![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3842406%2F5e9ef40a887d489c3348f801bb84f27b%2FUntitled.png?generation=1698414051863680&alt=media\")
Hi,
\nIs it feasible to extract individual sub-images from a larger image and create a dataset where each sub-image is assigned the same label? In other words, can we assume that any cancer present in the original image is represented consistently across all the sub-images?
\n
thanks!
", "messages": [ { "id": 2502584, "postDate": "2023-10-28T09:38:06.420Z", "content": "From my experience outside this competition you cannot say that cancer will be represented on each \"sub-image\" where each sub-image is actually a separate tissue part on the scanned glass.
\nThe whole process of WSIs preparation consists of:
\nThat's why different \"sub-images\" fom one WSI are actually different slices of the sample and thus they may have different structure and one may have cancer cells while the other - doesn't.
\nSo the correct approach would be: if at least on one \"sub-image\" you find particular cancer cells - you should label the whole original WSI with this label
", "rawMarkdown": "From my experience outside this competition you cannot say that cancer will be represented on each \"sub-image\" where each sub-image is actually a separate tissue part on the scanned glass.\n\nThe whole process of WSIs preparation consists of:\n1. biopsy or other tissue sample extraction from human body\n2. preparation slices of the tissue sample\n3. slices are put on the glass (**more than one slices can be on one glass**)\n4. the glass is scanned by special histoscanner\n5. scanner puts out WSI\n\nThat's why different \"sub-images\" fom one WSI are actually different slices of the sample and thus they may have different structure and one may have cancer cells while the other - doesn't.\n\nSo the correct approach would be: if at least on one \"sub-image\" you find particular cancer cells - you should label the whole original WSI with this label", "votes": 1, "replies": [ { "id": 2503034, "postDate": "2023-10-28T17:32:28.473Z", "content": "Thank you for the clarification. I understand your points. However, after reviewing the dataset, I believe that a significant portion of it may indeed contain redundant data. This redundancy has the potential to adversely impact the performance of the classification algorithm if we do not take measures to separate and crop these redundant slices into distinct images. From a classification perspective, it can be challenging to augment such behavior. Therefore, it may be preferable to retain some outliers in case the slices do not accurately represent the cancer type.
", "rawMarkdown": "Thank you for the clarification. I understand your points. However, after reviewing the dataset, I believe that a significant portion of it may indeed contain redundant data. This redundancy has the potential to adversely impact the performance of the classification algorithm if we do not take measures to separate and crop these redundant slices into distinct images. From a classification perspective, it can be challenging to augment such behavior. Therefore, it may be preferable to retain some outliers in case the slices do not accurately represent the cancer type.", "votes": 1 } ] }, { "id": 2501488, "postDate": "2023-10-27T13:41:55.457Z", "content": "Hi,
\nIs it feasible to extract individual sub-images from a larger image and create a dataset where each sub-image is assigned the same label? In other words, can we assume that any cancer present in the original image is represented consistently across all the sub-images?
\n
thanks!
", "rawMarkdown": "Hi,\n\nIs it feasible to extract individual sub-images from a larger image and create a dataset where each sub-image is assigned the same label? In other words, can we assume that any cancer present in the original image is represented consistently across all the sub-images?\n\n\n\nthanks!\n", "votes": 1 } ], "comments": [ { "id": 2502584, "author_name": "Prosperous Silver Alpaca", "author_url": "", "post_date": "2023-10-28T09:38:06.420000", "content": "From my experience outside this competition you cannot say that cancer will be represented on each \"sub-image\" where each sub-image is actually a separate tissue part on the scanned glass.
\nThe whole process of WSIs preparation consists of:
\nThat's why different \"sub-images\" fom one WSI are actually different slices of the sample and thus they may have different structure and one may have cancer cells while the other - doesn't.
\nSo the correct approach would be: if at least on one \"sub-image\" you find particular cancer cells - you should label the whole original WSI with this label
", "votes": 1, "replies": [ { "id": 2503034, "author_name": "Cyrus", "author_url": "", "post_date": "2023-10-28T17:32:28.473000", "content": "Thank you for the clarification. I understand your points. However, after reviewing the dataset, I believe that a significant portion of it may indeed contain redundant data. This redundancy has the potential to adversely impact the performance of the classification algorithm if we do not take measures to separate and crop these redundant slices into distinct images. From a classification perspective, it can be challenging to augment such behavior. Therefore, it may be preferable to retain some outliers in case the slices do not accurately represent the cancer type.
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "2502584": "From my experience outside this competition you cannot say that cancer will be represented on each \"sub-image\" where each sub-image is actually a separate tissue part on the scanned glass.\n\nThe whole process of WSIs preparation consists of:\n1. biopsy or other tissue sample extraction from human body\n2. preparation slices of the tissue sample\n3. slices are put on the glass (**more than one slices can be on one glass**)\n4. the glass is scanned by special histoscanner\n5. scanner puts out WSI\n\nThat's why different \"sub-images\" fom one WSI are actually different slices of the sample and thus they may have different structure and one may have cancer cells while the other - doesn't.\n\nSo the correct approach would be: if at least on one \"sub-image\" you find particular cancer cells - you should label the whole original WSI with this label", "2501488": "Hi,\n\nIs it feasible to extract individual sub-images from a larger image and create a dataset where each sub-image is assigned the same label? In other words, can we assume that any cancer present in the original image is represented consistently across all the sub-images?\n\n\n\nthanks!\n" } }