Dear kagglers,
\nI am starting this topic to share some hard examples, I have stumbled upon.
\nIt made me think that preprocessing wound not be as straightforward as it seemed to be.
Artifacts: the very first 2 images.
\nCloseups: 3, 4, 6, 7, 8, 9, 10. - It might make sense to use random crop for augmentations!?
\nCircles on the images: 4, 9, 10, 11.
More lines and black square:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6259210%2F56f9f67c8ecfa27747537a9f0175e98b%2FScreenshot%20from%202022-12-05%2013-42-25.png?generation=1670265978595585&alt=media\")
White-like: I would assume extremely dense breast tissue.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6259210%2F55160814a6734a9e61b06ab0476c71e6%2FScreenshot%20from%202022-12-05%2016-22-23.png?generation=1670275398422817&alt=media\")
p.s. you are welcome to share your examples, I am going to continue my search.
", "messages": [ { "id": 2056019, "postDate": "2022-12-05T16:43:23.173Z", "content": "Dear kagglers,
\nI am starting this topic to share some hard examples, I have stumbled upon.
\nIt made me think that preprocessing wound not be as straightforward as it seemed to be.
Artifacts: the very first 2 images.
\nCloseups: 3, 4, 6, 7, 8, 9, 10. - It might make sense to use random crop for augmentations!?
\nCircles on the images: 4, 9, 10, 11.
More lines and black square:
\n
White-like: I would assume extremely dense breast tissue.
\n
p.s. you are welcome to share your examples, I am going to continue my search.
", "rawMarkdown": "Dear kagglers, \n\nI am starting this topic to share some hard examples, I have stumbled upon.\nIt made me think that preprocessing wound not be as straightforward as it seemed to be.\n\n\n\n**Artifacts:** the very first 2 images.\n**Closeups:** 3, 4, 6, 7, 8, 9, 10. - It might make sense to use random crop for augmentations!?\n**Circles on the images:** 4, 9, 10, 11.\n\n**More lines and black square:**\n\n\n\n**White-like**: I would assume extremely dense breast tissue.\n\n\n\np.s. you are welcome to share your examples, I am going to continue my search.", "votes": 19 }, { "id": 2056103, "postDate": "2022-12-05T18:26:50.100Z", "content": "The 1st, 2nd, and 6th images have compression paddles, these are not typical in screening mammograms and are likely erroneous in the dataset. The other images where the entire breast is not covered in the image is typically due to large breasts where the mammogram has to be taken in several 'tiles' to capture the whole breast tissue.
", "rawMarkdown": "The 1st, 2nd, and 6th images have compression paddles, these are not typical in screening mammograms and are likely erroneous in the dataset. The other images where the entire breast is not covered in the image is typically due to large breasts where the mammogram has to be taken in several 'tiles' to capture the whole breast tissue.", "votes": 8 }, { "id": 2057220, "postDate": "2022-12-06T21:17:37.043Z", "content": "This one includes artifact too: 1147_597771506
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F698363%2Fede91b470d6cddbe58522646a6898f4c%2F1147_597771506.png?generation=1670361399142063&alt=media\")
That is a pacemaker. It is implanted in the skin overlying the breast.
", "rawMarkdown": "That is a pacemaker. It is implanted in the skin overlying the breast.", "votes": 7 } ] }, { "id": 2057266, "postDate": "2022-12-06T22:59:37.897Z", "content": "Most of the \"white\" images are based on the gray-scale used. They are not generally \"dense\". The next-to-last one is dense. The second \"white\" image has a breast implant.
", "rawMarkdown": "Most of the \"white\" images are based on the gray-scale used. They are not generally \"dense\". The next-to-last one is dense. The second \"white\" image has a breast implant.", "votes": 4 }, { "id": 2056045, "postDate": "2022-12-05T17:14:48.887Z", "content": "The circles are stickers placed on the patient. Usually to note moles.
", "rawMarkdown": "The circles are stickers placed on the patient. Usually to note moles.", "votes": 4 }, { "id": 2059522, "postDate": "2022-12-09T00:59:34.857Z", "content": "\"I am … hard examples\",
\non a side note:
\ni suggest one should just learn a model first, then check the model prediction.
\nfrom the prediction results, decide if you want to add these rare objects/artifacts to augmentation or just choose to ignore them if they are not cause of poor performances.
\"Hard examples\" are identified via their prediction scores.
