Has anyone able to extract useful features from the DICOM images, which might correlate to a decrease in lung capacity. There is a really good article by @sandorkonya which explains a lot of possible routes like lung segmentation, texture analysis, full lung volume capacity, BMI etc.
", "messages": [ { "id": "931196", "postDate": "07/16/2020 03:49:39", "content": "Has anyone able to extract useful features from the DICOM images, which might correlate to a decrease in lung capacity. There is a really good article by @sandorkonya which explains a lot of possible routes like lung segmentation, texture analysis, full lung volume capacity, BMI etc.
", "rawMarkdown": "Has anyone able to extract useful features from the DICOM images, which might correlate to a decrease in lung capacity. There is a really good article by @sandorkonya which explains a lot of possible routes like lung segmentation, texture analysis, full lung volume capacity, BMI etc.", "votes": null }, { "id": "931268", "postDate": "07/16/2020 05:13:22", "content": "I mainly tried transfer learning. The results weren't great because of heavy overfitting. Using a combination of the image mask could help marginally. In general, unless someone is very good with image based regression, the images are not advisable.
", "rawMarkdown": "I mainly tried transfer learning. The results weren't great because of heavy overfitting. Using a combination of the image mask could help marginally. In general, unless someone is very good with image based regression, the images are not advisable.", "votes": null }, { "id": "931319", "postDate": "07/16/2020 05:50:26", "content": "@vivekgopalramaswamy,\nthere is a 3rd \"domain expert's insight\" in preparation where i am going to write about the vessels, hearth and airways, their possible correlations to the decrease of lung capacity... however, i did not see anyone yet calculating the lung volume, normalised with the parameters as i suggested and noone predicted the other parameters yet.\nThese all i've mentioned before even started to talk about possible image based features! Maybe alone they don't mean much but in combination to other features... who knows =)
\n\nI strongly believe that many image based features fail simply due to the lack of normalization... and i don't mean now min-max, i mean that the anatomical variances are so huge! If you do not have any information about the localisation of the changes - normalized through the patients - then your model is not going to generalize well.
\n\nKeep up the good work!
", "rawMarkdown": "vivekgopalramaswamy,\nthere is a 3rd \"domain expert's insight\" in preparation where i am going to write about the vessels, hearth and airways, their possible correlations to the decrease of lung capacity... however, i did not see anyone yet calculating the lung volume, normalised with the parameters as i suggested and noone predicted the other parameters yet.\nThese all i've mentioned before even started to talk about possible image based features! Maybe alone they don't mean much but in combination to other features... who knows =)\n\nI strongly believe that many image based features fail simply due to the lack of normalization... and i don't mean now min-max, i mean that the anatomical variances are so huge! If you do not have any information about the localisation of the changes - normalized through the patients - then your model is not going to generalize well. \n\nKeep up the good work!", "votes": null }, { "id": "931331", "postDate": "07/16/2020 05:58:48", "content": "@eladwar ,
\n\nno question, this challenge is a regression... but maybe one could try creating distinct classes based on the end result (declining, non declining, fast declining), similarly as in this article.
\n\nMaybe one could find features that are more characteriistin in the distinct classes... or only present in one of them...
\n\nBut, maybe there is not feature that can predict the decrease, maybe (as suggested by some experts) the changes compared with the 6-Month followup are the definite answer.
\n\nKeep up the good work!
", "rawMarkdown": "eladwar ,\n\nno question, this challenge is a regression... but maybe one could try creating distinct classes based on the end result (declining, non declining, fast declining), similarly as [in this article](https://academic.oup.com/rheumatology/article/54/8/1464/1799372).\n\nMaybe one could find features that are more characteriistin in the distinct classes... or only present in one of them... \n\nBut, maybe there is not feature that can predict the decrease, maybe (as suggested by some experts) the changes compared with the 6-Month followup are the definite answer.\n\nKeep up the good work!", "votes": null }, { "id": "931337", "postDate": "07/16/2020 06:08:18", "content": "Yes you are correct, image normalization issues are barriers to get good image based features. I am relatively new to the medical domain, but given your insights , I am excited to experiment with detecting lung volume from the CT images. And Awesome, I am waiting for your third insight.
", "rawMarkdown": "Yes you are correct, image normalization issues are barriers to get good image based features. I am relatively new to the medical domain, but given your insights , I am excited to experiment with detecting lung volume from the CT images. And Awesome, I am waiting for your third insight.", "votes": null }, { "id": "931340", "postDate": "07/16/2020 06:11:56", "content": "Awesome, @eladwar . Yes that was a kind of question in my mind to, if the images can play a important role in this challenge. Given the insights by @sandorkonya , I feel definitely some of the features can help in predicting the progress.
