{ "id": 181815, "title": "predicting a number of labels, at both the image and study level", "url": "/competitions/rsna-str-pulmonary-embolism-detection/discussion/181815", "author_name": "roshaan zafar", "post_date": "2020-09-10T08:44:57.159000", "votes": 2, "comment_count": 4, "views": 0, "content": "

Hi
\nI am beginner here so I wanted to know what is meant by at image level and study level? Does it mean the output we should have both contains the combination of image labels and labels from csv file? or what?
\nkindly please guide me, any sort of help will be beneficial
\nthank you

", "messages": [ { "id": 1005125, "postDate": "2020-09-10T09:10:58.503Z", "content": "

Each study has multiple images. We have to predict labels for images as well as studies.

\n", "rawMarkdown": "Each **study** has multiple **images**. We have to predict labels for images as well as studies. \n\n- Image level - predict for each image i.e SOPInstanceUID\nLabels to predict : *pe_present_on_image*\n\n- Study level - predict for each study i.e StudyInstanceUID\nLabels to predict : *negative_exam_for_pe* , *indeterminate*, *rv_lv_ratio_gte_1*, *rv_lv_ratio_lt_1*, *leftsided_pe*, *rightsided_pe*, *central_pe*, *chronic_pe*, *acute_and_chronic_pe*", "votes": 3, "replies": [ { "id": 1009519, "postDate": "2020-09-14T04:25:39.870Z", "content": "

and to submit do we need to map those predicted labels into one as in submission sample csv it has only one column belong to LABELS.
\nin simple words how many predicted labels we need to submit by submission csv?

", "rawMarkdown": "and to submit do we need to map those predicted labels into one as in submission sample csv it has only one column belong to LABELS.\nin simple words how many predicted labels we need to submit by submission csv?", "replies": [ { "id": 1009901, "postDate": "2020-09-14T10:59:03.547Z", "content": "

For each image - predict one label (PE on image)
\nFor each study - predict 9 labels (one per row):
\n negative_exam_for_pe ,
\n indeterminate,
\n rv_lv_ratio_gte_1,
\n rv_lv_ratio_lt_1,
\n leftsided_pe,
\n rightsided_pe,
\n central_pe,
\n chronic_pe,
\n acute_and_chronic_pe

\n

The \"Public\" submission file has 152,703 rows:
\n650 studies x 9 labels = 5,850 rows
\n146,853 images
\n146853+5850=152,703

\n

The real submission file (that we can never see) has 1517 studies (posted by Competition host), so is about 2.33 times the size.

\n

Typically the order of a submission file doesn't matter, but I haven't tested that in this competition. Typically you must have all rows with the correct id,label; even if you guess zero for some rows as a placeholder.

\n

For now, rather that waste processing time, I am running my tests on about 40000 hidden test images and leaving the rest of the submission file as a default value. That way I can see if I am making progress (and if my code actually works), without wasting GPU time.

\n

-Rich

", "rawMarkdown": "For each image - predict one label (PE on image)\nFor each study - predict 9 labels (one per row):\n negative_exam_for_pe , \n indeterminate,\n rv_lv_ratio_gte_1, \n rv_lv_ratio_lt_1, \n leftsided_pe, \n rightsided_pe, \n central_pe, \n chronic_pe, \n acute_and_chronic_pe\n\nThe \"Public\" submission file has 152,703 rows:\n650 studies x 9 labels = 5,850 rows\n146,853 images\n146853+5850=152,703\n\nThe real submission file (that we can never see) has 1517 studies (posted by Competition host), so is about 2.33 times the size.\n\nTypically the order of a submission file doesn't matter, but I haven't tested that in this competition. Typically you must have all rows with the correct id,label; even if you guess zero for some rows as a placeholder.\n\nFor now, rather that waste processing time, I am running my tests on about 40000 hidden test images and leaving the rest of the submission file as a default value. That way I can see if I am making progress (and if my code actually works), without wasting GPU time.\n\n-Rich\n", "votes": 3 }, { "id": 1011452, "postDate": "2020-09-15T13:28:58.830Z", "content": "

@richardepstein great explanation. It's easy to understand prediction at the slice/image level, we will give the image and its respective labels to model for training and then use the trained model for prediction at the image level but I am really confused about prediction at the study level since each study has multiple images and targets associated with it so how we will feed it to model as a single input/target row?

