{ "id": 486462, "title": "Related Studies Estimating Age from Head CT", "url": "/competitions/spr-head-ct-age-prediction-challenge/discussion/486462", "author_name": "patriot", "post_date": "2024-03-25T03:28:48.215000", "votes": 2, "comment_count": 6, "views": 0, "content": "

I would like to find out which parts of the CT should be focused on.

\n

https://pubmed.ncbi.nlm.nih.gov/33047535/
\nhttps://www.researchgate.net/publication/236610072_Age_estimation_In_Old_Individuals_by_CT_Scan_of_Skull
\nfrom bone?
\nhttps://www.mpg.de/19260712/determining-age-from-brain-scans
\nfrom White matter?
\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC7228157/
\nfrom Brain parenchymal fraction?

", "messages": [ { "id": 2715023, "postDate": "2024-03-25T08:18:37.023Z", "content": "

I believe the challenge is to let your model decide what is best to focus on. The 'Generate windowed PNG' notebook in the Code tab uses windowing of 40, 80 which implies a focus on the brain tissue. However, I think it is better to not focus on any part explicitly (apart from windowing) and let your model figure out which parts to focus on. It could focus on both the facial structure and the brain tissue - the features that determine age can vary from scan to scan, as some cases may be hard to judge on just a single feature derived from a single part of the CT scan.

", "rawMarkdown": "I believe the challenge is to let your model decide what is best to focus on. The 'Generate windowed PNG' notebook in the Code tab uses windowing of 40, 80 which implies a focus on the brain tissue. However, I think it is better to not focus on any part explicitly (apart from windowing) and let your model figure out which parts to focus on. It could focus on both the facial structure **and** the brain tissue - the features that determine age can vary from scan to scan, as some cases may be hard to judge on just a single feature derived from a single part of the CT scan.", "votes": 2, "replies": [ { "id": 2715052, "postDate": "2024-03-25T08:33:57.927Z", "content": "

https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/117480

\n

Stacking multiple windows to create a pseudo-color image or adding window optimization to learnable parameters may be possible options.

", "rawMarkdown": "https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/117480\n\nStacking multiple windows to create a pseudo-color image or adding window optimization to learnable parameters may be possible options.", "votes": 2, "replies": [ { "id": 2715060, "postDate": "2024-03-25T08:42:10.223Z", "content": "

That's a good point. This would help let the model decide which parts to focus on all by itself. Personally, I like the idea of stacking multiple windows, but this would increase computational cost and memory footprint by a lot. Although this might lead to better convergence, the cost outweighs the benefits for me. Nonetheless, it is worth trying.
\nAs for learnable windowing parameters: The same argument applies. It will likely help convergence, but at the cost of losing performance, especially during training. I preprocess the scans, which also applies the windowing. I then save the results to disk, so that I only need to process scans once - that helped speeding up training in my case.

", "rawMarkdown": "That's a good point. This would help let the model decide which parts to focus on all by itself. Personally, I like the idea of stacking multiple windows, but this would increase computational cost and memory footprint by a lot. Although this might lead to better convergence, the cost outweighs the benefits for me. Nonetheless, it is worth trying.\nAs for learnable windowing parameters: The same argument applies. It will likely help convergence, but at the cost of losing performance, especially during training. I preprocess the scans, which also applies the windowing. I then save the results to disk, so that I only need to process scans once - that helped speeding up training in my case.", "votes": 2 } ] }, { "id": 2720300, "postDate": "2024-03-28T08:26:23.040Z", "content": "

I think that having a reasonable understanding of why is important in that challenge. We have three main anatomical players distinguishable by signal range: skull, brain and cerebrospinal fluid. Skull undergoes changes with ageing. Frontal bones are getting thicker particularly in women. Texture changes in bones because of osteoporosis. Brain is getting smaller with more space in the ventricular system and thiner cortex and more CSF in the cortical gyri. Cerebrospinal fluid space is expanding both inside and outside. Lateral ventricles and gyri are expanding. If we can project the atlas into the data we should get some meaningful answer.

", "rawMarkdown": "I think that having a reasonable understanding of why is important in that challenge. We have three main anatomical players distinguishable by signal range: skull, brain and cerebrospinal fluid. Skull undergoes changes with ageing. Frontal bones are getting thicker particularly in women. Texture changes in bones because of osteoporosis. Brain is getting smaller with more space in the ventricular system and thiner cortex and more CSF in the cortical gyri. Cerebrospinal fluid space is expanding both inside and outside. Lateral ventricles and gyri are expanding. If we can project the atlas into the data we should get some meaningful answer.", "votes": 3, "replies": [ { "id": 2720460, "postDate": "2024-03-28T11:12:55.640Z", "content": "

Great input. I am not working in the medical field, so this information is really valuable. I will consider adding multiple windows after I'm done training my current model.

", "rawMarkdown": "Great input. I am not working in the medical field, so this information is really valuable. I will consider adding multiple windows after I'm done training my current model." }, { "id": 2723425, "postDate": "2024-03-30T07:52:37.093Z", "content": "

I am happy to help. The interdisciplinary character of this competition is really exciting. Hopefully there is more to come. I would be very interested in the degree of improvement once we add anatomical details.

