The supplied images are 96x96 but the competition description states that cancer is present in the center 32x32 area when indicated by the labels. My understanding is that we are searching the test data fro cancer in the center 32x32, but it seems that everyone is using the full image in their models. Am I reading the evaluation criteria wrong? Is it that cancer is in the center 32x32 of the training images but we are looking for cancer anywhere in the 96x96 test data images?
", "messages": [ { "id": "444291", "postDate": "12/23/2018 18:35:41", "content": "The supplied images are 96x96 but the competition description states that cancer is present in the center 32x32 area when indicated by the labels. My understanding is that we are searching the test data fro cancer in the center 32x32, but it seems that everyone is using the full image in their models. Am I reading the evaluation criteria wrong? Is it that cancer is in the center 32x32 of the training images but we are looking for cancer anywhere in the 96x96 test data images?
", "rawMarkdown": "The supplied images are 96x96 but the competition description states that cancer is present in the center 32x32 area when indicated by the labels. My understanding is that we are searching the test data fro cancer in the center 32x32, but it seems that everyone is using the full image in their models. Am I reading the evaluation criteria wrong? Is it that cancer is in the center 32x32 of the training images but we are looking for cancer anywhere in the 96x96 test data images?", "votes": null }, { "id": "444465", "postDate": "12/24/2018 04:28:50", "content": "just additional information for nn
", "rawMarkdown": "just additional information for nn", "votes": null }, { "id": "444820", "postDate": "12/24/2018 22:43:37", "content": "Is it not noise though? The label has no bearing on outer area. How could the nn really “know” if any of the data is useful?
", "rawMarkdown": "Is it not noise though? The label has no bearing on outer area. How could the nn really “know” if any of the data is useful?", "votes": null }, { "id": "444960", "postDate": "12/25/2018 08:23:25", "content": "Same issue
", "rawMarkdown": "Same issue", "votes": null }, { "id": "445208", "postDate": "12/26/2018 00:16:25", "content": "I have same question
", "rawMarkdown": "I have same question", "votes": null }, { "id": "445482", "postDate": "12/26/2018 14:56:35", "content": "may be some features outside the 32x32 area that may give u more information
", "rawMarkdown": "may be some features outside the 32x32 area that may give u more information", "votes": null }, { "id": "445984", "postDate": "12/27/2018 09:56:04", "content": "I have the same issue. I started to train with 96x96 but will also train with 32x32 and compare the results.
", "rawMarkdown": "I have the same issue. I started to train with 96x96 but will also train with 32x32 and compare the results.", "votes": null }, { "id": "446315", "postDate": "12/27/2018 21:19:53", "content": "I tried 32x32 but I think Im having compound issues with my model and solution output. My biggest concern is that, while 96x96 training may give us more info, if the final solution evaluation is based on the center 32x32 but the model thinks its \"sees\" cancer in some other quadrant it may report a false positive.
", "rawMarkdown": "I tried 32x32 but I think Im having compound issues with my model and solution output. My biggest concern is that, while 96x96 training may give us more info, if the final solution evaluation is based on the center 32x32 but the model thinks its \"sees\" cancer in some other quadrant it may report a false positive.", "votes": null }, { "id": "447027", "postDate": "12/29/2018 02:42:13", "content": "According to the PCam dataset on GitHub:
\n\n\"A positive label indicates that the center 32x32px region of a patch contains at least one pixel of tumor tissue. Tumor tissue in the outer region of the patch does not influence the label\"
\n\nStill, when I switched from using 32x32 to 96x96 sized images, my validation and testing accuracy go from low 80's to low 90's. I'm not sure why.
", "rawMarkdown": "According to the PCam dataset on GitHub:\n\n*\"A positive label indicates that the center 32x32px region of a patch contains at least one pixel of tumor tissue. Tumor tissue in the outer region of the patch does not influence the label\"*\n\nStill, when I switched from using 32x32 to 96x96 sized images, my validation and testing accuracy go from low 80's to low 90's. I'm not sure why.", "votes": null }, { "id": "454748", "postDate": "01/12/2019 06:06:20", "content": "According to competition we have to detect tumor on the centre area of 32X32, the rest is not useful, it is given so that we do not require zero padding. The data is labeled such that, it is true only when there is maligned tumor in the given specific area not outside it.
", "rawMarkdown": "According to competition we have to detect tumor on the centre area of 32X32, the rest is not useful, it is given so that we do not require zero padding. The data is labeled such that, it is true only when there is maligned tumor in the given specific area not outside it.", "votes": null }, { "id": "456240", "postDate": "01/15/2019 11:58:13", "content": "Hey Reid, I also tried with 32x32 , but the validation accuracy was nearly 0,80. I have found that the best validation accuracies are in the range 90x90 - 96x96.
", "rawMarkdown": "Hey Reid, I also tried with 32x32 , but the validation accuracy was nearly 0,80. I have found that the best validation accuracies are in the range 90x90 - 96x96.", "votes": null }, { "id": "499055", "postDate": "03/24/2019 09:10:48", "content": "I used 32x32 but score is lower than 96x96.
