{ "id": 294176, "title": "Quick review: Cellpose", "url": "/competitions/sartorius-cell-instance-segmentation/discussion/294176", "author_name": "", "post_date": "2021-12-08T22:54:41.560305500Z", "votes": 19, "comment_count": 1, "views": 0, "content": "

I read through Cellpose paper.
\nI listed up the terminology and technology for better understanding the concept.

\n\n
    \n
  1. it uses feature concatenation on the upsampling path,
    2. \n
  2. building block is replaced by residual block and
    3. \n
  3. global average pooling is used to get style of image.
    4. \n
\n", "messages": [ { "id": "1612416", "postDate": "12/08/2021 22:54:41", "content": "

I read through Cellpose paper.
\nI listed up the terminology and technology for better understanding the concept.

\n\n
    \n
  1. it uses feature concatenation on the upsampling path,
    2. \n
  2. building block is replaced by residual block and
    3. \n
  3. global average pooling is used to get style of image.
    4. \n
\n", "rawMarkdown": "I read through [Cellpose paper](https://www.biorxiv.org/content/10.1101/2020.02.02.931238v1.full.pdf).\nI listed up the terminology and technology for better understanding the concept.\n\n\n- [**Stardist**](https://arxiv.org/pdf/1908.03636.pdf)\nDeep-learning tool for phase-contrast cell images. It is introduced as an example. For comparison, Stardist and [**Mask-RCNN**](https://arxiv.org/pdf/1703.06870.pdf) are chosen as previous sota methods.\n\n- [**nucleAIzer**](https://www.sciencedirect.com/science/article/pii/S2405471220301174)\nDeep-learning tool for phase-contrast cell images. It is introduced by an example.\n\n- **Energy function**\nCore concept. It is defined in the paper. Let's say the heat-source is virtually set at the center of the cell. By heat diffusion, inside cells various types of heat distribution are formed according to shapes of cells, which can be determined by thermal equilibrium.\nThen we can calculate spatial gradient of heat **flow** (vector representation). This flow is given to the input of the neural network.\n\n- [**U-net**](https://arxiv.org/pdf/1505.04597.pdf)\nThe architecture is based on U-net. But the different points between U-net and Cellpose architecure are for Cellpose\n1. it uses feature concatenation on the upsampling path,\n2. building block is replaced by [**residual block**](https://paperswithcode.com/method/residual-block) and \n3. [**global average pooling**](https://paperswithcode.com/method/global-average-pooling) is used to get *style* of image.\n\n- **Test time enhancements**\nTest time resizing, ROI quality estimation, model ensembling, image tiling and image augmentation are done *to further increase the predictive power of the model*.\n\n- [**fruits, rocks and jellyfish**](https://en.wikipedia.org/wiki/Jellyfish)\nA small set of nonmicroscopy images are included to training dataset to *allow the network to generalize more widely and more robustly* .\n\n- [**t-SNE**](https://towardsdatascience.com/t-sne-python-example-1ded9953f26)\nFor visualization of the structure of this dataset.\n\n- **[PyQt](https://doc.qt.io/qtforpython/) and [pyqtgraph](https://pyqtgraph.readthedocs.io/en/latest/)**\nTo develop GUI.", "votes": null }, { "id": "1614314", "postDate": "12/10/2021 23:05:12", "content": "

Thank you very much!
\nThis discussion brought me my first gold medal😄
\nI will keep it up so as to publish such topic.

", "rawMarkdown": "Thank you very much!\nThis discussion brought me my first gold medal😄\nI will keep it up so as to publish such topic.", "votes": null } ], "comments": [ { "id": 1614314, "author_name": "osamurai", "author_url": "", "post_date": "12/10/2021 23:05:12", "content": "

Thank you very much!
\nThis discussion brought me my first gold medal😄
\nI will keep it up so as to publish such topic.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1612416": "I read through [Cellpose paper](https://www.biorxiv.org/content/10.1101/2020.02.02.931238v1.full.pdf).\nI listed up the terminology and technology for better understanding the concept.\n\n\n- [**Stardist**](https://arxiv.org/pdf/1908.03636.pdf)\nDeep-learning tool for phase-contrast cell images. It is introduced as an example. For comparison, Stardist and [**Mask-RCNN**](https://arxiv.org/pdf/1703.06870.pdf) are chosen as previous sota methods.\n\n- [**nucleAIzer**](https://www.sciencedirect.com/science/article/pii/S2405471220301174)\nDeep-learning tool for phase-contrast cell images. It is introduced by an example.\n\n- **Energy function**\nCore concept. It is defined in the paper. Let's say the heat-source is virtually set at the center of the cell. By heat diffusion, inside cells various types of heat distribution are formed according to shapes of cells, which can be determined by thermal equilibrium.\nThen we can calculate spatial gradient of heat **flow** (vector representation). This flow is given to the input of the neural network.\n\n- [**U-net**](https://arxiv.org/pdf/1505.04597.pdf)\nThe architecture is based on U-net. But the different points between U-net and Cellpose architecure are for Cellpose\n1. it uses feature concatenation on the upsampling path,\n2. building block is replaced by [**residual block**](https://paperswithcode.com/method/residual-block) and \n3. [**global average pooling**](https://paperswithcode.com/method/global-average-pooling) is used to get *style* of image.\n\n- **Test time enhancements**\nTest time resizing, ROI quality estimation, model ensembling, image tiling and image augmentation are done *to further increase the predictive power of the model*.\n\n- [**fruits, rocks and jellyfish**](https://en.wikipedia.org/wiki/Jellyfish)\nA small set of nonmicroscopy images are included to training dataset to *allow the network to generalize more widely and more robustly* .\n\n- [**t-SNE**](https://towardsdatascience.com/t-sne-python-example-1ded9953f26)\nFor visualization of the structure of this dataset.\n\n- **[PyQt](https://doc.qt.io/qtforpython/) and [pyqtgraph](https://pyqtgraph.readthedocs.io/en/latest/)**\nTo develop GUI.", "1614314": "Thank you very much!\nThis discussion brought me my first gold medal😄\nI will keep it up so as to publish such topic." }, "source": "meta" }