", "rawMarkdown": "I am happy to help. The interdisciplinary character of this competition is really exciting. Hopefully there is more to come. I would be very interested in the degree of improvement once we add anatomical details." } ] } ] }, { "id": 2714714, "postDate": "2024-03-25T03:28:48.217Z", "content": "

I would like to find out which parts of the CT should be focused on.

\n

https://pubmed.ncbi.nlm.nih.gov/33047535/
\nhttps://www.researchgate.net/publication/236610072_Age_estimation_In_Old_Individuals_by_CT_Scan_of_Skull
\nfrom bone?
\nhttps://www.mpg.de/19260712/determining-age-from-brain-scans
\nfrom White matter?
\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC7228157/
\nfrom Brain parenchymal fraction?

", "rawMarkdown": "I would like to find out which parts of the CT should be focused on.\n\nhttps://pubmed.ncbi.nlm.nih.gov/33047535/ \nhttps://www.researchgate.net/publication/236610072_Age_estimation_In_Old_Individuals_by_CT_Scan_of_Skull\nfrom bone?\nhttps://www.mpg.de/19260712/determining-age-from-brain-scans\nfrom White matter?\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC7228157/\nfrom Brain parenchymal fraction?\n\n\n\n", "votes": 2 } ], "comments": [ { "id": 2715023, "author_name": "Skelp", "author_url": "", "post_date": "2024-03-25T08:18:37.023000", "content": "

I believe the challenge is to let your model decide what is best to focus on. The 'Generate windowed PNG' notebook in the Code tab uses windowing of 40, 80 which implies a focus on the brain tissue. However, I think it is better to not focus on any part explicitly (apart from windowing) and let your model figure out which parts to focus on. It could focus on both the facial structure and the brain tissue - the features that determine age can vary from scan to scan, as some cases may be hard to judge on just a single feature derived from a single part of the CT scan.

", "votes": 2, "replies": [ { "id": 2715052, "author_name": "patriot", "author_url": "", "post_date": "2024-03-25T08:33:57.927000", "content": "

https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/117480

\n

Stacking multiple windows to create a pseudo-color image or adding window optimization to learnable parameters may be possible options.

", "votes": 2, "replies": [ { "id": 2715060, "author_name": "Skelp", "author_url": "", "post_date": "2024-03-25T08:42:10.223000", "content": "

That's a good point. This would help let the model decide which parts to focus on all by itself. Personally, I like the idea of stacking multiple windows, but this would increase computational cost and memory footprint by a lot. Although this might lead to better convergence, the cost outweighs the benefits for me. Nonetheless, it is worth trying.
\nAs for learnable windowing parameters: The same argument applies. It will likely help convergence, but at the cost of losing performance, especially during training. I preprocess the scans, which also applies the windowing. I then save the results to disk, so that I only need to process scans once - that helped speeding up training in my case.

", "votes": 2, "replies": [] } ] }, { "id": 2720300, "author_name": "neuroguide", "author_url": "", "post_date": "2024-03-28T08:26:23.040000", "content": "

I think that having a reasonable understanding of why is important in that challenge. We have three main anatomical players distinguishable by signal range: skull, brain and cerebrospinal fluid. Skull undergoes changes with ageing. Frontal bones are getting thicker particularly in women. Texture changes in bones because of osteoporosis. Brain is getting smaller with more space in the ventricular system and thiner cortex and more CSF in the cortical gyri. Cerebrospinal fluid space is expanding both inside and outside. Lateral ventricles and gyri are expanding. If we can project the atlas into the data we should get some meaningful answer.

", "votes": 3, "replies": [ { "id": 2720460, "author_name": "Skelp", "author_url": "", "post_date": "2024-03-28T11:12:55.640000", "content": "

Great input. I am not working in the medical field, so this information is really valuable. I will consider adding multiple windows after I'm done training my current model.

", "votes": 0, "replies": [] }, { "id": 2723425, "author_name": "neuroguide", "author_url": "", "post_date": "2024-03-30T07:52:37.093000", "content": "

I am happy to help. The interdisciplinary character of this competition is really exciting. Hopefully there is more to come. I would be very interested in the degree of improvement once we add anatomical details.

", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2715023": "I believe the challenge is to let your model decide what is best to focus on. The 'Generate windowed PNG' notebook in the Code tab uses windowing of 40, 80 which implies a focus on the brain tissue. However, I think it is better to not focus on any part explicitly (apart from windowing) and let your model figure out which parts to focus on. It could focus on both the facial structure **and** the brain tissue - the features that determine age can vary from scan to scan, as some cases may be hard to judge on just a single feature derived from a single part of the CT scan.", "2714714": "I would like to find out which parts of the CT should be focused on.\n\nhttps://pubmed.ncbi.nlm.nih.gov/33047535/ \nhttps://www.researchgate.net/publication/236610072_Age_estimation_In_Old_Individuals_by_CT_Scan_of_Skull\nfrom bone?\nhttps://www.mpg.de/19260712/determining-age-from-brain-scans\nfrom White matter?\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC7228157/\nfrom Brain parenchymal fraction?\n\n\n\n" } }