![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F983147%2Fd33b7efc54601e16ab18433c245b8c11%2Fleak.png?generation=1573702109414798&alt=media\")
trick1: If you sort by [patient ID, 'imageposition2'] you will find the lable is continue\ntrick2: patientID have an overlap in stage1 (not work at stage2)
\n\nYou can take this into feature enginnering. \nFor example, patientID with lable encoder, agg groupy, count encoding and lable encoding.\nYou can extract some time series feature like lag, diff, next/last, etc..
\n\nAll you need is a LightGBM/XGboost/Catboost or do some postprocessing.\nI perform a stacking with these features. It works well.
\n\n
PS: I think seutao's sequence model is the best solution. All roads lead to Rome.
", "messages": [ { "id": 672647, "postDate": "2019-11-14T03:34:09.680Z", "content": "trick1: If you sort by [patient ID, 'imageposition2'] you will find the lable is continue\ntrick2: patientID have an overlap in stage1 (not work at stage2)
\n\nYou can take this into feature enginnering. \nFor example, patientID with lable encoder, agg groupy, count encoding and lable encoding.\nYou can extract some time series feature like lag, diff, next/last, etc..
\n\nAll you need is a LightGBM/XGboost/Catboost or do some postprocessing.\nI perform a stacking with these features. It works well.
\n\n
PS: I think seutao's sequence model is the best solution. All roads lead to Rome.
", "rawMarkdown": "trick1: If you sort by [patient ID, 'imageposition2'] you will find the lable is continue\ntrick2: patientID have an overlap in stage1 (not work at stage2)\n\nYou can take this into feature enginnering. \nFor example, patientID with lable encoder, agg groupy, count encoding and lable encoding.\nYou can extract some time series feature like lag, diff, next/last, etc..\n\nAll you need is a LightGBM/XGboost/Catboost or do some postprocessing.\nI perform a stacking with these features. It works well. \n\n\n\nPS: I think seutao's sequence model is the best solution. All roads lead to Rome.", "votes": 19 }, { "id": 672751, "postDate": "2019-11-14T05:50:32.280Z", "content": "One patient might has multil series, so group by study_id make sense.
", "rawMarkdown": "One patient might has multil series, so group by study_id make sense.", "votes": 3, "replies": [ { "id": 675581, "postDate": "2019-11-18T09:02:35.397Z", "content": "@lanjunyelan \nCould u please help understand below things, i m struggling to understand the purpose of order by Z axis and group by Various fields like StudyInstanceUID in terms of what useful information does it provides.
\n\n1) When we do order by ImagePostition2 ,In what way it helps the classification model.\n2) Is sequencing useful for single slice trained classification model using just windowing.\n3) Single PatientID is found under different studyID and seriesInstanceUID and has got different Target,what does it implies\n4) In what way we use group by StudyInstance UID on sorted Z axis data
", "rawMarkdown": "@lanjunyelan \nCould u please help understand below things, i m struggling to understand the purpose of order by Z axis and group by Various fields like StudyInstanceUID in terms of what useful information does it provides.\n\n1) When we do order by ImagePostition2 ,In what way it helps the classification model.\n2) Is sequencing useful for single slice trained classification model using just windowing.\n3) Single PatientID is found under different studyID and seriesInstanceUID and has got different Target,what does it implies\n4) In what way we use group by StudyInstance UID on sorted Z axis data" }, { "id": 678174, "postDate": "2019-11-21T03:55:13.250Z", "content": "I'm sorry for replying late. There are some domain knowledge in this competition.\n1. Input of the model is the combination of three continuous images from the same SeriesInstanceUID. There are more information in three continuous images than single image\n2. Yes, There are continuous label in series.\n3. I guess one patient might have multil CTs.\n4. pandas groupby and sort
", "rawMarkdown": "I'm sorry for replying late. There are some domain knowledge in this competition.\n1. Input of the model is the combination of three continuous images from the same SeriesInstanceUID. There are more information in three continuous images than single image\n2. Yes, There are continuous label in series.\n3. I guess one patient might have multil CTs.\n4. pandas groupby and sort" } ] }, { "id": 672650, "postDate": "2019-11-14T03:38:48.213Z", "content": "wow, great!
", "rawMarkdown": "wow, great!", "votes": 1, "replies": [ { "id": 672651, "postDate": "2019-11-14T03:39:36.257Z", "content": "Thanks, 阿水老师.
", "rawMarkdown": "Thanks, 阿水老师.", "votes": 1 } ] }, { "id": 673039, "postDate": "2019-11-14T12:42:17.413Z", "content": "Indeed, this is more or less our post-processing as well (more details at https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/117232#latest-673036)
", "rawMarkdown": "Indeed, this is more or less our post-processing as well (more details at https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/117232#latest-673036)" }, { "id": 672893, "postDate": "2019-11-14T08:55:21.183Z", "content": "Alas, I was so focused on modeling I only started working on this idea during stage2, out of curiosity... Good to see that some people had this kind of idea too and that it worked. Thanks for sharing your ideas though, I didn't think of the lag for example !
