\nPsammoma bodies, if present the diagnosis has a high probability of being LGSC:
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15149404%2F5ded130b69ac5efb2120c1005f0c261f%2Fpsammoma.jpg?generation=1702393344261705&alt=media\")
\nThey can range in size and number and cause many scratches of the images.
Hi everyone, I would like to share how a pathologist generally handles this task.
\nThrough the laboratory, a tissue sample/specimen is received, processed, and stained with hematoxylin and eosin (H&E) before being given to the pathologist, who uses their expertise and experience to arrive at a diagnosis of the specimen.
\nThis is essentially classification task because we already know that the specimen contains a tumor, we just have to decide which \"class\" of tumor it is. Each of the classes has a set of morphological criteria that includes presence or absence of a certain feature, which can be found in the WHO classification of tumors and many other textbooks.
\nFrom the given data, there are large whole slide images, and small tissue microarray (TMA) images. The TMA images are essentially a small patch of the labelled tumor class, while the whole slide images contain many patches of the tumor class, along with the other morphologies that may or may not be relevant to deciding the tumor class. For example, high grade serous carcinoma is notoriously \"ugly\" looking because it has \"high-grade\" nuclei and can have extensive areas of necrosis (dead cells). The model should factor in areas of tumor and areas of necrosis to arrive at the correct classification, kind of like classification within a classification.
\nAdded to the work burden, is that processing of specimens is prone to variations in the staining, cutting, optics, etc., that the classification model must be able to handle well.
Hi everyone, I would like to share how a pathologist generally handles this task.
\nThrough the laboratory, a tissue sample/specimen is received, processed, and stained with hematoxylin and eosin (H&E) before being given to the pathologist, who uses their expertise and experience to arrive at a diagnosis of the specimen.
\nThis is essentially classification task because we already know that the specimen contains a tumor, we just have to decide which \"class\" of tumor it is. Each of the classes has a set of morphological criteria that includes presence or absence of a certain feature, which can be found in the WHO classification of tumors and many other textbooks.
\nFrom the given data, there are large whole slide images, and small tissue microarray (TMA) images. The TMA images are essentially a small patch of the labelled tumor class, while the whole slide images contain many patches of the tumor class, along with the other morphologies that may or may not be relevant to deciding the tumor class. For example, high grade serous carcinoma is notoriously \"ugly\" looking because it has \"high-grade\" nuclei and can have extensive areas of necrosis (dead cells). The model should factor in areas of tumor and areas of necrosis to arrive at the correct classification, kind of like classification within a classification.
\nAdded to the work burden, is that processing of specimens is prone to variations in the staining, cutting, optics, etc., that the classification model must be able to handle well.
I've been reviewing the training images and have some additional tips that might be useful:
\nThere are many images that only contain a tiny amount of the specified tumor, and probably could be excluded or used as null. These are: 281, 3222, 5264, 9154, 12244, 26124, 31793, 32192, 33839, 41099, 52308, 54506, 63836.
\nImage number 1289 has way too many artifacts and may be excluded.
\nImage number 15583 should be labelled as MC not LGSC.
\nImage number 32035 can be excluded since it's distorted, and there are other examples of HGSC with the same morphology.
\nImage number 34822 is quite a unique example of EC.
Other things I noticed are:
\nPsammoma bodies, if present the diagnosis has a high probability of being LGSC:
\n
\nThey can range in size and number and cause many scratches of the images.
