{
  "id": 279303,
  "title": "Previous Cell/Histopathology related competitions - ",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/279303",
  "author_name": "Dr. Amritpal Singh",
  "post_date": "2021-10-17T16:27:14.392000",
  "votes": 9,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi Kagglers, <br>\nI was looking through previous competitions, and trying to find out some domain knowledge in this region. I found 3 competitions on cell/histopathology-related areas. All 3 competitions were related to the classification of histopathology images, and not segmentation. so, I tried to look for insights that we can carry forward in segmentation tasks as well. </p>\n<h2>1. <a href=\"https://www.kaggle.com/c/hpa-single-cell-image-classification/overview\" target=\"_blank\">Human Protein Atlas - Single Cell Classification</a></h2>\n<p>Find individual human cell differences in microscope images</p>\n<ul>\n<li><p>🔬 1st place -  🦠 Fair Cell Activation Network and Swin Transformer  🦠 - by <a href=\"https://www.kaggle.com/bestfitting\" target=\"_blank\">@bestfitting</a></p>\n<ul>\n<li><p>Augmentations used</p>\n<ul>\n<li>flip, transpose, scale, rotate, crop</li>\n<li>Adding mitotic spindles with high confidence to other images to generate more positive samples of this type(lead to a boost with 0.02)</li></ul></li>\n<li><p>Test Time augmentation: default,flipud,fliplr,transpose.</p></li>\n<li><p>Another interesting findings shared by author were - </p>\n<ul>\n<li>4.1 The Fair Cell Activation Network(FCAN) can increase cell level recall which is very important to this competition.</li>\n<li>4.2 The vision transformer models have shown promising capability.</li>\n<li>4.3 <strong>Larger model not always means better result</strong> as most pre-trained models are designed for ImageNet, our models should find relationship of relative position of pixels instead of abstract semantic.</li>\n<li>4.4 I found little differences among JPEG, PNG and 8bit 16bit formats.</li></ul></li></ul></li>\n<li><p>🔬 2nd place -  <a href=\"https://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/238645\" target=\"_blank\">link</a> by  🦠  <a href=\"https://www.kaggle.com/steamedsheep\" target=\"_blank\">@steamedsheep</a>  🦠 </p>\n<ul>\n<li>16xTTA(scale, rotate, flip at random), 256 size cell-tiles</li>\n<li><strong>Authors also trained a segmentation model</strong>, and used some interesting Post-processing: - used to deal with cells on border<ul>\n<li>\"We slightly changed label_cell function from the original implementation. We found that in many cases, <strong>border cells are segmented in the wrong way</strong>: some of them are combined together with border cells that have no nuclei (or it’s outside of the image). We <strong>tweaked the watershed distance threshold</strong> in order to separate cells masks a little bit further from each other than they were before, then we ignored the masks on the border that became separated from the main cell. Furthermore, we removed the border cells with nuclei whose area was less than a half of the median area of the non-border nuclei on the image. And we also removed the cells that did not have the corresponding nuclei. </li></ul></li></ul></li>\n</ul>\n<h2>2. Recursion Cellular Image Classification</h2>\n<p>cell signal: Disentangling biological signal from experimental noise in cellular images</p>\n<ul>\n<li><p>🔬 1st place - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110543\" target=\"_blank\">link</a>  🦠 by <a href=\"https://www.kaggle.com/maciejsypetkowski\" target=\"_blank\">@maciejsypetkowski</a>  🦠 </p>\n<ul>\n<li>Progressive pseudo-labeling</li></ul></li>\n<li><p>🔬 2nd place - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110457\" target=\"_blank\">link</a>  🦠 by <a href=\"https://www.kaggle.com/pathbinder\" target=\"_blank\">@pathbinder</a> 🦠 </p></li>\n<li><p>🔬 3rd place solution - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110340\" target=\"_blank\">link</a> 🦠  🦠 </p></li>\n<li><p>🔬 4th place solution - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110337\" target=\"_blank\">link</a>  🦠 by <a href=\"https://www.kaggle.com/ren4yu\" target=\"_blank\">@ren4yu</a> 🦠 </p></li>\n<li><p>🔬 7th place - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110335\" target=\"_blank\">link</a>  🦠 by <a href=\"https://www.kaggle.com/zaharch\" target=\"_blank\">@zaharch</a> 🦠 </p>\n<ul>\n<li>Normalization of images per experiment and channel, with small randomization</li>\n<li>Test-time-augmentation (TTA)</li>\n<li>Ensemble</li>\n<li>Incremental