All of the notebooks seem to remove class 14. I have also seen the discussions about how most of the radiologists only report class 14. But, I don't understand why removing class 14 would improve the models' performance
", "messages": [ { "id": 1189264, "postDate": "2021-02-06T21:11:50.453Z", "content": "All of the notebooks seem to remove class 14. I have also seen the discussions about how most of the radiologists only report class 14. But, I don't understand why removing class 14 would improve the models' performance
", "rawMarkdown": "All of the notebooks seem to remove class 14. I have also seen the discussions about how most of the radiologists only report class 14. But, I don't understand why removing class 14 would improve the models' performance", "votes": 4 }, { "id": 1191724, "postDate": "2021-02-08T16:21:41.117Z", "content": "So removing class 14 seem to be the easiest way to deal with the images where there are no annotations. .
\nBut I think there is something more to this and I think some of those class 14 could have real 'No findings' while others may just been wrongly annotated or just no annotated at all. Check out a discussion thread I started about it.
\nhttps://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/217886
I think some of those 'No findings' needs to be included in training data.
\nHope this helps.
Ah I understand your point and ur I do agree with ur thread to some extent, thanks for ur response :)
", "rawMarkdown": "Ah I understand your point and ur I do agree with ur thread to some extent, thanks for ur response :)" } ] }, { "id": 1190134, "postDate": "2021-02-07T14:06:25.023Z", "content": "Maybe you can conduct two experiments, one for training on abnormal images only, one for training on both of the normal and abnormal images, and see what happens? Looking forward to your reply if you have done this.
", "rawMarkdown": "Maybe you can conduct two experiments, one for training on abnormal images only, one for training on both of the normal and abnormal images, and see what happens? Looking forward to your reply if you have done this.", "replies": [ { "id": 1190660, "postDate": "2021-02-07T21:58:41.737Z", "content": "I think your response could have been more helpful to be honest :)
", "rawMarkdown": "I think your response could have been more helpful to be honest :)", "votes": 1 }, { "id": 1193064, "postDate": "2021-02-09T12:41:59.027Z", "content": "Actually, I have conducted this experiment, and the result is that it may be better to train on abnormal images only. Training on both normal and abnormal images improves the CV. Unfortunately, It doesn't result in an improvement on PB.
", "rawMarkdown": "Actually, I have conducted this experiment, and the result is that it may be better to train on abnormal images only. Training on both normal and abnormal images improves the CV. Unfortunately, It doesn't result in an improvement on PB.", "votes": 3 } ] } ], "comments": [ { "id": 1191724, "author_name": "sourabhsc", "author_url": "", "post_date": "2021-02-08T16:21:41.117000", "content": "So removing class 14 seem to be the easiest way to deal with the images where there are no annotations. .
\nBut I think there is something more to this and I think some of those class 14 could have real 'No findings' while others may just been wrongly annotated or just no annotated at all. Check out a discussion thread I started about it.
\nhttps://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/217886
I think some of those 'No findings' needs to be included in training data.
\nHope this helps.
Ah I understand your point and ur I do agree with ur thread to some extent, thanks for ur response :)
", "votes": 0, "replies": [] } ] }, { "id": 1190134, "author_name": "ZehuiGong", "author_url": "", "post_date": "2021-02-07T14:06:25.023000", "content": "Maybe you can conduct two experiments, one for training on abnormal images only, one for training on both of the normal and abnormal images, and see what happens? Looking forward to your reply if you have done this.
", "votes": 0, "replies": [ { "id": 1190660, "author_name": "Mostafa Ibrahim", "author_url": "", "post_date": "2021-02-07T21:58:41.737000", "content": "I think your response could have been more helpful to be honest :)
", "votes": 1, "replies": [] }, { "id": 1193064, "author_name": "ZehuiGong", "author_url": "", "post_date": "2021-02-09T12:41:59.027000", "content": "Actually, I have conducted this experiment, and the result is that it may be better to train on abnormal images only. Training on both normal and abnormal images improves the CV. Unfortunately, It doesn't result in an improvement on PB.
", "votes": 3, "replies": [] } ] } ], "raw_markdown_by_id": { "1189264": "All of the notebooks seem to remove class 14. I have also seen the discussions about how most of the radiologists only report class 14. But, I don't understand why removing class 14 would improve the models' performance", "1191724": "So removing class 14 seem to be the easiest way to deal with the images where there are no annotations. .\n\n But I think there is something more to this and I think some of those class 14 could have real 'No findings' while others may just been wrongly annotated or just no annotated at all. Check out a discussion thread I started about it. \nhttps://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/discussion/217886 \n\nI think some of those 'No findings' needs to be included in training data. \nHope this helps.", "1190134": "Maybe you can conduct two experiments, one for training on abnormal images only, one for training on both of the normal and abnormal images, and see what happens? Looking forward to your reply if you have done this." } }