Giant cells, if present, the diagnosis always seems to be HGSC:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15149404%2Fdd2bc556f7a4a0fddd29dddb31c7715f%2Fgiant%20cells.jpg?generation=1702393565049743&alt=media\")
Metaplasia, if present, favors EC instead of HGSC. The most common is squamous metaplasia which can look like this:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15149404%2Fccb7f45cb4d78701e7f8686ab80ac947%2Fsquamet.jpg?generation=1702394378159652&alt=media\")
Thank you for your work. Incorrect data has a significant impact on models, especially when there is limited training data. I will modify train.csv based on your suggestions and provide you with feedback on the result soon. I hope you can continue to provide us with more information.🥰
", "rawMarkdown": "Thank you for your work. Incorrect data has a significant impact on models, especially when there is limited training data. I will modify train.csv based on your suggestions and provide you with feedback on the result soon. I hope you can continue to provide us with more information.🥰", "votes": 1, "replies": [ { "id": 2560799, "postDate": "2023-12-14T02:38:02.860Z", "content": "Now I have some good news. Based on your suggestion, I removed number 1289 32035 and changed the label of 15583, which brought me 0.03 improvement. As I said, incorrect data can have a catastrophic impact on the model. I hope to receive more help from you.😄
", "rawMarkdown": "Now I have some good news. Based on your suggestion, I removed number 1289 32035 and changed the label of 15583, which brought me 0.03 improvement. As I said, incorrect data can have a catastrophic impact on the model. I hope to receive more help from you.😄", "votes": 9 } ] }, { "id": 2559098, "postDate": "2023-12-12T15:56:57.950Z", "content": "WOW, thank you!
", "rawMarkdown": "WOW, thank you!" }, { "id": 2571185, "postDate": "2023-12-23T01:47:25.260Z", "content": "I did some experiments according to your advice. I removed number 1289 32035 and changed the label of 15583 and it didnt improve my LB…then I removed more number like 281, 3222, 5264, 9154, e.t.c, lb score just increased by 0.01…hope this information would help
", "rawMarkdown": "I did some experiments according to your advice. I removed number 1289 32035 and changed the label of 15583 and it didnt improve my LB...then I removed more number like 281, 3222, 5264, 9154, e.t.c, lb score just increased by 0.01...hope this information would help \n", "votes": 8 } ] }, { "id": 2507584, "postDate": "2023-11-01T05:41:01.747Z", "content": "I can confirm that the tiles and detail view seem to be really important for this task,
\n as explained in baseline with Lightningâš¡TIMM scores 0.4+ on LB become TOP 5%
Hi there ! , are cancerous area is uniformly spread all over images or just concentrated on just only some part ?
", "rawMarkdown": "Hi there ! , are cancerous area is uniformly spread all over images or just concentrated on just only some part ?", "votes": 4, "replies": [ { "id": 2484099, "postDate": "2023-10-16T08:16:41.303Z", "content": "It varies widely. Some of the images are purely tumor tissue, but in some areas the tumor is concentrated (solid), other areas are cystic (large spaces filled with fluid), other areas are infiltrative (tumor cells in between normal stroma). Some of the images have a tiny area of tumor, while the rest of the image is normal, benign, or fluid.
", "rawMarkdown": "It varies widely. Some of the images are purely tumor tissue, but in some areas the tumor is concentrated (solid), other areas are cystic (large spaces filled with fluid), other areas are infiltrative (tumor cells in between normal stroma). Some of the images have a tiny area of tumor, while the rest of the image is normal, benign, or fluid.", "votes": 17 } ] }, { "id": 2498345, "postDate": "2023-10-25T08:49:27.053Z", "content": "Hi there!
\nI wanna know that is the Other class means normal heathy cell , or it means other kind of subtypes of ovarian cancer?
\nTHX!
I think the challenge refers to something that looks like a tumor, but not one of the listed subtypes. Normal cells and tissues could be ignored.
", "rawMarkdown": "I think the challenge refers to something that looks like a tumor, but not one of the listed subtypes. Normal cells and tissues could be ignored.", "votes": 5 } ] }, { "id": 2474288, "postDate": "2023-10-09T05:37:02.090Z", "content": "Hi, the same specimen can result in images of slightly different darkness due to different staining environment, like time of staining, temperature, etc. Is it right?
", "rawMarkdown": "Hi, the same specimen can result in images of slightly different darkness due to different staining environment, like time of staining, temperature, etc. Is it right?", "votes": 1, "replies": [ { "id": 2474364, "postDate": "2023-10-09T06:45:36.287Z", "content": "These images are from the same type of specimen, an ovary with a tumor. Since the source of the images are from \"more than 20 centers across four continents\" it is very likely that there are differences of the images due to laboratory environment, such as:
\nHi there
\nare the tma images were taken from same patient in case of similar ovarian cancer subtype?
\ne.g in training data, we have 25 images, 5 images for each subtype. So i have a doubt if 5 images of same subtype taken from same patient?
I think it's safe to say that all of the images were from different patients.
