{ "id": 351356, "title": "More background information on lung FTUs", "url": "/competitions/hubmap-organ-segmentation/discussion/351356", "author_name": "", "post_date": "2022-09-09T16:30:25.025157800Z", "votes": 21, "comment_count": 8, "views": 0, "content": "

Hi all,
\nWe wanted to share some more information about lung FTUs which might clear some confusions about the masks. The data includes masks of both atelectatic (collapsed) and inflated alveoli (un-collapsed). The alveolar appearance on the image slides depends on how the tissue samples were prepared. For the inflated alveoli, which have a 3D ‘cup’ shaped structure, how the tissue is sectioned can cause variability as well. If the alveoli were sectioned in a horizontal manner, their shape will appear more like a complete circle. Whereas if the alveoli were sectioned vertically, they may appear more as a U-shape. Here are some examples to help clarify:

\n

This is an example of data with atelectatic alveoli:
\n\"\"

\n

This is an example of data with inflated alveoli likely cut in a horizontal manner where the mask also captures the alveolar space:
\n\"\"

\n

This is an example of data with inflated alveoli cut in a vertical manner:
\n\"\"

\n

We hope this clears some confusion regarding the variability in lung FTUs.

\n

Best,
\nYash

", "messages": [ { "id": "1932464", "postDate": "09/09/2022 16:30:25", "content": "

Hi all,
\nWe wanted to share some more information about lung FTUs which might clear some confusions about the masks. The data includes masks of both atelectatic (collapsed) and inflated alveoli (un-collapsed). The alveolar appearance on the image slides depends on how the tissue samples were prepared. For the inflated alveoli, which have a 3D ‘cup’ shaped structure, how the tissue is sectioned can cause variability as well. If the alveoli were sectioned in a horizontal manner, their shape will appear more like a complete circle. Whereas if the alveoli were sectioned vertically, they may appear more as a U-shape. Here are some examples to help clarify:

\n

This is an example of data with atelectatic alveoli:
\n\"\"

\n

This is an example of data with inflated alveoli likely cut in a horizontal manner where the mask also captures the alveolar space:
\n\"\"

\n

This is an example of data with inflated alveoli cut in a vertical manner:
\n\"\"

\n

We hope this clears some confusion regarding the variability in lung FTUs.

\n

Best,
\nYash

", "rawMarkdown": "Hi all,\nWe wanted to share some more information about lung FTUs which might clear some confusions about the masks. The data includes masks of both atelectatic (collapsed) and inflated alveoli (un-collapsed). The alveolar appearance on the image slides depends on how the tissue samples were prepared. For the inflated alveoli, which have a 3D ‘cup’ shaped structure, how the tissue is sectioned can cause variability as well. If the alveoli were sectioned in a horizontal manner, their shape will appear more like a complete circle. Whereas if the alveoli were sectioned vertically, they may appear more as a U-shape. Here are some examples to help clarify: \n\nThis is an example of data with atelectatic alveoli:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1536542%2F28715b71f79b8e4c2ab02f4b48183019%2F1.png?generation=1662740626971517&alt=media)\n\nThis is an example of data with inflated alveoli likely cut in a horizontal manner where the mask also captures the alveolar space:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1536542%2Fea9d69eb5773829a8f9d75bf2df6ab14%2F2.png?generation=1662740636207401&alt=media)\n\nThis is an example of data with inflated alveoli cut in a vertical manner:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1536542%2Fc39c4412abbb931b57f9ff2e6a2764da%2F3.png?generation=1662740644142930&alt=media)\n\nWe hope this clears some confusion regarding the variability in lung FTUs.\n\nBest,\nYash", "votes": null }, { "id": "1932528", "postDate": "09/09/2022 17:36:37", "content": "

Thanks. this means that we have to look at the whole image to decide how the lung FTU would look like.
\nIt probably explain why whole image is better than sliding window.

