\nRadiologist R8 marked box (1687.0, 367.0, 2398.0, 1779.0) with both 'Infiltration' and 'ILD'
\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6203332%2F71db88c052e346039ac9c95194616a5e%2FScreen%20Shot%202021-01-07%20at%201.57.37.png?generation=1609979179861245&alt=media\")
Hi,
\nI encountered at least one occasion of a Radiologist annotating the exact same bounding box with two different classes (abnormalities).
\nI don't have medical background so I don't know if it's medically possible (different abnormalities in the same lung area).
How do we handle this?
\nCan the detection model learn different classes from the same bboxes?
Look at image_id: b065d3b4da92b8015b47224341661fd8
\nRadiologist R8 marked box (1687.0, 367.0, 2398.0, 1779.0) with both 'Infiltration' and 'ILD'
\n
Hi @amirsher
\nSome dense lesion area contains many labels, our radiologists created a box then assign many labels to that box.
Hi,
\nI encountered at least one occasion of a Radiologist annotating the exact same bounding box with two different classes (abnormalities).
\nI don't have medical background so I don't know if it's medically possible (different abnormalities in the same lung area).
How do we handle this?
\nCan the detection model learn different classes from the same bboxes?
Look at image_id: b065d3b4da92b8015b47224341661fd8
\nRadiologist R8 marked box (1687.0, 367.0, 2398.0, 1779.0) with both 'Infiltration' and 'ILD'
\n
Ideally if you consider a multi-label classification task togheter with localization of bboxes you could overcome this problem. Anyway in my experiments I'm not having very good results..
", "rawMarkdown": "Ideally if you consider a multi-label classification task togheter with localization of bboxes you could overcome this problem. Anyway in my experiments I'm not having very good results.." }, { "id": 1146291, "postDate": "2021-01-09T16:53:41.440Z", "content": "https://www.mayoclinic.org/diseases-conditions/interstitial-lung-disease/symptoms-causes/syc-20353108
\nILD -> Interstitial (in-tur-STISH-ul) lung disease describes a large group of disorders, most of which cause progressive scarring of lung tissue. The scarring associated with interstitial lung disease eventually affects your ability to breathe and get enough oxygen into your bloodstream.
Seems like ILD is a blanket term for a group of disorders. Maybe the radiologist classified it as ILD and later realised that \"Infiltration\" better describes the finding. This is just my guess.
", "rawMarkdown": "[https://www.mayoclinic.org/diseases-conditions/interstitial-lung-disease/symptoms-causes/syc-20353108](url)\nILD -> Interstitial (in-tur-STISH-ul) lung disease describes **a large group of disorders**, most of which cause progressive scarring of lung tissue. The scarring associated with interstitial lung disease eventually affects your ability to breathe and get enough oxygen into your bloodstream.\n\nSeems like ILD is a blanket term for a group of disorders. Maybe the radiologist classified it as ILD and later realised that \"Infiltration\" better describes the finding. This is just my guess.", "replies": [ { "id": 1146461, "postDate": "2021-01-09T19:38:07.097Z", "content": "So, is it a good practice for such cases of duplicate labels, to remove the ILD label from the training?
", "rawMarkdown": "So, is it a good practice for such cases of duplicate labels, to remove the ILD label from the training?" }, { "id": 1150201, "postDate": "2021-01-12T13:01:07.300Z", "content": "Might be a good idea since ILD is a more generic classification.
", "rawMarkdown": "Might be a good idea since ILD is a more generic classification." } ] } ], "comments": [ { "id": 1141975, "author_name": "DungNB", "author_url": "", "post_date": "2021-01-07T03:37:51.880000", "content": "Hi @amirsher
\nSome dense lesion area contains many labels, our radiologists created a box then assign many labels to that box.
Ideally if you consider a multi-label classification task togheter with localization of bboxes you could overcome this problem. Anyway in my experiments I'm not having very good results..
", "votes": 0, "replies": [] }, { "id": 1146291, "author_name": "Abhimanyu Dikshit", "author_url": "", "post_date": "2021-01-09T16:53:41.440000", "content": "https://www.mayoclinic.org/diseases-conditions/interstitial-lung-disease/symptoms-causes/syc-20353108
\nILD -> Interstitial (in-tur-STISH-ul) lung disease describes a large group of disorders, most of which cause progressive scarring of lung tissue. The scarring associated with interstitial lung disease eventually affects your ability to breathe and get enough oxygen into your bloodstream.
Seems like ILD is a blanket term for a group of disorders. Maybe the radiologist classified it as ILD and later realised that \"Infiltration\" better describes the finding. This is just my guess.
", "votes": 0, "replies": [ { "id": 1146461, "author_name": "Amir Sher", "author_url": "", "post_date": "2021-01-09T19:38:07.097000", "content": "So, is it a good practice for such cases of duplicate labels, to remove the ILD label from the training?
", "votes": 0, "replies": [] }, { "id": 1150201, "author_name": "Abhimanyu Dikshit", "author_url": "", "post_date": "2021-01-12T13:01:07.300000", "content": "Might be a good idea since ILD is a more generic classification.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1141975": "Hi @amirsher\nSome dense lesion area contains many labels, our radiologists created a box then assign many labels to that box.", "1141863": "Hi,\nI encountered at least one occasion of a Radiologist annotating the exact same bounding box with two different classes (abnormalities).\nI don't have medical background so I don't know if it's medically possible (different abnormalities in the same lung area).\n\nHow do we handle this? \nCan the detection model learn different classes from the same bboxes?\n\nLook at image_id: b065d3b4da92b8015b47224341661fd8\nRadiologist R8 marked box (1687.0, 367.0, 2398.0, 1779.0) with both 'Infiltration' and 'ILD'\n", "1148678": "Ideally if you consider a multi-label classification task togheter with localization of bboxes you could overcome this problem. Anyway in my experiments I'm not having very good results..", "1146291": "[https://www.mayoclinic.org/diseases-conditions/interstitial-lung-disease/symptoms-causes/syc-20353108](url)\nILD -> Interstitial (in-tur-STISH-ul) lung disease describes **a large group of disorders**, most of which cause progressive scarring of lung tissue. The scarring associated with interstitial lung disease eventually affects your ability to breathe and get enough oxygen into your bloodstream.\n\nSeems like ILD is a blanket term for a group of disorders. Maybe the radiologist classified it as ILD and later realised that \"Infiltration\" better describes the finding. This is just my guess." } }