", "rawMarkdown": "I think it's safe to say that all of the images were from different patients.", "votes": 4 } ] }, { "id": 2479101, "postDate": "2023-10-12T12:06:12.063Z", "content": "Hi, thank you for this professional view!
\nWhat are the key characteristics and criteria that pathologists use to identify and classify whole slide images (WSIs) as outliers, distinct from any specific cancer subtype, and what are the primary differences or unique features that distinguish these outliers from the rest of the WSIs in the context of ovarian cancer classification?
", "rawMarkdown": "Hi, thank you for this professional view!\n\nWhat are the key characteristics and criteria that pathologists use to identify and classify whole slide images (WSIs) as outliers, distinct from any specific cancer subtype, and what are the primary differences or unique features that distinguish these outliers from the rest of the WSIs in the context of ovarian cancer classification?", "votes": 2, "replies": [ { "id": 2479279, "postDate": "2023-10-12T14:20:00.593Z", "content": "I've been slowly going through the images, and I think some of the so-called outliers will be more of a \"rule out\" kind of decision by the AI model. The only way to train the model to recognize the rare outliers is by finding more labeled data, which is an added challenge.
\nFor example, there are Brenner tumors, which are not too common, but can be malignant, thus grouped as an ovarian cancer, and have a very distinct urothelial-like epithelium. Other distinct examples include Carcinosarcoma, Undifferentiated carcinoma, and Dedifferentiated carcinoma.
I've been reviewing the current WHO classification of ovarian tumors, https://www.pathologyoutlines.com/topic/ovarytumorwhoclassif.html, which the majority of pathologists use, and the only ovarian cancer that has a specific subtype in endometrioid carcinoma. It contains the **seromucinous **subtype, which used to be called seromucinous carcinoma, because there is significant morphological overlap with the conventional endometrial carcinoma. The distinguishable feature is that the tumor would be mostly composed of serous and endocervical-type mucinous epithelium, rather than the darker purple endometrial type.
\nI'll have to review some of ovary pathology specialty textbooks to see what other outlier variants there are.
Thank you for sharing valuable pathologist's perspective here. This clarifies many questions I had in interpreting the data. I was also wondering about ground truth errors and your post confirms that some level of noise is to be expected in the provided labels. I had not realized that TMA images contain mostly tumor areas until I read this post.
", "rawMarkdown": "Thank you for sharing valuable pathologist's perspective here. This clarifies many questions I had in interpreting the data. I was also wondering about ground truth errors and your post confirms that some level of noise is to be expected in the provided labels. I had not realized that TMA images contain mostly tumor areas until I read this post." }, { "id": 2558408, "postDate": "2023-12-12T06:29:37.297Z", "content": "Are the 25 TMA images given as training data enlarged images of the tumor part?
\nIn other words, is most of the tissue in the TMA image a tumor?
If the TMA image shows an enlargement of the tumor area, do you think the same trend would be true for the test data?
", "rawMarkdown": "Are the 25 TMA images given as training data enlarged images of the tumor part?\nIn other words, is most of the tissue in the TMA image a tumor?\n\nIf the TMA image shows an enlargement of the tumor area, do you think the same trend would be true for the test data?", "isDeleted": true, "replies": [ { "id": 2558503, "postDate": "2023-12-12T07:33:06.200Z", "content": "Essentially yes, they are enlarged tumor parts and should be downsampled since they are taken at twice the magnification.
\nAlthough most of the TMA images are filled with tumor cells, a few have about one third of the area consisting of non-tumor cells or fluids or spaces.
\nI think the format of the TMA images should be consistent whether they are test or train.
Thank you for your information.
", "rawMarkdown": "Thank you for your information.", "isDeleted": true }, { "id": 2558659, "postDate": "2023-12-12T09:03:25.453Z", "content": "\n\nthey are enlarged tumor parts
\n
Do you mean rather a crop tissue sample with the majority of cancer inside, correct?