", "rawMarkdown": "Awesome, @eladwar . Yes that was a kind of question in my mind to, if the images can play a important role in this challenge. Given the insights by @sandorkonya , I feel definitely some of the features can help in predicting the progress.", "votes": null } ], "comments": [ { "id": 931268, "author_name": "eladwar", "author_url": "", "post_date": "07/16/2020 05:13:22", "content": "I mainly tried transfer learning. The results weren't great because of heavy overfitting. Using a combination of the image mask could help marginally. In general, unless someone is very good with image based regression, the images are not advisable.
", "votes": null, "replies": [ { "id": 931331, "author_name": "sandorkonya", "author_url": "", "post_date": "07/16/2020 05:58:48", "content": "@eladwar ,
\n\nno question, this challenge is a regression... but maybe one could try creating distinct classes based on the end result (declining, non declining, fast declining), similarly as in this article.
\n\nMaybe one could find features that are more characteriistin in the distinct classes... or only present in one of them...
\n\nBut, maybe there is not feature that can predict the decrease, maybe (as suggested by some experts) the changes compared with the 6-Month followup are the definite answer.
\n\nKeep up the good work!
", "votes": null, "replies": [] }, { "id": 931340, "author_name": "vivekgopalramaswamy", "author_url": "", "post_date": "07/16/2020 06:11:56", "content": "Awesome, @eladwar . Yes that was a kind of question in my mind to, if the images can play a important role in this challenge. Given the insights by @sandorkonya , I feel definitely some of the features can help in predicting the progress.
", "votes": null, "replies": [] } ] }, { "id": 931319, "author_name": "sandorkonya", "author_url": "", "post_date": "07/16/2020 05:50:26", "content": "@vivekgopalramaswamy,\nthere is a 3rd \"domain expert's insight\" in preparation where i am going to write about the vessels, hearth and airways, their possible correlations to the decrease of lung capacity... however, i did not see anyone yet calculating the lung volume, normalised with the parameters as i suggested and noone predicted the other parameters yet.\nThese all i've mentioned before even started to talk about possible image based features! Maybe alone they don't mean much but in combination to other features... who knows =)
\n\nI strongly believe that many image based features fail simply due to the lack of normalization... and i don't mean now min-max, i mean that the anatomical variances are so huge! If you do not have any information about the localisation of the changes - normalized through the patients - then your model is not going to generalize well.
\n\nKeep up the good work!
", "votes": null, "replies": [ { "id": 931337, "author_name": "vivekgopalramaswamy", "author_url": "", "post_date": "07/16/2020 06:08:18", "content": "Yes you are correct, image normalization issues are barriers to get good image based features. I am relatively new to the medical domain, but given your insights , I am excited to experiment with detecting lung volume from the CT images. And Awesome, I am waiting for your third insight.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "931196": "Has anyone able to extract useful features from the DICOM images, which might correlate to a decrease in lung capacity. There is a really good article by @sandorkonya which explains a lot of possible routes like lung segmentation, texture analysis, full lung volume capacity, BMI etc.", "931268": "I mainly tried transfer learning. The results weren't great because of heavy overfitting. Using a combination of the image mask could help marginally. In general, unless someone is very good with image based regression, the images are not advisable.", "931319": "vivekgopalramaswamy,\nthere is a 3rd \"domain expert's insight\" in preparation where i am going to write about the vessels, hearth and airways, their possible correlations to the decrease of lung capacity... however, i did not see anyone yet calculating the lung volume, normalised with the parameters as i suggested and noone predicted the other parameters yet.\nThese all i've mentioned before even started to talk about possible image based features! Maybe alone they don't mean much but in combination to other features... who knows =)\n\nI strongly believe that many image based features fail simply due to the lack of normalization... and i don't mean now min-max, i mean that the anatomical variances are so huge! If you do not have any information about the localisation of the changes - normalized through the patients - then your model is not going to generalize well. \n\nKeep up the good work!", "931331": "eladwar ,\n\nno question, this challenge is a regression... but maybe one could try creating distinct classes based on the end result (declining, non declining, fast declining), similarly as [in this article](https://academic.oup.com/rheumatology/article/54/8/1464/1799372).\n\nMaybe one could find features that are more characteriistin in the distinct classes... or only present in one of them... \n\nBut, maybe there is not feature that can predict the decrease, maybe (as suggested by some experts) the changes compared with the 6-Month followup are the definite answer.\n\nKeep up the good work!", "931337": "Yes you are correct, image normalization issues are barriers to get good image based features. I am relatively new to the medical domain, but given your insights , I am excited to experiment with detecting lung volume from the CT images. And Awesome, I am waiting for your third insight.", "931340": "Awesome, @eladwar . Yes that was a kind of question in my mind to, if the images can play a important role in this challenge. Given the insights by @sandorkonya , I feel definitely some of the features can help in predicting the progress." }, "source": "meta" }