Most of the notebooks are using just the csv files and not using DICOM file at all. Despite that they are getting results closer to the highest rank on leaderboard.
\n @pradyut23 has already asked a related question(https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/182955).
\nThere must be rationale why organisers provided the CT scans???????
Isnt it rationale to think that adding Image features in a smart way, will improve the score ???
\nThank you, if you are answering this question. And forgive if this a very begineer's question.
Most of the notebooks are using just the csv files and not using DICOM file at all. Despite that they are getting results closer to the highest rank on leaderboard.
\n @pradyut23 has already asked a related question(https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/182955).
\nThere must be rationale why organisers provided the CT scans???????
Isnt it rationale to think that adding Image features in a smart way, will improve the score ???
\nThank you, if you are answering this question. And forgive if this a very begineer's question.
The organizers hypothesized that the CT scans contain information that can allow a ML model to predict a patient's FVC in the future (or make existing predictions more accurate). It may well be that this hypothesis is wrong. Or, competitors just haven't found a model that can extract that information.
", "rawMarkdown": "The organizers hypothesized that the CT scans contain information that can allow a ML model to predict a patient's FVC in the future (or make existing predictions more accurate). It may well be that this hypothesis is wrong. Or, competitors just haven't found a model that can extract that information.", "votes": null }, { "id": "1020955", "postDate": "09/21/2020 14:32:45", "content": "From the contest Overview Description it seems they are implying that even experienced Radiologists may not be able to recognize sufficient info from CT images to reliably predict the progression in individual cases. If this is so, then we can likely expect that useful features from the CT scans will be quite difficult to find. Nevertheless, I suspect that in order for winners' results to actually be useful, CT scan features will be key.
", "rawMarkdown": "From the contest Overview Description it seems they are implying that even experienced Radiologists may not be able to recognize sufficient info from CT images to reliably predict the progression in individual cases. If this is so, then we can likely expect that useful features from the CT scans will be quite difficult to find. Nevertheless, I suspect that in order for winners' results to actually be useful, CT scan features will be key.", "votes": null }, { "id": "1021343", "postDate": "09/21/2020 19:10:43", "content": "You should be aware that the LB is highly misleading. We are dealing with a small number of test data samples to begin with, and the LB is calculated on a much smaller subset of these. Trust your local cross-validation score and not the leaderboard.
", "rawMarkdown": "You should be aware that the LB is highly misleading. We are dealing with a small number of test data samples to begin with, and the LB is calculated on a much smaller subset of these. Trust your local cross-validation score and not the leaderboard.", "votes": null }, { "id": "1021926", "postDate": "09/22/2020 08:05:06", "content": "@samlin20 @r2j1d2 If we had to use CNNs, how can CNNs be used for time series image data?
", "rawMarkdown": "samlin20 @r2j1d2 If we had to use CNNs, how can CNNs be used for time series image data?", "votes": null }, { "id": "1022108", "postDate": "09/22/2020 10:14:07", "content": "The image features have not exactly helped most competitors ( at least those reasonably active on the forums ) and neither have they helped us. Our model actually performs somewhat better without image feats than with.
", "rawMarkdown": "The image features have not exactly helped most competitors ( at least those reasonably active on the forums ) and neither have they helped us. Our model actually performs somewhat better without image feats than with.", "votes": null }, { "id": "1022805", "postDate": "09/22/2020 18:20:45", "content": "@nxrprime Thank you for your insights. What are your thoughts on why Images aren't helpful in this competition?
", "rawMarkdown": "nxrprime Thank you for your insights. What are your thoughts on why Images aren't helpful in this competition?", "votes": null }, { "id": "1022941", "postDate": "09/22/2020 20:40:33", "content": "Those image data are not time series, my friend.