hard pseudo-labelling (PL)</li></ul></li>\n</ul>\n<h2>3. <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/overview\" target=\"_blank\">Histopathologic Cancer Detection</a></h2>\n<p>Classification competitions, Identify metastatic tissue in histopathologic scans of lymph node sections</p>\n<p>Tricks specific to histopathology - </p>\n<ul>\n<li><p>🦠 stain normalization of images  🦠 - <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87400\" target=\"_blank\">link</a></p></li>\n<li><p>4 TTA - <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87400\" target=\"_blank\">link</a></p></li>\n<li><p>🔬 17 place solution - <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87397\" target=\"_blank\">link</a> -  🦠 by <a href=\"https://www.kaggle.com/ivanpan\" target=\"_blank\">@ivanpan</a> 🦠 </p>\n<ul>\n<li>1) I used an ensemble of 5 se_resnet50 models. Each of these models were trained exactly the same, just on different subsets of training examples. More complicated ensembles (with a bigger LB score) performed worse, so I guess they were overfitting to the LB.</li>\n<li>2) Splitting by <strong>WSI(WSI normalization) helped,</strong> judging by submissions before splitting with WSI and after.</li>\n<li>3) More intensive TTA helped. <strong>I used 16-TTA, which performed better than 4-TTA.</strong></li>\n<li>4) Obviously resizing to 196x196 helped, but I'm not sure that 196x196 is the best size.</li>\n<li>5) I also used ReduceLROnPlateau (2 epocs), but have no idea whether it helped or not</li></ul></li>\n<li><p>WSI normalisation of slides - <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87069\" target=\"_blank\">link</a></p></li>\n<li><p>Since this competition had colored images, a lot of augmentations with for hue, ColorJitter were used.</p></li>\n</ul>\n<p>Please upvote the amazing work done by these people. I hope this post was helpful to all. </p>",
  "messages": [
    {
      "id": 1547840,
      "postDate": "2021-10-17T16:27:14.393Z",
      "content": "<p>Hi Kagglers, <br>\nI was looking through previous competitions, and trying to find out some domain knowledge in this region. I found 3 competitions on cell/histopathology-related areas. All 3 competitions were related to the classification of histopathology images, and not segmentation. so, I tried to look for insights that we can carry forward in segmentation tasks as well. </p>\n<h2>1. <a href=\"https://www.kaggle.com/c/hpa-single-cell-image-classification/overview\" target=\"_blank\">Human Protein Atlas - Single Cell Classification</a></h2>\n<p>Find individual human cell differences in microscope images</p>\n<ul>\n<li><p>🔬 1st place -  🦠 Fair Cell Activation Network and Swin Transformer  🦠 - by <a href=\"https://www.kaggle.com/bestfitting\" target=\"_blank\">@bestfitting</a></p>\n<ul>\n<li><p>Augmentations used</p>\n<ul>\n<li>flip, transpose, scale, rotate, crop</li>\n<li>Adding mitotic spindles with high confidence to other images to generate more positive samples of this type(lead to a boost with 0.02)</li></ul></li>\n<li><p>Test Time augmentation: default,flipud,fliplr,transpose.</p></li>\n<li><p>Another interesting findings shared by author were - </p>\n<ul>\n<li>4.1 The Fair Cell Activation Network(FCAN) can increase cell level recall which is very important to this competition.</li>\n<li>4.2 The vision transformer models have shown promising capability.</li>\n<li>4.3 <strong>Larger model not always means better result</strong> as most pre-trained models are designed for ImageNet, our models should find relationship of relative position of pixels instead of abstract semantic.</li>\n<li>4.4 I found little differences among JPEG, PNG and 8bit 16bit formats.</li></ul></li></ul></li>\n<li><p>🔬 2nd place -  <a href=\"https://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/238645\" target=\"_blank\">link</a> by  🦠  <a href=\"https://www.kaggle.com/steamedsheep\" target=\"_blank\">@steamedsheep</a>  🦠 </p>\n<ul>\n<li>16xTTA(scale, rotate, flip at random), 256 size cell-tiles</li>\n<li><strong>Authors also trained a segmentation model</strong>, and used some interesting Post-processing: - used to deal with cells on border<ul>\n<li>\"We slightly changed label_cell function from the original implementation. We found that in many cases, <strong>border cells are segmented in the wrong way</strong>: some of them are combined together with border cells that have no nuclei (or it’s outside of the image). We <strong>tweaked the watershed distance threshold</strong> in order to