\n\n\nshould be downsampled since they are taken at twice the magnification
\n
but the provided images of TMA and not TMA have the same pixel/cell ratio, correct? they do not need to be treated differently…
", "rawMarkdown": "> they are enlarged tumor parts\n\nDo you mean rather a crop tissue sample with the majority of cancer inside, correct?\n\n> should be downsampled since they are taken at twice the magnification\n\nbut the provided images of TMA and not TMA have the same pixel/cell ratio, correct? they do not need to be treated differently..." }, { "id": 2558725, "postDate": "2023-12-12T10:08:47.867Z", "content": "It's easier to visualize it like this image:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15149404%2Fac97dcf4af9aacfe6893543b6ac64988%2Ftma.jpg?generation=1702372526194114&alt=media\")
Every little circle is a sample of mostly tumor and can be any type of tumor. These were scanned as a whole with a 40x magnification objective lens, and then automatically cropped into multiple TMA images.
\nOn the other hand, whole tissue slides look like this and each was scanned with a 20x magnification objective lens:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F15149404%2F7c9830748cb64eabced76a2fe3da518c%2Fwsi.jpg?generation=1702375482122844&alt=media\")
Each image has a micrometers per pixel property, and it might not be set accordingly.
\nUsing a less scientific way, I viewed the slides using QuPath, and the TMA images have noticeable better details when zooming in.
\nI measured a small resting lymphocyte (a cell that can be roughly used as a gauge) in both TMA and non-TMA images and noticed that in TMA images they have roughly twice the area compared to non-TMA.
\nIt is probably better to take the data as it's described and resize the TMA files by a factor of 0.5 (downsample).
This is awesome explanation with illustrations!
", "rawMarkdown": "This is awesome explanation with illustrations! " }, { "id": 2563067, "postDate": "2023-12-16T02:13:21.087Z", "content": "Hi,@Noli Alonso.Thanks for your nice explanation! I have the same feeling that we need to treat TMA and WSI separately due to different magnification, as you suggested, maybe downsample is a good way to handle this…
\nHowever i still strongly believe we'd better add more TMA and I've been looking for similar datasets online for a long time…any ideas where i can find some tma about these Ovarian Cancer?
", "rawMarkdown": "Hi,@Noli Alonso.Thanks for your nice explanation! I have the same feeling that we need to treat TMA and WSI separately due to different magnification, as you suggested, maybe downsample is a good way to handle this...\n\nHowever i still strongly believe we'd better add more TMA and I've been looking for similar datasets online for a long time...any ideas where i can find some tma about these Ovarian Cancer?" }, { "id": 2563286, "postDate": "2023-12-16T07:15:16.813Z", "content": "It's easier to think of TMA images as physically annotated whole slide images, which guarantees that the specified tumor is included in the image. I don't think there would be many datasets of TMA images alone, because that would have a limited / specific purpose use, rather than WSI which has many uses.
\nWith the annotations provided by the competition, or with the ones I provided in a dataset, you can generate something equivalent to a TMA image, by selecting TMA-sized patches located within tumor areas of a WSI.
you're right i will try it later
", "rawMarkdown": "you're right i will try it later" } ] } ] } ], "comments": [ { "id": 2559062, "author_name": "Noli Alonso", "author_url": "", "post_date": "2023-12-12T15:22:48.487000", "content": "I've been reviewing the training images and have some additional tips that might be useful:
\nThere are many images that only contain a tiny amount of the specified tumor, and probably could be excluded or used as null. These are: 281, 3222, 5264, 9154, 12244, 26124, 31793, 32192, 33839, 41099, 52308, 54506, 63836.
\nImage number 1289 has way too many artifacts and may be excluded.
\nImage number 15583 should be labelled as MC not LGSC.
\nImage number 32035 can be excluded since it's distorted, and there are other examples of HGSC with the same morphology.
\nImage number 34822 is quite a unique example of EC.
Other things I noticed are:
\nPsammoma bodies, if present the diagnosis has a high probability of being LGSC:
\n
\nThey can range in size and number and cause many scratches of the images.