", "rawMarkdown": "Those image data are not time series, my friend.", "votes": null } ], "comments": [ { "id": 1020955, "author_name": "r2j1d2", "author_url": "", "post_date": "09/21/2020 14:32:45", "content": "From the contest Overview Description it seems they are implying that even experienced Radiologists may not be able to recognize sufficient info from CT images to reliably predict the progression in individual cases. If this is so, then we can likely expect that useful features from the CT scans will be quite difficult to find. Nevertheless, I suspect that in order for winners' results to actually be useful, CT scan features will be key.
", "votes": null, "replies": [] }, { "id": 1021343, "author_name": "bigironsphere", "author_url": "", "post_date": "09/21/2020 19:10:43", "content": "You should be aware that the LB is highly misleading. We are dealing with a small number of test data samples to begin with, and the LB is calculated on a much smaller subset of these. Trust your local cross-validation score and not the leaderboard.
", "votes": null, "replies": [] }, { "id": 1022108, "author_name": "nxrprime", "author_url": "", "post_date": "09/22/2020 10:14:07", "content": "The image features have not exactly helped most competitors ( at least those reasonably active on the forums ) and neither have they helped us. Our model actually performs somewhat better without image feats than with.
", "votes": null, "replies": [ { "id": 1022805, "author_name": "amritpal333", "author_url": "", "post_date": "09/22/2020 18:20:45", "content": "@nxrprime Thank you for your insights. What are your thoughts on why Images aren't helpful in this competition?
", "votes": null, "replies": [] } ] }, { "id": 1020885, "author_name": "samlin20", "author_url": "", "post_date": "09/21/2020 13:44:03", "content": "The organizers hypothesized that the CT scans contain information that can allow a ML model to predict a patient's FVC in the future (or make existing predictions more accurate). It may well be that this hypothesis is wrong. Or, competitors just haven't found a model that can extract that information.
", "votes": null, "replies": [ { "id": 1021926, "author_name": "amritpal333", "author_url": "", "post_date": "09/22/2020 08:05:06", "content": "@samlin20 @r2j1d2 If we had to use CNNs, how can CNNs be used for time series image data?
", "votes": null, "replies": [] }, { "id": 1022941, "author_name": "samlin20", "author_url": "", "post_date": "09/22/2020 20:40:33", "content": "Those image data are not time series, my friend.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1019257": "Most of the notebooks are using just the csv files and not using DICOM file at all. Despite that they are getting results closer to the highest rank on leaderboard.\n @pradyut23 has already asked a related question(https://www.kaggle.com/c/osic-pulmonary-fibrosis-progression/discussion/182955). \nThere must be rationale why organisers provided the CT scans???????\n\nIsnt it rationale to think that adding Image features in a smart way, will improve the score ??? \nThank you, if you are answering this question. And forgive if this a very begineer's question.", "1020885": "The organizers hypothesized that the CT scans contain information that can allow a ML model to predict a patient's FVC in the future (or make existing predictions more accurate). It may well be that this hypothesis is wrong. Or, competitors just haven't found a model that can extract that information.", "1020955": "From the contest Overview Description it seems they are implying that even experienced Radiologists may not be able to recognize sufficient info from CT images to reliably predict the progression in individual cases. If this is so, then we can likely expect that useful features from the CT scans will be quite difficult to find. Nevertheless, I suspect that in order for winners' results to actually be useful, CT scan features will be key.", "1021343": "You should be aware that the LB is highly misleading. We are dealing with a small number of test data samples to begin with, and the LB is calculated on a much smaller subset of these. Trust your local cross-validation score and not the leaderboard.", "1021926": "samlin20 @r2j1d2 If we had to use CNNs, how can CNNs be used for time series image data?", "1022108": "The image features have not exactly helped most competitors ( at least those reasonably active on the forums ) and neither have they helped us. Our model actually performs somewhat better without image feats than with.", "1022805": "nxrprime Thank you for your insights. What are your thoughts on why Images aren't helpful in this competition?", "1022941": "Those image data are not time series, my friend." }, "source": "meta" }