separate cells masks a little bit further from each other than they were before, then we ignored the masks on the border that became separated from the main cell. Furthermore, we removed the border cells with nuclei whose area was less than a half of the median area of the non-border nuclei on the image. And we also removed the cells that did not have the corresponding nuclei. </li></ul></li></ul></li>\n</ul>\n<h2>2. Recursion Cellular Image Classification</h2>\n<p>cell signal: Disentangling biological signal from experimental noise in cellular images</p>\n<ul>\n<li><p>🔬 1st place - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110543\" target=\"_blank\">link</a>  🦠 by <a href=\"https://www.kaggle.com/maciejsypetkowski\" target=\"_blank\">@maciejsypetkowski</a>  🦠 </p>\n<ul>\n<li>Progressive pseudo-labeling</li></ul></li>\n<li><p>🔬 2nd place - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110457\" target=\"_blank\">link</a>  🦠 by <a href=\"https://www.kaggle.com/pathbinder\" target=\"_blank\">@pathbinder</a> 🦠 </p></li>\n<li><p>🔬 3rd place solution - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110340\" target=\"_blank\">link</a> 🦠  🦠 </p></li>\n<li><p>🔬 4th place solution - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110337\" target=\"_blank\">link</a>  🦠 by <a href=\"https://www.kaggle.com/ren4yu\" target=\"_blank\">@ren4yu</a> 🦠 </p></li>\n<li><p>🔬 7th place - <a href=\"https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110335\" target=\"_blank\">link</a>  🦠 by <a href=\"https://www.kaggle.com/zaharch\" target=\"_blank\">@zaharch</a> 🦠 </p>\n<ul>\n<li>Normalization of images per experiment and channel, with small randomization</li>\n<li>Test-time-augmentation (TTA)</li>\n<li>Ensemble</li>\n<li>Incremental hard pseudo-labelling (PL)</li></ul></li>\n</ul>\n<h2>3. <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/overview\" target=\"_blank\">Histopathologic Cancer Detection</a></h2>\n<p>Classification competitions, Identify metastatic tissue in histopathologic scans of lymph node sections</p>\n<p>Tricks specific to histopathology - </p>\n<ul>\n<li><p>🦠 stain normalization of images  🦠 - <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87400\" target=\"_blank\">link</a></p></li>\n<li><p>4 TTA - <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87400\" target=\"_blank\">link</a></p></li>\n<li><p>🔬 17 place solution - <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87397\" target=\"_blank\">link</a> -  🦠 by <a href=\"https://www.kaggle.com/ivanpan\" target=\"_blank\">@ivanpan</a> 🦠 </p>\n<ul>\n<li>1) I used an ensemble of 5 se_resnet50 models. Each of these models were trained exactly the same, just on different subsets of training examples. More complicated ensembles (with a bigger LB score) performed worse, so I guess they were overfitting to the LB.</li>\n<li>2) Splitting by <strong>WSI(WSI normalization) helped,</strong> judging by submissions before splitting with WSI and after.</li>\n<li>3) More intensive TTA helped. <strong>I used 16-TTA, which performed better than 4-TTA.</strong></li>\n<li>4) Obviously resizing to 196x196 helped, but I'm not sure that 196x196 is the best size.</li>\n<li>5) I also used ReduceLROnPlateau (2 epocs), but have no idea whether it helped or not</li></ul></li>\n<li><p>WSI normalisation of slides - <a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87069\" target=\"_blank\">link</a></p></li>\n<li><p>Since this competition had colored images, a lot of augmentations with for hue, ColorJitter were used.</p></li>\n</ul>\n<p>Please upvote the amazing work done by these people. I hope this post was helpful to all. </p>",
      "rawMarkdown": "Hi Kagglers, \nI was looking through previous competitions, and trying to find out some domain knowledge in this region. I found 3 competitions on cell/histopathology-related areas. All 3 competitions were related to the classification of histopathology images, and not segmentation. so, I tried to look for insights that we can carry forward in segmentation tasks as well. \n\n##1. [Human Protein Atlas - Single Cell Classification](https://www.kaggle.com/c/hpa-single-cell-image-classification/overview)\nFind individual human cell differences in microscope images\n- 🔬 1st place -  🦠 Fair Cell Activation Network and Swin Transformer  🦠 - by @bestfitting\n\n   - Augmentations used\n      - flip, transpose, scale, rotate, crop\n      - Adding mitotic spindles with high confidence to other images to generate more positive samples of this type(lead to a boost with 0.02)\n   - Test Time augmentation: default,flipud,fliplr,transpose.