Giant cells, if present, the diagnosis always seems to be HGSC:
\n
Metaplasia, if present, favors EC instead of HGSC. The most common is squamous metaplasia which can look like this:
\n
Thank you for your work. Incorrect data has a significant impact on models, especially when there is limited training data. I will modify train.csv based on your suggestions and provide you with feedback on the result soon. I hope you can continue to provide us with more information.🥰
", "votes": 1, "replies": [ { "id": 2560799, "author_name": "zznznb", "author_url": "", "post_date": "2023-12-14T02:38:02.860000", "content": "Now I have some good news. Based on your suggestion, I removed number 1289 32035 and changed the label of 15583, which brought me 0.03 improvement. As I said, incorrect data can have a catastrophic impact on the model. I hope to receive more help from you.😄
", "votes": 9, "replies": [] } ] }, { "id": 2559098, "author_name": "Jirka", "author_url": "", "post_date": "2023-12-12T15:56:57.950000", "content": "WOW, thank you!
", "votes": 0, "replies": [] }, { "id": 2571185, "author_name": "Blumenkranz*e27", "author_url": "", "post_date": "2023-12-23T01:47:25.260000", "content": "I did some experiments according to your advice. I removed number 1289 32035 and changed the label of 15583 and it didnt improve my LB…then I removed more number like 281, 3222, 5264, 9154, e.t.c, lb score just increased by 0.01…hope this information would help
", "votes": 8, "replies": [] } ] }, { "id": 2507584, "author_name": "Jirka", "author_url": "", "post_date": "2023-11-01T05:41:01.747000", "content": "I can confirm that the tiles and detail view seem to be really important for this task,
\n as explained in baseline with Lightningâš¡TIMM scores 0.4+ on LB become TOP 5%
Hi there ! , are cancerous area is uniformly spread all over images or just concentrated on just only some part ?
", "votes": 4, "replies": [ { "id": 2484099, "author_name": "Noli Alonso", "author_url": "", "post_date": "2023-10-16T08:16:41.303000", "content": "It varies widely. Some of the images are purely tumor tissue, but in some areas the tumor is concentrated (solid), other areas are cystic (large spaces filled with fluid), other areas are infiltrative (tumor cells in between normal stroma). Some of the images have a tiny area of tumor, while the rest of the image is normal, benign, or fluid.
", "votes": 17, "replies": [] } ] }, { "id": 2498345, "author_name": "Mr.Fire", "author_url": "", "post_date": "2023-10-25T08:49:27.053000", "content": "Hi there!
\nI wanna know that is the Other class means normal heathy cell , or it means other kind of subtypes of ovarian cancer?
\nTHX!
I think the challenge refers to something that looks like a tumor, but not one of the listed subtypes. Normal cells and tissues could be ignored.
", "votes": 5, "replies": [] } ] }, { "id": 2474288, "author_name": "quantrader", "author_url": "", "post_date": "2023-10-09T05:37:02.090000", "content": "Hi, the same specimen can result in images of slightly different darkness due to different staining environment, like time of staining, temperature, etc. Is it right?
", "votes": 1, "replies": [ { "id": 2474364, "author_name": "Noli Alonso", "author_url": "", "post_date": "2023-10-09T06:45:36.287000", "content": "These images are from the same type of specimen, an ovary with a tumor. Since the source of the images are from \"more than 20 centers across four continents\" it is very likely that there are differences of the images due to laboratory environment, such as:
\nHi there
\nare the tma images were taken from same patient in case of similar ovarian cancer subtype?
\ne.g in training data, we have 25 images, 5 images for each subtype. So i have a doubt if 5 images of same subtype taken from same patient?
I think it's safe to say that all of the images were from different patients.
", "votes": 4, "replies": [] } ] }, { "id": 2479101, "author_name": "agape", "author_url": "", "post_date": "2023-10-12T12:06:12.063000", "content": "Hi, thank you for this professional view!
\nWhat are the key characteristics and criteria that pathologists use to identify and classify whole slide images (WSIs) as outliers, distinct from any specific cancer subtype, and what are the primary differences or unique features that distinguish these outliers from the rest of the WSIs in the context of ovarian cancer classification?
", "votes": 2, "replies": [ { "id": 2479279, "author_name": "Noli Alonso", "author_url": "", "post_date": "2023-10-12T14:20:00.593000", "content": "I've been slowly going through the images, and I think some of the so-called outliers will be more of a \"rule out\" kind of decision by the AI model. The only way to train the model to recognize the rare outliers is by finding more labeled data, which is an added challenge.