", "rawMarkdown": "I used 32x32 but score is lower than 96x96.", "votes": null } ], "comments": [ { "id": 444465, "author_name": "justajoke", "author_url": "", "post_date": "12/24/2018 04:28:50", "content": "just additional information for nn
", "votes": null, "replies": [ { "id": 444820, "author_name": "reidtc", "author_url": "", "post_date": "12/24/2018 22:43:37", "content": "Is it not noise though? The label has no bearing on outer area. How could the nn really “know” if any of the data is useful?
", "votes": null, "replies": [] }, { "id": 445482, "author_name": "justajoke", "author_url": "", "post_date": "12/26/2018 14:56:35", "content": "may be some features outside the 32x32 area that may give u more information
", "votes": null, "replies": [] } ] }, { "id": 444960, "author_name": "arthasmenethil", "author_url": "", "post_date": "12/25/2018 08:23:25", "content": "Same issue
", "votes": null, "replies": [] }, { "id": 445208, "author_name": "samyssmile", "author_url": "", "post_date": "12/26/2018 00:16:25", "content": "I have same question
", "votes": null, "replies": [] }, { "id": 445984, "author_name": "robotdreams", "author_url": "", "post_date": "12/27/2018 09:56:04", "content": "I have the same issue. I started to train with 96x96 but will also train with 32x32 and compare the results.
", "votes": null, "replies": [ { "id": 446315, "author_name": "reidtc", "author_url": "", "post_date": "12/27/2018 21:19:53", "content": "I tried 32x32 but I think Im having compound issues with my model and solution output. My biggest concern is that, while 96x96 training may give us more info, if the final solution evaluation is based on the center 32x32 but the model thinks its \"sees\" cancer in some other quadrant it may report a false positive.
", "votes": null, "replies": [] }, { "id": 456240, "author_name": "robotdreams", "author_url": "", "post_date": "01/15/2019 11:58:13", "content": "Hey Reid, I also tried with 32x32 , but the validation accuracy was nearly 0,80. I have found that the best validation accuracies are in the range 90x90 - 96x96.
", "votes": null, "replies": [] } ] }, { "id": 447027, "author_name": "dwoooo", "author_url": "", "post_date": "12/29/2018 02:42:13", "content": "According to the PCam dataset on GitHub:
\n\n\"A positive label indicates that the center 32x32px region of a patch contains at least one pixel of tumor tissue. Tumor tissue in the outer region of the patch does not influence the label\"
\n\nStill, when I switched from using 32x32 to 96x96 sized images, my validation and testing accuracy go from low 80's to low 90's. I'm not sure why.
", "votes": null, "replies": [] }, { "id": 454748, "author_name": "asking28", "author_url": "", "post_date": "01/12/2019 06:06:20", "content": "According to competition we have to detect tumor on the centre area of 32X32, the rest is not useful, it is given so that we do not require zero padding. The data is labeled such that, it is true only when there is maligned tumor in the given specific area not outside it.
", "votes": null, "replies": [] }, { "id": 499055, "author_name": "wuzhongli", "author_url": "", "post_date": "03/24/2019 09:10:48", "content": "I used 32x32 but score is lower than 96x96.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "444291": "The supplied images are 96x96 but the competition description states that cancer is present in the center 32x32 area when indicated by the labels. My understanding is that we are searching the test data fro cancer in the center 32x32, but it seems that everyone is using the full image in their models. Am I reading the evaluation criteria wrong? Is it that cancer is in the center 32x32 of the training images but we are looking for cancer anywhere in the 96x96 test data images?", "444465": "just additional information for nn", "444820": "Is it not noise though? The label has no bearing on outer area. How could the nn really “know” if any of the data is useful?", "444960": "Same issue", "445208": "I have same question", "445482": "may be some features outside the 32x32 area that may give u more information", "445984": "I have the same issue. I started to train with 96x96 but will also train with 32x32 and compare the results.", "446315": "I tried 32x32 but I think Im having compound issues with my model and solution output. My biggest concern is that, while 96x96 training may give us more info, if the final solution evaluation is based on the center 32x32 but the model thinks its \"sees\" cancer in some other quadrant it may report a false positive.", "447027": "According to the PCam dataset on GitHub:\n\n*\"A positive label indicates that the center 32x32px region of a patch contains at least one pixel of tumor tissue. Tumor tissue in the outer region of the patch does not influence the label\"*\n\nStill, when I switched from using 32x32 to 96x96 sized images, my validation and testing accuracy go from low 80's to low 90's. I'm not sure why.", "454748": "According to competition we have to detect tumor on the centre area of 32X32, the rest is not useful, it is given so that we do not require zero padding. The data is labeled such that, it is true only when there is maligned tumor in the given specific area not outside it.", "456240": "Hey Reid, I also tried with 32x32 , but the validation accuracy was nearly 0,80. I have found that the best validation accuracies are in the range 90x90 - 96x96.", "499055": "I used 32x32 but score is lower than 96x96." }, "source": "meta" }