\n

It solve my puzzle: why efficient-net is especially good at lung in my experiments (it uses channel squeeze and excite attention, which is based on GAP over the whole image. the attention activation heatmap can confirm this)

", "rawMarkdown": "Thanks. this means that we have to look at the whole image to decide how the lung FTU would look like.\nIt probably explain why whole image is better than sliding window.\n\nIt solve my puzzle: why efficient-net is especially good at lung in my experiments (it uses channel squeeze and excite attention, which is based on GAP over the whole image. the attention activation heatmap can confirm this)", "votes": null }, { "id": "1932532", "postDate": "09/09/2022 17:46:44", "content": "

Thanks for the explanation. I have a question. How does the annotator decide which parts are the alveolar space in the third example?

", "rawMarkdown": "Thanks for the explanation. I have a question. How does the annotator decide which parts are the alveolar space in the third example?", "votes": null }, { "id": "1932535", "postDate": "09/09/2022 17:48:00", "content": "

Yeah, I also observed that my efficientnet models are better than other transformer based models on lung samples. I thought transformers are supposed to be better at capturing long range dependencies?

", "rawMarkdown": "Yeah, I also observed that my efficientnet models are better than other transformer based models on lung samples. I thought transformers are supposed to be better at capturing long range dependencies?", "votes": null }, { "id": "1932573", "postDate": "09/09/2022 18:45:30", "content": "

Thank you, interesting

", "rawMarkdown": "Thank you, interesting", "votes": null }, { "id": "1932786", "postDate": "09/10/2022 00:03:28", "content": "

u need a global context cls token to capture whole image

", "rawMarkdown": "u need a global context cls token to capture whole image", "votes": null }, { "id": "1933124", "postDate": "09/10/2022 09:26:04", "content": "

It probably explain why whole image is better than sliding window. Yeah, there will be amount of noise when training with tiles, my lung is only about 0.08 in hubmap, but almost 0.59 in total.
\nPerhaps I'll need a teammate doing nicely in lung, haha😂

", "rawMarkdown": "`It probably explain why whole image is better than sliding window.` Yeah, there will be amount of noise when training with tiles, my lung is only about 0.08 in hubmap, but almost 0.59 in total.\nPerhaps I'll need a teammate doing nicely in lung, haha😂", "votes": null }, { "id": "1933685", "postDate": "09/10/2022 17:12:19", "content": "

for LB=0.83, hupmap-lung is 0.10

", "rawMarkdown": "for LB=0.83, hupmap-lung is 0.10", "votes": null }, { "id": "1934465", "postDate": "09/11/2022 11:54:46", "content": "

Hi all. @yashvrdnjain is it really the best way we can annotate lungs FTU's? It looks like if several medical experts will annotate this competition lung data the way it was done, iou of their annotations will be close to zero (at least at inflated ones).

", "rawMarkdown": "Hi all. @yashvrdnjain is it really the best way we can annotate lungs FTU's? It looks like if several medical experts will annotate this competition lung data the way it was done, iou of their annotations will be close to zero (at least at inflated ones).", "votes": null } ], "comments": [ { "id": 1932528, "author_name": "hengck23", "author_url": "", "post_date": "09/09/2022 17:36:37", "content": "

Thanks. this means that we have to look at the whole image to decide how the lung FTU would look like.
\nIt probably explain why whole image is better than sliding window.

\n

It solve my puzzle: why efficient-net is especially good at lung in my experiments (it uses channel squeeze and excite attention, which is based on GAP over the whole image. the attention activation heatmap can confirm this)

", "votes": null, "replies": [ { "id": 1932535, "author_name": "gunesevitan", "author_url": "", "post_date": "09/09/2022 17:48:00", "content": "

Yeah, I also observed that my efficientnet models are better than other transformer based models on lung samples. I thought transformers are supposed to be better at capturing long range dependencies?