", "rawMarkdown": "\"I am ... hard examples\", \n\non a side note:\n1. the model see samples in feature space. So \"visually outliers\" may not be outliers at all if the model can differentiate it. \n2. i am more interested where these outlier sample is in an TSNE plot for a learned model.\n3. a learned model will have activated feature for \"some leaned image characteristic\" and usually zero for background.\nif the artifacts or foreign objects are not common, they probably will give zero values in predictions.\n\ni suggest one should just learn a model first, then check the model prediction.\nfrom the prediction results, decide if you want to add these rare objects/artifacts to augmentation or just choose to ignore them if they are not cause of poor performances.\n\n\"Hard examples\" are identified via their prediction scores.", "votes": 1, "replies": [ { "id": 2059596, "postDate": "2022-12-09T04:17:22.813Z", "rawMarkdown": "", "isDeleted": true }, { "id": 2059620, "postDate": "2022-12-09T04:41:55.913Z", "content": "no. by checking the CAM activation map, there is no false positive due to detection of foreign tag objects
", "rawMarkdown": "no. by checking the CAM activation map, there is no false positive due to detection of foreign tag objects", "votes": 2 } ] }, { "id": 2059510, "postDate": "2022-12-09T00:10:28.947Z", "content": "nice work keep it up
", "rawMarkdown": "nice work keep it up\n", "votes": 1 } ], "comments": [ { "id": 2056103, "author_name": "Hari T.", "author_url": "", "post_date": "2022-12-05T18:26:50.100000", "content": "The 1st, 2nd, and 6th images have compression paddles, these are not typical in screening mammograms and are likely erroneous in the dataset. The other images where the entire breast is not covered in the image is typically due to large breasts where the mammogram has to be taken in several 'tiles' to capture the whole breast tissue.
", "votes": 8, "replies": [] }, { "id": 2057220, "author_name": "MPWARE", "author_url": "", "post_date": "2022-12-06T21:17:37.043000", "content": "This one includes artifact too: 1147_597771506
\n
That is a pacemaker. It is implanted in the skin overlying the breast.
", "votes": 7, "replies": [] } ] }, { "id": 2057266, "author_name": "quadcore/Richard Epstein", "author_url": "", "post_date": "2022-12-06T22:59:37.897000", "content": "Most of the \"white\" images are based on the gray-scale used. They are not generally \"dense\". The next-to-last one is dense. The second \"white\" image has a breast implant.
", "votes": 4, "replies": [] }, { "id": 2056045, "author_name": "quadcore/Richard Epstein", "author_url": "", "post_date": "2022-12-05T17:14:48.887000", "content": "The circles are stickers placed on the patient. Usually to note moles.
", "votes": 4, "replies": [] }, { "id": 2059522, "author_name": "hengck23", "author_url": "", "post_date": "2022-12-09T00:59:34.857000", "content": "\"I am … hard examples\",
\non a side note:
\ni suggest one should just learn a model first, then check the model prediction.
\nfrom the prediction results, decide if you want to add these rare objects/artifacts to augmentation or just choose to ignore them if they are not cause of poor performances.
\"Hard examples\" are identified via their prediction scores.
", "votes": 1, "replies": [ { "id": 2059596, "author_name": "", "author_url": "", "post_date": "2022-12-09T04:17:22.813000", "content": "", "votes": 0, "replies": [] }, { "id": 2059620, "author_name": "hengck23", "author_url": "", "post_date": "2022-12-09T04:41:55.913000", "content": "no. by checking the CAM activation map, there is no false positive due to detection of foreign tag objects
", "votes": 2, "replies": [] } ] }, { "id": 2059510, "author_name": "myo gyi(myo min htet)", "author_url": "", "post_date": "2022-12-09T00:10:28.947000", "content": "nice work keep it up
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "2056019": "Dear kagglers, \n\nI am starting this topic to share some hard examples, I have stumbled upon.\nIt made me think that preprocessing wound not be as straightforward as it seemed to be.\n\n\n\n**Artifacts:** the very first 2 images.\n**Closeups:** 3, 4, 6, 7, 8, 9, 10. - It might make sense to use random crop for augmentations!?\n**Circles on the images:** 4, 9, 10, 11.\n\n**More lines and black square:**\n\n\n\n**White-like**: I would assume extremely dense breast tissue.\n\n\n\np.s. you are welcome to share your examples, I am going to continue my search.", "2056103": "The 1st, 2nd, and 6th images have compression paddles, these are not typical in screening mammograms and are likely erroneous in the dataset. The other images where the entire breast is not covered in the image is typically due to large breasts where the mammogram has to be taken in several 'tiles' to capture the whole breast tissue.", "2057220": "This one includes artifact too: 1147_597771506\n", "2057266": "Most of the \"white\" images are based on the gray-scale used. They are not generally \"dense\". The next-to-last one is dense. The second \"white\" image has a breast implant.", "2056045": "The circles are stickers placed on the patient. Usually to note moles.", "2059522": "\"I am ... hard examples\", \n\non a side note:\n1. the model see samples in feature space. So \"visually outliers\" may not be outliers at all if the model can differentiate it. \n2. i am more interested where these outlier sample is in an TSNE plot for a learned model.\n3. a learned model will have activated feature for \"some leaned image characteristic\" and usually zero for background.\nif the artifacts or foreign objects are not common, they probably will give zero values in predictions.\n\ni suggest one should just learn a model first, then check the model prediction.\nfrom the prediction results, decide if you want to add these rare objects/artifacts to augmentation or just choose to ignore them if they are not cause of poor performances.\n\n\"Hard examples\" are identified via their prediction scores.", "2059510": "nice work keep it up\n" } }