I didn't have enough time to do this.
\nI'll share my ideas with you. I hope this will help someone else.
I thought about doing a linear regression of the slope of the predictions of the training data based on a Laplace distribution, such as Hoover regression, to account for outliers, rather than a simple linear regression. I think there is a problem with sample data.
\nFrom a medical point of view, it is more reasonable to base the regression on the %FVC (=percent). However, since the method of creating %FVC differs from country to country, I was unsure to what extent it should be used.
\nI wanted to create a multiple-input, multiple-output model and set the slope and intercept of the predictions in the output layer. Due to the large amount of data, we had difficulty with the training model on the server and in a local environment.
\nWe did not understand how to get the variance from quantile regression. I feel that over-learning will proceed by manipulating the error function.
\nIt was a good practice to analyze CT images and medical data. I would like to study further with you to improve the accuracy.
", "messages": [ { "id": "1036138", "postDate": "10/03/2020 13:15:42", "content": "I didn't have enough time to do this.
\nI'll share my ideas with you. I hope this will help someone else.
I thought about doing a linear regression of the slope of the predictions of the training data based on a Laplace distribution, such as Hoover regression, to account for outliers, rather than a simple linear regression. I think there is a problem with sample data.
\nFrom a medical point of view, it is more reasonable to base the regression on the %FVC (=percent). However, since the method of creating %FVC differs from country to country, I was unsure to what extent it should be used.
\nI wanted to create a multiple-input, multiple-output model and set the slope and intercept of the predictions in the output layer. Due to the large amount of data, we had difficulty with the training model on the server and in a local environment.
\nWe did not understand how to get the variance from quantile regression. I feel that over-learning will proceed by manipulating the error function.
\nIt was a good practice to analyze CT images and medical data. I would like to study further with you to improve the accuracy.
", "rawMarkdown": "I didn't have enough time to do this.\nI'll share my ideas with you. I hope this will help someone else.\n\nI thought about doing a linear regression of the slope of the predictions of the training data based on a Laplace distribution, such as Hoover regression, to account for outliers, rather than a simple linear regression. I think there is a problem with sample data.\n\nFrom a medical point of view, it is more reasonable to base the regression on the %FVC (=percent). However, since the method of creating %FVC differs from country to country, I was unsure to what extent it should be used.\n\nI wanted to create a multiple-input, multiple-output model and set the slope and intercept of the predictions in the output layer. Due to the large amount of data, we had difficulty with the training model on the server and in a local environment.\n\nWe did not understand how to get the variance from quantile regression. I feel that over-learning will proceed by manipulating the error function.\n\nIt was a good practice to analyze CT images and medical data. I would like to study further with you to improve the accuracy.", "votes": null }, { "id": "1037136", "postDate": "10/04/2020 16:24:49", "content": "Nice idea shared.
", "rawMarkdown": "Nice idea shared.", "votes": null }, { "id": "1037164", "postDate": "10/04/2020 16:45:03", "content": "The challenge is that outliers are actually kind of a normal outcome. There is a lot of uncertainty on how the pacient will progress, the FVC can decay exponentially, stay the same or in many cases even improve. For instance, I tryed to remove extreme outliers manually and the model got worst. I think that finding some method that accounts for outliers but also for the 'normal' behaviour is the key for this comp
", "rawMarkdown": "The challenge is that outliers are actually kind of a normal outcome. There is a lot of uncertainty on how the pacient will progress, the FVC can decay exponentially, stay the same or in many cases even improve. For instance, I tryed to remove extreme outliers manually and the model got worst. I think that finding some method that accounts for outliers but also for the 'normal' behaviour is the key for this comp", "votes": null }, { "id": "1038020", "postDate": "10/05/2020 13:40:43", "content": "Thanks for your comment.
\nThank you for telling me. I will refer to your valuable attempts. I realized how difficult it is to work with medical data.
\nA useful method that can be applied to input as well as output: ….
\nI couldn't think of any other way to represent it other than with a linear model, which is a small sample size, so I couldn't come up with it in my capacity.