\nFor example, there are Brenner tumors, which are not too common, but can be malignant, thus grouped as an ovarian cancer, and have a very distinct urothelial-like epithelium. Other distinct examples include Carcinosarcoma, Undifferentiated carcinoma, and Dedifferentiated carcinoma.
I've been reviewing the current WHO classification of ovarian tumors, https://www.pathologyoutlines.com/topic/ovarytumorwhoclassif.html, which the majority of pathologists use, and the only ovarian cancer that has a specific subtype in endometrioid carcinoma. It contains the **seromucinous **subtype, which used to be called seromucinous carcinoma, because there is significant morphological overlap with the conventional endometrial carcinoma. The distinguishable feature is that the tumor would be mostly composed of serous and endocervical-type mucinous epithelium, rather than the darker purple endometrial type.
\nI'll have to review some of ovary pathology specialty textbooks to see what other outlier variants there are.
Thank you for sharing valuable pathologist's perspective here. This clarifies many questions I had in interpreting the data. I was also wondering about ground truth errors and your post confirms that some level of noise is to be expected in the provided labels. I had not realized that TMA images contain mostly tumor areas until I read this post.
", "votes": 0, "replies": [] }, { "id": 2558408, "author_name": "", "author_url": "", "post_date": "2023-12-12T06:29:37.297000", "content": "Are the 25 TMA images given as training data enlarged images of the tumor part?
\nIn other words, is most of the tissue in the TMA image a tumor?
If the TMA image shows an enlargement of the tumor area, do you think the same trend would be true for the test data?
", "votes": 0, "replies": [ { "id": 2558503, "author_name": "Noli Alonso", "author_url": "", "post_date": "2023-12-12T07:33:06.200000", "content": "Essentially yes, they are enlarged tumor parts and should be downsampled since they are taken at twice the magnification.
\nAlthough most of the TMA images are filled with tumor cells, a few have about one third of the area consisting of non-tumor cells or fluids or spaces.
\nI think the format of the TMA images should be consistent whether they are test or train.
Thank you for your information.
", "votes": 0, "replies": [] }, { "id": 2558659, "author_name": "Jirka", "author_url": "", "post_date": "2023-12-12T09:03:25.453000", "content": "\n\nthey are enlarged tumor parts
\n
Do you mean rather a crop tissue sample with the majority of cancer inside, correct?
\n\n\nshould be downsampled since they are taken at twice the magnification
\n
but the provided images of TMA and not TMA have the same pixel/cell ratio, correct? they do not need to be treated differently…
", "votes": 0, "replies": [] }, { "id": 2558725, "author_name": "Noli Alonso", "author_url": "", "post_date": "2023-12-12T10:08:47.867000", "content": "It's easier to visualize it like this image:
\n
Every little circle is a sample of mostly tumor and can be any type of tumor. These were scanned as a whole with a 40x magnification objective lens, and then automatically cropped into multiple TMA images.
\nOn the other hand, whole tissue slides look like this and each was scanned with a 20x magnification objective lens:
\n
Each image has a micrometers per pixel property, and it might not be set accordingly.
\nUsing a less scientific way, I viewed the slides using QuPath, and the TMA images have noticeable better details when zooming in.
\nI measured a small resting lymphocyte (a cell that can be roughly used as a gauge) in both TMA and non-TMA images and noticed that in TMA images they have roughly twice the area compared to non-TMA.
\nIt is probably better to take the data as it's described and resize the TMA files by a factor of 0.5 (downsample).
This is awesome explanation with illustrations!
", "votes": 0, "replies": [] }, { "id": 2563067, "author_name": "Blumenkranz*e27", "author_url": "", "post_date": "2023-12-16T02:13:21.087000", "content": "Hi,@Noli Alonso.Thanks for your nice explanation! I have the same feeling that we need to treat TMA and WSI separately due to different magnification, as you suggested, maybe downsample is a good way to handle this…
\nHowever i still strongly believe we'd better add more TMA and I've been looking for similar datasets online for a long time…any ideas where i can find some tma about these Ovarian Cancer?
", "votes": 0, "replies": [] }, { "id": 2563286, "author_name": "Noli Alonso", "author_url": "", "post_date": "2023-12-16T07:15:16.813000", "content": "It's easier to think of TMA images as physically annotated whole slide images, which guarantees that the specified tumor is included in the image. I don't think there would be many datasets of TMA images alone, because that would have a limited / specific purpose use, rather than WSI which has many uses.