\n\n   - Another interesting findings shared by author were - \n\n        - 4.1 The Fair Cell Activation Network(FCAN) can increase cell level recall which is very important to this competition.\n        - 4.2 The vision transformer models have shown promising capability.\n        - 4.3 **Larger model not always means better result** as most pre-trained models are designed for ImageNet, our models should find relationship of relative position of pixels instead of abstract semantic.\n        - 4.4 I found little differences among JPEG, PNG and 8bit 16bit formats.\n\n-  🔬 2nd place -  [link](https://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/238645) by  🦠  @steamedsheep  🦠 \n  - 16xTTA(scale, rotate, flip at random), 256 size cell-tiles\n  - **Authors also trained a segmentation model**, and used some interesting Post-processing: - used to deal with cells on border\n        - \"We slightly changed label_cell function from the original implementation. We found that in many cases, **border cells are segmented in the wrong way**: some of them are combined together with border cells that have no nuclei (or it’s outside of the image). We **tweaked the watershed distance threshold** in order to separate cells masks a little bit further from each other than they were before, then we ignored the masks on the border that became separated from the main cell. Furthermore, we removed the border cells with nuclei whose area was less than a half of the median area of the non-border nuclei on the image. And we also removed the cells that did not have the corresponding nuclei. \n\n## 2. Recursion Cellular Image Classification\ncell signal: Disentangling biological signal from experimental noise in cellular images\n\n-  🔬 1st place - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110543)  🦠 by @maciejsypetkowski  🦠 \n  - Progressive pseudo-labeling\n\n-  🔬 2nd place - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110457)  🦠 by @pathbinder 🦠 \n\n-  🔬 3rd place solution - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110340) 🦠  🦠 \n\n-  🔬 4th place solution - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110337)  🦠 by @ren4yu 🦠 \n\n-  🔬 7th place - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110335)  🦠 by @zaharch 🦠 \n  - Normalization of images per experiment and channel, with small randomization\n  -  Test-time-augmentation (TTA)\n  - Ensemble\n  - Incremental hard pseudo-labelling (PL)\n\n\n## 3. [Histopathologic Cancer Detection](https://www.kaggle.com/c/histopathologic-cancer-detection/overview)\nClassification competitions, Identify metastatic tissue in histopathologic scans of lymph node sections\n\nTricks specific to histopathology - \n-  🦠 stain normalization of images  🦠 - [link](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87400)\n- 4 TTA - [link](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87400)\n\n-  🔬 17 place solution - [link](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87397) -  🦠 by @ivanpan 🦠 \n\n   - 1) I used an ensemble of 5 se_resnet50 models. Each of these models were trained exactly the same, just on different subsets of training examples. More complicated ensembles (with a bigger LB score) performed worse, so I guess they were overfitting to the LB.\n   - 2) Splitting by **WSI(WSI normalization) helped,** judging by submissions before splitting with WSI and after.\n   - 3) More intensive TTA helped. **I used 16-TTA, which performed better than 4-TTA.**\n   - 4) Obviously resizing to 196x196 helped, but I'm not sure that 196x196 is the best size.\n   - 5) I also used ReduceLROnPlateau (2 epocs), but have no idea whether it helped or not\n\n \n- WSI normalisation of slides - [link](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87069)\n\n- Since this competition had colored images, a lot of augmentations with for hue, ColorJitter were used.\n\nPlease upvote the amazing work done by these people. I hope this post was helpful to all. \n",
      "votes": 9
