What I want to do is to detect the heart and lungs.
\nThen crop the pictures of the heart, left and right lungs.
\nYou can use cropped pictures to detect diseases.
\nI'm not sure if this will be better.
\nThe advantage of this is that the original image size can be used for training when the GPU is not good enough.
\nRemove noise specifically to train the picture of the lungs and the picture of the heart.
\nWe hand-labeling 4394 pictures with diseases.
\nWhen labeling, we will try to include disease labeling in the picture.
\nThe picture of the heart contains aortic tumors and cardiac hypertrophy.
\nThe picture of the lungs is included in addition to other diseases.
\nHowever, some diseases that are too much beyond the lungs will be ignored.
\nThere will be 5 models at the end.
\nA model that predicts the disease in the entire picture.
\nPredict the model of lung and heart position.
\nUse cropped pictures to focus on predicting left lung disease.
\nUse cropped pictures to focus on predicting right lung disease.
\nUse cropped pictures to focus on predicting heart disease.
\nI want to know if this idea can improve performance but the GPU usage will be too large.
\nAnd everyone uses different methods.
\nIf you think this method is feasible, you can try it.
\nI am also trying.
\nIf you have improved your score with this method, share how much you improved.
\nFinally, thank my team members for accompanying me with manual labeling.
\nnotebook:
\nVinBigData-Detect the heart and lungs\n
dataset:
\nHeart and lung labels.
\nxml
\ntxt
\ncsv
Heart and lung 1024 jpg.
\nheart
\nlung_1
\nlung_2
What I want to do is to detect the heart and lungs.
\nThen crop the pictures of the heart, left and right lungs.
\nYou can use cropped pictures to detect diseases.
\nI'm not sure if this will be better.
\nThe advantage of this is that the original image size can be used for training when the GPU is not good enough.
\nRemove noise specifically to train the picture of the lungs and the picture of the heart.
\nWe hand-labeling 4394 pictures with diseases.
\nWhen labeling, we will try to include disease labeling in the picture.
\nThe picture of the heart contains aortic tumors and cardiac hypertrophy.
\nThe picture of the lungs is included in addition to other diseases.
\nHowever, some diseases that are too much beyond the lungs will be ignored.
\nThere will be 5 models at the end.
\nA model that predicts the disease in the entire picture.
\nPredict the model of lung and heart position.
\nUse cropped pictures to focus on predicting left lung disease.
\nUse cropped pictures to focus on predicting right lung disease.
\nUse cropped pictures to focus on predicting heart disease.
\nI want to know if this idea can improve performance but the GPU usage will be too large.
\nAnd everyone uses different methods.
\nIf you think this method is feasible, you can try it.
\nI am also trying.
\nIf you have improved your score with this method, share how much you improved.
\nFinally, thank my team members for accompanying me with manual labeling.
\nnotebook:
\nVinBigData-Detect the heart and lungs\n
dataset:
\nHeart and lung labels.
\nxml
\ntxt
\ncsv
Heart and lung 1024 jpg.
\nheart
\nlung_1
\nlung_2
Did I just read a poem?
", "rawMarkdown": "Did I just read a poem?", "votes": 6, "replies": [ { "id": 1232388, "postDate": "2021-03-09T17:59:45.640Z", "content": "It's a very lung' poem ;)
", "rawMarkdown": "It's a very lung' poem ;)", "votes": 4 }, { "id": 1233426, "postDate": "2021-03-10T11:51:27.053Z", "content": "Haha! It turns out that writing like this would be like writing a poem.😄
", "rawMarkdown": "Haha! It turns out that writing like this would be like writing a poem.😄" } ] }, { "id": 1246977, "postDate": "2021-03-21T09:48:29.900Z", "content": "there is a paper that uses segmentation of heart and lung as auxiliary loss for chest xray abnormalities classification. you can google about it (i forget the title)
", "rawMarkdown": "there is a paper that uses segmentation of heart and lung as auxiliary loss for chest xray abnormalities classification. you can google about it (i forget the title)", "votes": 1 }, { "id": 1233076, "postDate": "2021-03-10T05:51:31.813Z", "content": "@h053473666 Thanks for all the effort which went in to hand labelling .