", "rawMarkdown": "@richardepstein great explanation. It's easy to understand prediction at the slice/image level, we will give the image and its respective labels to model for training and then use the trained model for prediction at the image level but I am really confused about prediction at the study level since each study has multiple images and targets associated with it so how we will feed it to model as a single input/target row?" } ] } ] }, { "id": 1005089, "postDate": "2020-09-10T08:44:57.160Z", "content": "

Hi
\nI am beginner here so I wanted to know what is meant by at image level and study level? Does it mean the output we should have both contains the combination of image labels and labels from csv file? or what?
\nkindly please guide me, any sort of help will be beneficial
\nthank you

", "rawMarkdown": "Hi \nI am beginner here so I wanted to know what is meant by at image level and study level? Does it mean the output we should have both contains the combination of image labels and labels from csv file? or what?\nkindly please guide me, any sort of help will be beneficial\nthank you", "votes": 2 } ], "comments": [ { "id": 1005125, "author_name": "Jebastin Nadar", "author_url": "", "post_date": "2020-09-10T09:10:58.503000", "content": "

Each study has multiple images. We have to predict labels for images as well as studies.

\n", "votes": 3, "replies": [ { "id": 1009519, "author_name": "manpreet singh", "author_url": "", "post_date": "2020-09-14T04:25:39.870000", "content": "

and to submit do we need to map those predicted labels into one as in submission sample csv it has only one column belong to LABELS.
\nin simple words how many predicted labels we need to submit by submission csv?

", "votes": 0, "replies": [ { "id": 1009901, "author_name": "quadcore/Richard Epstein", "author_url": "", "post_date": "2020-09-14T10:59:03.547000", "content": "

For each image - predict one label (PE on image)
\nFor each study - predict 9 labels (one per row):
\n negative_exam_for_pe ,
\n indeterminate,
\n rv_lv_ratio_gte_1,
\n rv_lv_ratio_lt_1,
\n leftsided_pe,
\n rightsided_pe,
\n central_pe,
\n chronic_pe,
\n acute_and_chronic_pe

\n

The \"Public\" submission file has 152,703 rows:
\n650 studies x 9 labels = 5,850 rows
\n146,853 images
\n146853+5850=152,703

\n

The real submission file (that we can never see) has 1517 studies (posted by Competition host), so is about 2.33 times the size.

\n

Typically the order of a submission file doesn't matter, but I haven't tested that in this competition. Typically you must have all rows with the correct id,label; even if you guess zero for some rows as a placeholder.

\n

For now, rather that waste processing time, I am running my tests on about 40000 hidden test images and leaving the rest of the submission file as a default value. That way I can see if I am making progress (and if my code actually works), without wasting GPU time.

\n

-Rich

", "votes": 3, "replies": [] }, { "id": 1011452, "author_name": "Abdur Rehman", "author_url": "", "post_date": "2020-09-15T13:28:58.830000", "content": "

@richardepstein great explanation. It's easy to understand prediction at the slice/image level, we will give the image and its respective labels to model for training and then use the trained model for prediction at the image level but I am really confused about prediction at the study level since each study has multiple images and targets associated with it so how we will feed it to model as a single input/target row?

", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "1005125": "Each **study** has multiple **images**. We have to predict labels for images as well as studies. \n\n- Image level - predict for each image i.e SOPInstanceUID\nLabels to predict : *pe_present_on_image*\n\n- Study level - predict for each study i.e StudyInstanceUID\nLabels to predict : *negative_exam_for_pe* , *indeterminate*, *rv_lv_ratio_gte_1*, *rv_lv_ratio_lt_1*, *leftsided_pe*, *rightsided_pe*, *central_pe*, *chronic_pe*, *acute_and_chronic_pe*", "1005089": "Hi \nI am beginner here so I wanted to know what is meant by at image level and study level? Does it mean the output we should have both contains the combination of image labels and labels from csv file? or what?\nkindly please guide me, any sort of help will be beneficial\nthank you" } }