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1547840": "Hi Kagglers, \nI was looking through previous competitions, and trying to find out some domain knowledge in this region. I found 3 competitions on cell/histopathology-related areas. All 3 competitions were related to the classification of histopathology images, and not segmentation. so, I tried to look for insights that we can carry forward in segmentation tasks as well. \n\n##1. [Human Protein Atlas - Single Cell Classification](https://www.kaggle.com/c/hpa-single-cell-image-classification/overview)\nFind individual human cell differences in microscope images\n- 🔬 1st place -  🦠 Fair Cell Activation Network and Swin Transformer  🦠 - by @bestfitting\n\n   - Augmentations used\n      - flip, transpose, scale, rotate, crop\n      - Adding mitotic spindles with high confidence to other images to generate more positive samples of this type(lead to a boost with 0.02)\n   - Test Time augmentation: default,flipud,fliplr,transpose.\n\n   - Another interesting findings shared by author were - \n\n        - 4.1 The Fair Cell Activation Network(FCAN) can increase cell level recall which is very important to this competition.\n        - 4.2 The vision transformer models have shown promising capability.\n        - 4.3 **Larger model not always means better result** as most pre-trained models are designed for ImageNet, our models should find relationship of relative position of pixels instead of abstract semantic.\n        - 4.4 I found little differences among JPEG, PNG and 8bit 16bit formats.\n\n-  🔬 2nd place -  [link](https://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/238645) by  🦠  @steamedsheep  🦠 \n  - 16xTTA(scale, rotate, flip at random), 256 size cell-tiles\n  - **Authors also trained a segmentation model**, and used some interesting Post-processing: - used to deal with cells on border\n        - \"We slightly changed label_cell function from the original implementation. We found that in many cases, **border cells are segmented in the wrong way**: some of them are combined together with border cells that have no nuclei (or it’s outside of the image). We **tweaked the watershed distance threshold** in order to separate cells masks a little bit further from each other than they were before, then we ignored the masks on the border that became separated from the main cell. Furthermore, we removed the border cells with nuclei whose area was less than a half of the median area of the non-border nuclei on the image. And we also removed the cells that did not have the corresponding nuclei. \n\n## 2. Recursion Cellular Image Classification\ncell signal: Disentangling biological signal from experimental noise in cellular images\n\n-  🔬 1st place - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110543)  🦠 by @maciejsypetkowski  🦠 \n  - Progressive pseudo-labeling\n\n-  🔬 2nd place - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110457)  🦠 by @pathbinder 🦠 \n\n-  🔬 3rd place solution - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110340) 🦠  🦠 \n\n-  🔬 4th place solution - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110337)  🦠 by @ren4yu 🦠 \n\n-  🔬 7th place - [link](https://www.kaggle.com/c/recursion-cellular-image-classification/discussion/110335)  🦠 by @zaharch 🦠 \n  - Normalization of images per experiment and channel, with small randomization\n  -  Test-time-augmentation (TTA)\n  - Ensemble\n  - Incremental hard pseudo-labelling (PL)\n\n\n## 3. [Histopathologic Cancer Detection](https://www.kaggle.com/c/histopathologic-cancer-detection/overview)\nClassification competitions, Identify metastatic tissue in histopathologic scans of lymph node sections\n\nTricks specific to histopathology - \n-  🦠 stain normalization of images  🦠 - [link](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87400)\n- 4 TTA - [link](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87400)\n\n-  🔬 17 place solution - [link](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87397) -  🦠 by @ivanpan 🦠 \n\n   - 1) I used an ensemble of 5 se_resnet50 models. Each of these models were trained exactly the same, just on different subsets of training examples. More complicated ensembles (with a bigger LB score) performed worse, so I guess they were overfitting to the LB.\n   - 2) Splitting by **WSI(WSI normalization) helped,** judging by submissions before splitting with WSI and after.\n   - 3) More intensive TTA helped. **I used 16-TTA, which performed better than 4-TTA.**\n   - 4) Obviously resizing to 196x196 helped, but I'm not sure that 196x196 is the best size.\n   - 5) I also used ReduceLROnPlateau (2 epocs), but have no idea whether it helped or not\n\n \n- WSI normalisation of slides - [link](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/87069)\n\n- Since this competition had colored images, a lot of augmentations with for hue, ColorJitter were used.\n\nPlease upvote the amazing work done by these people. I hope this post was helpful to all. \n"
  }
}