", "rawMarkdown": "Alas, I was so focused on modeling I only started working on this idea during stage2, out of curiosity... Good to see that some people had this kind of idea too and that it worked. Thanks for sharing your ideas though, I didn't think of the lag for example ! " }, { "id": 672783, "postDate": "2019-11-14T06:38:23.137Z", "content": "", "rawMarkdown": "@baomengjiao 👍" }, { "id": 672655, "postDate": "2019-11-14T03:43:32.083Z", "content": "great
", "rawMarkdown": "great" }, { "id": 673567, "postDate": "2019-11-15T07:01:27.087Z", "content": "thanks for sharing
", "rawMarkdown": "thanks for sharing" } ], "comments": [ { "id": 672751, "author_name": "yelan", "author_url": "", "post_date": "2019-11-14T05:50:32.280000", "content": "One patient might has multil series, so group by study_id make sense.
", "votes": 3, "replies": [ { "id": 675581, "author_name": "Jaideep", "author_url": "", "post_date": "2019-11-18T09:02:35.397000", "content": "@lanjunyelan \nCould u please help understand below things, i m struggling to understand the purpose of order by Z axis and group by Various fields like StudyInstanceUID in terms of what useful information does it provides.
\n\n1) When we do order by ImagePostition2 ,In what way it helps the classification model.\n2) Is sequencing useful for single slice trained classification model using just windowing.\n3) Single PatientID is found under different studyID and seriesInstanceUID and has got different Target,what does it implies\n4) In what way we use group by StudyInstance UID on sorted Z axis data
", "votes": 0, "replies": [] }, { "id": 678174, "author_name": "yelan", "author_url": "", "post_date": "2019-11-21T03:55:13.250000", "content": "I'm sorry for replying late. There are some domain knowledge in this competition.\n1. Input of the model is the combination of three continuous images from the same SeriesInstanceUID. There are more information in three continuous images than single image\n2. Yes, There are continuous label in series.\n3. I guess one patient might have multil CTs.\n4. pandas groupby and sort
", "votes": 0, "replies": [] } ] }, { "id": 672650, "author_name": "Finlay", "author_url": "", "post_date": "2019-11-14T03:38:48.213000", "content": "wow, great!
", "votes": 1, "replies": [ { "id": 672651, "author_name": "spongebob", "author_url": "", "post_date": "2019-11-14T03:39:36.257000", "content": "Thanks, 阿水老师.
", "votes": 1, "replies": [] } ] }, { "id": 673039, "author_name": "tarobxl", "author_url": "", "post_date": "2019-11-14T12:42:17.413000", "content": "Indeed, this is more or less our post-processing as well (more details at https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/117232#latest-673036)
", "votes": 0, "replies": [] }, { "id": 672893, "author_name": "Benjamin Dubreu", "author_url": "", "post_date": "2019-11-14T08:55:21.183000", "content": "Alas, I was so focused on modeling I only started working on this idea during stage2, out of curiosity... Good to see that some people had this kind of idea too and that it worked. Thanks for sharing your ideas though, I didn't think of the lag for example !
", "votes": 0, "replies": [] }, { "id": 672783, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2019-11-14T06:38:23.137000", "content": "", "votes": 0, "replies": [] }, { "id": 672655, "author_name": "Zuping Wu", "author_url": "", "post_date": "2019-11-14T03:43:32.083000", "content": "great
", "votes": 0, "replies": [] }, { "id": 673567, "author_name": "yangDDD", "author_url": "", "post_date": "2019-11-15T07:01:27.087000", "content": "thanks for sharing
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "672647": "trick1: If you sort by [patient ID, 'imageposition2'] you will find the lable is continue\ntrick2: patientID have an overlap in stage1 (not work at stage2)\n\nYou can take this into feature enginnering. \nFor example, patientID with lable encoder, agg groupy, count encoding and lable encoding.\nYou can extract some time series feature like lag, diff, next/last, etc..\n\nAll you need is a LightGBM/XGboost/Catboost or do some postprocessing.\nI perform a stacking with these features. It works well. \n\n\n\nPS: I think seutao's sequence model is the best solution. All roads lead to Rome.", "672751": "One patient might has multil series, so group by study_id make sense.", "672650": "wow, great!", "673039": "Indeed, this is more or less our post-processing as well (more details at https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/117232#latest-673036)", "672893": "Alas, I was so focused on modeling I only started working on this idea during stage2, out of curiosity... Good to see that some people had this kind of idea too and that it worked. Thanks for sharing your ideas though, I didn't think of the lag for example ! ", "672783": "@baomengjiao 👍", "672655": "great", "673567": "thanks for sharing" } }