance.blogspot.com/2026/09/what-are-models-of-self-supervised.html\" target=\"_blank\">https://mlguidance.blogspot.com/2026/09/what-are-models-of-self-supervised.html</a></p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1173428,
      "author_name": "Darien Schettler",
      "author_url": "",
      "post_date": "2021-01-27T21:40:12.137000",
      "content": "<p>Upvotes. Will take a look. Thanks for being this up!</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1172713": "I just wanted to drop a link to some relevant work that we've done on self-supervised pretraining for single cells in fluorescent microscopy images:\nhttps://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007348\nhttps://github.com/alexxijielu/paired_cell_inpainting\n\nMy github repository was popular for the previous full-image classification challenge, and I hope it will help people this time around too - there are some scripts to download jpeg images from the HPA, and to segment these images into single cell crops. \n\nAlthough I don't have time to join the competition (and I'm not much of an engineer anyway), I'm very curious on if the self-supervised learning method we've developed will be a viable pre-training strategy for this competition. We've never investigated it as part of a supervised end-to-end neural network: we've only purposed it for unsupervised learning. We do know that our \"paired cell inpainting\" method produces features that are superior to other pretrained networks (e.g. ImageNet) in discriminating protein localization in single-cells without fine-tuning, and that they can be purposed to pick up on differences in single cells in the same image with clustering and distance-based statistics (check out the paper for details on this).",
    "3527604": "Autoencoders, contrastive learning, BYOL, DINO… different ways to make models learn from images without labels.\nhttps://mlguidance.blogspot.com/2026/09/what-are-models-of-self-supervised.html",
    "1173428": "Upvotes. Will take a look. Thanks for being this up!"
  }
}