\nWith the annotations provided by the competition, or with the ones I provided in a dataset, you can generate something equivalent to a TMA image, by selecting TMA-sized patches located within tumor areas of a WSI.
you're right i will try it later
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2474181": "Hi everyone, I would like to share how a pathologist generally handles this task.\nThrough the laboratory, a tissue sample/specimen is received, processed, and stained with hematoxylin and eosin (H&E) before being given to the pathologist, who uses their expertise and experience to arrive at a diagnosis of the specimen. \nThis is essentially classification task because we already know that the specimen contains a tumor, we just have to decide which \"class\" of tumor it is. Each of the classes has a set of morphological criteria that includes presence or absence of a certain feature, which can be found in the WHO classification of tumors and many other textbooks.\nFrom the given data, there are large whole slide images, and small tissue microarray (TMA) images. The TMA images are essentially a small patch of the labelled tumor class, while the whole slide images contain many patches of the tumor class, along with the other morphologies that may or may not be relevant to deciding the tumor class. For example, high grade serous carcinoma is notoriously \"ugly\" looking because it has \"high-grade\" nuclei and can have extensive areas of necrosis (dead cells). The model should factor in areas of tumor and areas of necrosis to arrive at the correct classification, kind of like classification within a classification.\nAdded to the work burden, is that processing of specimens is prone to variations in the staining, cutting, optics, etc., that the classification model must be able to handle well.\n", "2559062": "I've been reviewing the training images and have some additional tips that might be useful:\nThere are many images that only contain a tiny amount of the specified tumor, and probably could be excluded or used as null. These are: 281, 3222, 5264, 9154, 12244, 26124, 31793, 32192, 33839, 41099, 52308, 54506, 63836.\nImage number 1289 has way too many artifacts and may be excluded.\nImage number 15583 should be labelled as MC not LGSC.\nImage number 32035 can be excluded since it's distorted, and there are other examples of HGSC with the same morphology.\nImage number 34822 is quite a unique example of EC.\n\nOther things I noticed are:\nPsammoma bodies, if present the diagnosis has a high probability of being LGSC:\n\nThey can range in size and number and cause many scratches of the images.\n\nGiant cells, if present, the diagnosis always seems to be HGSC:\n\n\nMetaplasia, if present, favors EC instead of HGSC. The most common is squamous metaplasia which can look like this:\n\n", "2507584": "I can confirm that the tiles and detail view seem to be really important for this task,\n as explained in [baseline with Lightningâš¡TIMM scores 0.4+ on LB become TOP 5%](https://kaggle.com/competitions/UBC-OCEAN/discussion/452165)", "2484058": "Hi there ! , are cancerous area is uniformly spread all over images or just concentrated on just only some part ?", "2498345": "Hi there!\nI wanna know that is the Other class means normal heathy cell , or it means other kind of subtypes of ovarian cancer?\nTHX!", "2474288": "Hi, the same specimen can result in images of slightly different darkness due to different staining environment, like time of staining, temperature, etc. Is it right?", "2485450": "Hi there\nare the tma images were taken from same patient in case of similar ovarian cancer subtype?\ne.g in training data, we have 25 images, 5 images for each subtype. So i have a doubt if 5 images of same subtype taken from same patient?\n", "2479101": "Hi, thank you for this professional view!\n\nWhat are the key characteristics and criteria that pathologists use to identify and classify whole slide images (WSIs) as outliers, distinct from any specific cancer subtype, and what are the primary differences or unique features that distinguish these outliers from the rest of the WSIs in the context of ovarian cancer classification?", "2580661": "Thank you for sharing valuable pathologist's perspective here. This clarifies many questions I had in interpreting the data. I was also wondering about ground truth errors and your post confirms that some level of noise is to be expected in the provided labels. I had not realized that TMA images contain mostly tumor areas until I read this post.", "2558408": "Are the 25 TMA images given as training data enlarged images of the tumor part?\nIn other words, is most of the tissue in the TMA image a tumor?\n\nIf the TMA image shows an enlargement of the tumor area, do you think the same trend would be true for the test data?" } }