", "votes": null, "replies": [] }, { "id": 1932786, "author_name": "hengck23", "author_url": "", "post_date": "09/10/2022 00:03:28", "content": "

u need a global context cls token to capture whole image

", "votes": null, "replies": [] }, { "id": 1933124, "author_name": "w3579628328", "author_url": "", "post_date": "09/10/2022 09:26:04", "content": "

It probably explain why whole image is better than sliding window. Yeah, there will be amount of noise when training with tiles, my lung is only about 0.08 in hubmap, but almost 0.59 in total.
\nPerhaps I'll need a teammate doing nicely in lung, haha😂

", "votes": null, "replies": [] }, { "id": 1933685, "author_name": "hengck23", "author_url": "", "post_date": "09/10/2022 17:12:19", "content": "

for LB=0.83, hupmap-lung is 0.10

", "votes": null, "replies": [] } ] }, { "id": 1932532, "author_name": "gunesevitan", "author_url": "", "post_date": "09/09/2022 17:46:44", "content": "

Thanks for the explanation. I have a question. How does the annotator decide which parts are the alveolar space in the third example?

", "votes": null, "replies": [] }, { "id": 1932573, "author_name": "saberghaderi", "author_url": "", "post_date": "09/09/2022 18:45:30", "content": "

Thank you, interesting

", "votes": null, "replies": [] }, { "id": 1934465, "author_name": "bakeryproducts", "author_url": "", "post_date": "09/11/2022 11:54:46", "content": "

Hi all. @yashvrdnjain is it really the best way we can annotate lungs FTU's? It looks like if several medical experts will annotate this competition lung data the way it was done, iou of their annotations will be close to zero (at least at inflated ones).

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1932464": "Hi all,\nWe wanted to share some more information about lung FTUs which might clear some confusions about the masks. The data includes masks of both atelectatic (collapsed) and inflated alveoli (un-collapsed). The alveolar appearance on the image slides depends on how the tissue samples were prepared. For the inflated alveoli, which have a 3D ‘cup’ shaped structure, how the tissue is sectioned can cause variability as well. If the alveoli were sectioned in a horizontal manner, their shape will appear more like a complete circle. Whereas if the alveoli were sectioned vertically, they may appear more as a U-shape. Here are some examples to help clarify: \n\nThis is an example of data with atelectatic alveoli:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1536542%2F28715b71f79b8e4c2ab02f4b48183019%2F1.png?generation=1662740626971517&alt=media)\n\nThis is an example of data with inflated alveoli likely cut in a horizontal manner where the mask also captures the alveolar space:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1536542%2Fea9d69eb5773829a8f9d75bf2df6ab14%2F2.png?generation=1662740636207401&alt=media)\n\nThis is an example of data with inflated alveoli cut in a vertical manner:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1536542%2Fc39c4412abbb931b57f9ff2e6a2764da%2F3.png?generation=1662740644142930&alt=media)\n\nWe hope this clears some confusion regarding the variability in lung FTUs.\n\nBest,\nYash", "1932528": "Thanks. this means that we have to look at the whole image to decide how the lung FTU would look like.\nIt probably explain why whole image is better than sliding window.\n\nIt solve my puzzle: why efficient-net is especially good at lung in my experiments (it uses channel squeeze and excite attention, which is based on GAP over the whole image. the attention activation heatmap can confirm this)", "1932532": "Thanks for the explanation. I have a question. How does the annotator decide which parts are the alveolar space in the third example?", "1932535": "Yeah, I also observed that my efficientnet models are better than other transformer based models on lung samples. I thought transformers are supposed to be better at capturing long range dependencies?", "1932573": "Thank you, interesting", "1932786": "u need a global context cls token to capture whole image", "1933124": "`It probably explain why whole image is better than sliding window.` Yeah, there will be amount of noise when training with tiles, my lung is only about 0.08 in hubmap, but almost 0.59 in total.\nPerhaps I'll need a teammate doing nicely in lung, haha😂", "1933685": "for LB=0.83, hupmap-lung is 0.10", "1934465": "Hi all. @yashvrdnjain is it really the best way we can annotate lungs FTU's? It looks like if several medical experts will annotate this competition lung data the way it was done, iou of their annotations will be close to zero (at least at inflated ones)." }, "source": "meta" }