Hi,\n I am confused about the overview description of this as acute hemorrhage prediction. Many cases have craniectomys and drains, which I would have thought indicated treatment and thereby a follow up scans?
\nThanks!
I'm not a medical expert and have close to non-medical expertise. However, in neuroscience research there are quite often discussions about what time-ranges to apply when talking about stroke. Especially, how to deal with different time-windows between neuro-psychological assessment and CT or MRI recordings. \nAnd according to this random source (http://casemed.case.edu/clerkships/neurology/Web%20Neurorad/MultHemorrhageTypes.htm), the classification of acute is quite arbitrary and means up to 4 days after the hemorrhage. So I think your intuition might be right, that there are some CTs which were taken after initial treatment.
", "rawMarkdown": "I'm not a medical expert and have close to non-medical expertise. However, in neuroscience research there are quite often discussions about what time-ranges to apply when talking about stroke. Especially, how to deal with different time-windows between neuro-psychological assessment and CT or MRI recordings. \nAnd according to this random source ([http://casemed.case.edu/clerkships/neurology/Web%20Neurorad/MultHemorrhageTypes.htm](http://casemed.case.edu/clerkships/neurology/Web%20Neurorad/MultHemorrhageTypes.htm)), the classification of acute is quite arbitrary and means up to 4 days after the hemorrhage. So I think your intuition might be right, that there are some CTs which were taken after initial treatment.", "votes": 1, "replies": [ { "id": 637733, "postDate": "2019-10-01T08:14:47.253Z", "content": "Thanks for the link!
\nIt seems to me that if the purpose of the model is to be applicable in the \"hyperacute\" setting - that is when the initial diagnosis is made then it is counterproductive to have craniectomy+drains in the dataset. Clearly presence of these will correlate with hemorrhage and become focus points for the models - but they wont be present in the actual data where this model could be useful. It's possible I am wrong but this is how it seems to me at the moment.
Totally agree with you. I think we might get into an issue here, where the best models might not be the most useful ones in real-life.\nI trying to account for this (at least a bit) by using some simple segmentation approach (as in one of my kernels) to extract only the patient's brain. But I have no idea yet, whether the approach actually works and what it does to Deep Learning models.
", "rawMarkdown": "Totally agree with you. I think we might get into an issue here, where the best models might not be the most useful ones in real-life.\nI trying to account for this (at least a bit) by using some simple segmentation approach (as in one of my kernels) to extract only the patient's brain. But I have no idea yet, whether the approach actually works and what it does to Deep Learning models." }, { "id": 637844, "postDate": "2019-10-01T09:55:02.997Z", "content": "Good luck! The high number of studies poses a challenge to masking etc. With 19000 studies just 1 minute of preprocessing is 2 weeks! How do you manage that?
", "rawMarkdown": "Good luck! The high number of studies poses a challenge to masking etc. With 19000 studies just 1 minute of preprocessing is 2 weeks! How do you manage that?" }, { "id": 637915, "postDate": "2019-10-01T11:18:11.920Z", "content": "I used something similar as in here (shameless self-plug), the kernel has unfortunately still some issues. Background separation, which is basically given by the windowing parameters, whole filling and dilation etc. just takes a second. It's definitely not perfect, that's why I need to test it out. It took about 12h to process all slices (not using anything parallel).
", "rawMarkdown": "I used something similar as in [here (shameless self-plug)](https://www.kaggle.com/srsteinkamp/intuitions-ideas-basic-preprocessing), the kernel has unfortunately still some issues. Background separation, which is basically given by the windowing parameters, whole filling and dilation etc. just takes a second. It's definitely not perfect, that's why I need to test it out. It took about 12h to process all slices (not using anything parallel). ", "votes": 1 } ] }, { "id": 637249, "postDate": "2019-09-30T19:56:05.113Z", "content": "Hi,\n I am confused about the overview description of this as acute hemorrhage prediction. Many cases have craniectomys and drains, which I would have thought indicated treatment and thereby a follow up scans?
\nThanks!
I'm not a medical expert and have close to non-medical expertise. However, in neuroscience research there are quite often discussions about what time-ranges to apply when talking about stroke. Especially, how to deal with different time-windows between neuro-psychological assessment and CT or MRI recordings. \nAnd according to this random source (http://casemed.case.edu/clerkships/neurology/Web%20Neurorad/MultHemorrhageTypes.htm), the classification of acute is quite arbitrary and means up to 4 days after the hemorrhage. So I think your intuition might be right, that there are some CTs which were taken after initial treatment.
", "votes": 1, "replies": [ { "id": 637733, "author_name": "Soren", "author_url": "", "post_date": "2019-10-01T08:14:47.253000", "content": "Thanks for the link!
\nIt seems to me that if the purpose of the model is to be applicable in the \"hyperacute\" setting - that is when the initial diagnosis is made then it is counterproductive to have craniectomy+drains in the dataset. Clearly presence of these will correlate with hemorrhage and become focus points for the models - but they wont be present in the actual data where this model could be useful. It's possible I am wrong but this is how it seems to me at the moment.
Totally agree with you. I think we might get into an issue here, where the best models might not be the most useful ones in real-life.\nI trying to account for this (at least a bit) by using some simple segmentation approach (as in one of my kernels) to extract only the patient's brain. But I have no idea yet, whether the approach actually works and what it does to Deep Learning models.
", "votes": 0, "replies": [] }, { "id": 637844, "author_name": "Soren", "author_url": "", "post_date": "2019-10-01T09:55:02.997000", "content": "Good luck! The high number of studies poses a challenge to masking etc. With 19000 studies just 1 minute of preprocessing is 2 weeks! How do you manage that?
", "votes": 0, "replies": [] }, { "id": 637915, "author_name": "srs", "author_url": "", "post_date": "2019-10-01T11:18:11.920000", "content": "I used something similar as in here (shameless self-plug), the kernel has unfortunately still some issues. Background separation, which is basically given by the windowing parameters, whole filling and dilation etc. just takes a second. It's definitely not perfect, that's why I need to test it out. It took about 12h to process all slices (not using anything parallel).
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "637703": "I'm not a medical expert and have close to non-medical expertise. However, in neuroscience research there are quite often discussions about what time-ranges to apply when talking about stroke. Especially, how to deal with different time-windows between neuro-psychological assessment and CT or MRI recordings. \nAnd according to this random source ([http://casemed.case.edu/clerkships/neurology/Web%20Neurorad/MultHemorrhageTypes.htm](http://casemed.case.edu/clerkships/neurology/Web%20Neurorad/MultHemorrhageTypes.htm)), the classification of acute is quite arbitrary and means up to 4 days after the hemorrhage. So I think your intuition might be right, that there are some CTs which were taken after initial treatment.", "637249": "Hi,\n I am confused about the overview description of this as acute hemorrhage prediction. Many cases have craniectomys and drains, which I would have thought indicated treatment and thereby a follow up scans? \nThanks!" } }