\nConsidering the background of pulmonary fibrosis, it can be broadly classified into Non-Specific Interstitial Pneumonia, which reacts with steroid-unresponsive idiopathic pulmonary fibrosis, unchanged normal lung and organicized pneumonia, but the application of deep learning to these classification problems does not lead to the usefulness of deep learning.
Thanks for your comment. I wish you good luck.
", "rawMarkdown": "Thanks for your comment. I wish you good luck.", "votes": null } ], "comments": [ { "id": 1037164, "author_name": "jsaguiar", "author_url": "", "post_date": "10/04/2020 16:45:03", "content": "The challenge is that outliers are actually kind of a normal outcome. There is a lot of uncertainty on how the pacient will progress, the FVC can decay exponentially, stay the same or in many cases even improve. For instance, I tryed to remove extreme outliers manually and the model got worst. I think that finding some method that accounts for outliers but also for the 'normal' behaviour is the key for this comp
", "votes": null, "replies": [ { "id": 1038020, "author_name": "shoheiyoshimoto", "author_url": "", "post_date": "10/05/2020 13:40:43", "content": "Thanks for your comment.
\nThank you for telling me. I will refer to your valuable attempts. I realized how difficult it is to work with medical data.
\nA useful method that can be applied to input as well as output: ….
\nI couldn't think of any other way to represent it other than with a linear model, which is a small sample size, so I couldn't come up with it in my capacity.
\nConsidering the background of pulmonary fibrosis, it can be broadly classified into Non-Specific Interstitial Pneumonia, which reacts with steroid-unresponsive idiopathic pulmonary fibrosis, unchanged normal lung and organicized pneumonia, but the application of deep learning to these classification problems does not lead to the usefulness of deep learning.
Nice idea shared.
", "votes": null, "replies": [ { "id": 1038022, "author_name": "shoheiyoshimoto", "author_url": "", "post_date": "10/05/2020 13:42:04", "content": "Thanks for your comment. I wish you good luck.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1036138": "I didn't have enough time to do this.\nI'll share my ideas with you. I hope this will help someone else.\n\nI thought about doing a linear regression of the slope of the predictions of the training data based on a Laplace distribution, such as Hoover regression, to account for outliers, rather than a simple linear regression. I think there is a problem with sample data.\n\nFrom a medical point of view, it is more reasonable to base the regression on the %FVC (=percent). However, since the method of creating %FVC differs from country to country, I was unsure to what extent it should be used.\n\nI wanted to create a multiple-input, multiple-output model and set the slope and intercept of the predictions in the output layer. Due to the large amount of data, we had difficulty with the training model on the server and in a local environment.\n\nWe did not understand how to get the variance from quantile regression. I feel that over-learning will proceed by manipulating the error function.\n\nIt was a good practice to analyze CT images and medical data. I would like to study further with you to improve the accuracy.", "1037136": "Nice idea shared.", "1037164": "The challenge is that outliers are actually kind of a normal outcome. There is a lot of uncertainty on how the pacient will progress, the FVC can decay exponentially, stay the same or in many cases even improve. For instance, I tryed to remove extreme outliers manually and the model got worst. I think that finding some method that accounts for outliers but also for the 'normal' behaviour is the key for this comp", "1038020": "Thanks for your comment. \nThank you for telling me. I will refer to your valuable attempts. I realized how difficult it is to work with medical data.\nA useful method that can be applied to input as well as output: ....\nI couldn't think of any other way to represent it other than with a linear model, which is a small sample size, so I couldn't come up with it in my capacity.\nConsidering the background of pulmonary fibrosis, it can be broadly classified into Non-Specific Interstitial Pneumonia, which reacts with steroid-unresponsive idiopathic pulmonary fibrosis, unchanged normal lung and organicized pneumonia, but the application of deep learning to these classification problems does not lead to the usefulness of deep learning.", "1038022": "Thanks for your comment. I wish you good luck." }, "source": "meta" }