", "rawMarkdown": "@h053473666 Thanks for all the effort which went in to hand labelling .", "votes": 1, "replies": [ { "id": 1238585, "postDate": "2021-03-15T05:45:43.740Z", "content": "no worries~
", "rawMarkdown": "no worries~", "votes": 1 } ] } ], "comments": [ { "id": 1232239, "author_name": "Ahmet Erdem", "author_url": "", "post_date": "2021-03-09T15:26:19.857000", "content": "Did I just read a poem?
", "votes": 6, "replies": [ { "id": 1232388, "author_name": "AmorfEvo", "author_url": "", "post_date": "2021-03-09T17:59:45.640000", "content": "It's a very lung' poem ;)
", "votes": 4, "replies": [] }, { "id": 1233426, "author_name": "Alien", "author_url": "", "post_date": "2021-03-10T11:51:27.053000", "content": "Haha! It turns out that writing like this would be like writing a poem.😄
", "votes": 0, "replies": [] } ] }, { "id": 1246977, "author_name": "hengck23", "author_url": "", "post_date": "2021-03-21T09:48:29.900000", "content": "there is a paper that uses segmentation of heart and lung as auxiliary loss for chest xray abnormalities classification. you can google about it (i forget the title)
", "votes": 1, "replies": [] }, { "id": 1233076, "author_name": "Tensor Girl", "author_url": "", "post_date": "2021-03-10T05:51:31.813000", "content": "@h053473666 Thanks for all the effort which went in to hand labelling .
", "votes": 1, "replies": [ { "id": 1238585, "author_name": "Cocoma", "author_url": "", "post_date": "2021-03-15T05:45:43.740000", "content": "no worries~
", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "1232154": "What I want to do is to detect the heart and lungs.\n\nThen crop the pictures of the heart, left and right lungs.\n\nYou can use cropped pictures to detect diseases.\n\nI'm not sure if this will be better.\n\nThe advantage of this is that the original image size can be used for training when the GPU is not good enough.\n\nRemove noise specifically to train the picture of the lungs and the picture of the heart.\n\nWe hand-labeling 4394 pictures with diseases.\n\nWhen labeling, we will try to include disease labeling in the picture.\n\nThe picture of the heart contains aortic tumors and cardiac hypertrophy.\n\nThe picture of the lungs is included in addition to other diseases.\n\nHowever, some diseases that are too much beyond the lungs will be ignored.\n\nThere will be 5 models at the end.\n\nA model that predicts the disease in the entire picture.\n\nPredict the model of lung and heart position.\n\nUse cropped pictures to focus on predicting left lung disease.\n\nUse cropped pictures to focus on predicting right lung disease.\n\nUse cropped pictures to focus on predicting heart disease.\n\nI want to know if this idea can improve performance but the GPU usage will be too large.\n\nAnd everyone uses different methods.\n\nIf you think this method is feasible, you can try it.\n\nI am also trying.\n\nIf you have improved your score with this method, share how much you improved.\n\nFinally, thank my team members for accompanying me with manual labeling.\n\nnotebook:\n[VinBigData-Detect the heart and lungs\n](https://www.kaggle.com/h053473666/vinbigdata-detect-the-heart-and-lungs)\n\ndataset:\nHeart and lung labels.\n[xml](https://www.kaggle.com/h053473666/xml-all-4394-vinbigdata)\n[txt](https://www.kaggle.com/h053473666/vinbigdata-txt-yolov5)\n[csv](https://www.kaggle.com/h053473666/vinbigdata-detect-the-heart-and-lungs-csv)\n\nHeart and lung 1024 jpg.\n[heart](https://www.kaggle.com/h053473666/heart1024)\n[lung_1](https://www.kaggle.com/h053473666/lung-1-800800jpg)\n[lung_2](https://www.kaggle.com/h053473666/lung-2-800800jpg)", "1232239": "Did I just read a poem?", "1246977": "there is a paper that uses segmentation of heart and lung as auxiliary loss for chest xray abnormalities classification. you can google about it (i forget the title)", "1233076": "@h053473666 Thanks for all the effort which went in to hand labelling ." } }