{ "id": 355738, "title": "Noise in Segmentation Data", "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/355738", "author_name": "Anouk Stein, MD", "post_date": "2022-09-28T00:44:48.890000", "votes": 12, "comment_count": 3, "views": 0, "content": "

There are 4 cases with random nonzero pixels that prevent bounding box access via the function cv2.boundingRect(cv2.findNonZero(img)). These are the ones I found so far: '1.2.826.0.1.3680043.6125', '1.2.826.0.1.3680043.1542', '1.2.826.0.1.3680043.24891', '1.2.826.0.1.3680043.11827'

\n

Also '1.2.826.0.1.3680043.1363' has random labels for C5 and C6 on the first slice. There are likely other cases. I'm trying to find a simple way to denoise these.

", "messages": [ { "id": 1959113, "postDate": "2022-09-28T00:44:48.890Z", "content": "

There are 4 cases with random nonzero pixels that prevent bounding box access via the function cv2.boundingRect(cv2.findNonZero(img)). These are the ones I found so far: '1.2.826.0.1.3680043.6125', '1.2.826.0.1.3680043.1542', '1.2.826.0.1.3680043.24891', '1.2.826.0.1.3680043.11827'

\n

Also '1.2.826.0.1.3680043.1363' has random labels for C5 and C6 on the first slice. There are likely other cases. I'm trying to find a simple way to denoise these.

", "rawMarkdown": "There are 4 cases with random nonzero pixels that prevent bounding box access via the function cv2.boundingRect(cv2.findNonZero(img)). These are the ones I found so far: '1.2.826.0.1.3680043.6125', '1.2.826.0.1.3680043.1542', '1.2.826.0.1.3680043.24891', '1.2.826.0.1.3680043.11827'\n\nAlso '1.2.826.0.1.3680043.1363' has random labels for C5 and C6 on the first slice. There are likely other cases. I'm trying to find a simple way to denoise these.\n ", "votes": 12 }, { "id": 1960707, "postDate": "2022-09-28T18:06:29.373Z", "content": "

You could get rid of single pixel lines and other small features with something like this,
\nwhere img is a 512 x 512 array of segmentation values (0…)

\n
img2 = img.copy\nimg2[img2 > 0] = 1    # make the image 0 or 1 only\n#filter with 3x3 array of 1's  \nimg2=scipy.signal.convolve2d(img2.astype(np.float32), np.ones((3,3)), mode='same', boundary='symm')\n#remove anything that is 4 or less\nimg[img2 < 5] = 0  # remove single pixel lines and other small features\n
\n

For example, this would remove the middle 1 from a 3x3 section that looked like:

\n
010\n110\n010\n
\n

but would not remove the middle of:

\n
010\n111\n010\n
\n

As a side note, there are values in the segmentation file that go higher than 7 (I forget the exact maximum), so watch for 8s and 9s!

", "rawMarkdown": "You could get rid of single pixel lines and other small features with something like this,\nwhere img is a 512 x 512 array of segmentation values (0...)\n```\nimg2 = img.copy\nimg2[img2 > 0] = 1 # make the image 0 or 1 only\n#filter with 3x3 array of 1's \nimg2=scipy.signal.convolve2d(img2.astype(np.float32), np.ones((3,3)), mode='same', boundary='symm')\n#remove anything that is 4 or less\nimg[img2 < 5] = 0 # remove single pixel lines and other small features\n\n```\nFor example, this would remove the middle 1 from a 3x3 section that looked like:\n```\n010\n110\n010\n```\nbut would not remove the middle of:\n```\n010\n111\n010\n```\nAs a side note, there are values in the segmentation file that go higher than 7 (I forget the exact maximum), so watch for 8s and 9s!", "votes": 2, "replies": [ { "id": 1960931, "postDate": "2022-09-28T22:00:47.927Z", "content": "

Nice! Thank you!

", "rawMarkdown": "Nice! Thank you!" } ] }, { "id": 1960078, "postDate": "2022-09-28T13:22:06.403Z", "content": "

I've found the same. It's a lot more than 4 cases though. I'd estimate about 40% of the segmentations have small errors in them.

\n

\"8884.png\"

\n

\"mask2.png\"

", "rawMarkdown": "I've found the same. It's a lot more than 4 cases though. I'd estimate about 40% of the segmentations have small errors in them. \n\n[![8884.png](https://i.postimg.cc/6pPX8vfh/8884.png)](https://postimg.cc/QFcPPVPB)\n\n[![mask2.png](https://i.postimg.cc/CxrFbtSk/mask2.png)](https://postimg.cc/xXHD26k1)", "votes": 2 } ], "comments": [ { "id": 1960707, "author_name": "SolverWorld", "author_url": "", "post_date": "2022-09-28T18:06:29.373000", "content": "

You could get rid of single pixel lines and other small features with something like this,
\nwhere img is a 512 x 512 array of segmentation values (0…)

\n
img2 = img.copy\nimg2[img2 > 0] = 1    # make the image 0 or 1 only\n#filter with 3x3 array of 1's  \nimg2=scipy.signal.convolve2d(img2.astype(np.float32), np.ones((3,3)), mode='same', boundary='symm')\n#remove anything that is 4 or less\nimg[img2 < 5] = 0  # remove single pixel lines and other small features\n
\n

For example, this would remove the middle 1 from a 3x3 section that looked like:

\n
010\n110\n010\n
\n

but would not remove the middle of:

\n
010\n111\n010\n
\n

As a side note, there are values in the segmentation file that go higher than 7 (I forget the exact maximum), so watch for 8s and 9s!

", "votes": 2, "replies": [ { "id": 1960931, "author_name": "Anouk Stein, MD", "author_url": "", "post_date": "2022-09-28T22:00:47.927000", "content": "

Nice! Thank you!

", "votes": 0, "replies": [] } ] }, { "id": 1960078, "author_name": "Samuel Cortinhas", "author_url": "", "post_date": "2022-09-28T13:22:06.403000", "content": "

I've found the same. It's a lot more than 4 cases though. I'd estimate about 40% of the segmentations have small errors in them.

\n

\"8884.png\"

\n

\"mask2.png\"

", "votes": 2, "replies": [] } ], "raw_markdown_by_id": { "1959113": "There are 4 cases with random nonzero pixels that prevent bounding box access via the function cv2.boundingRect(cv2.findNonZero(img)). These are the ones I found so far: '1.2.826.0.1.3680043.6125', '1.2.826.0.1.3680043.1542', '1.2.826.0.1.3680043.24891', '1.2.826.0.1.3680043.11827'\n\nAlso '1.2.826.0.1.3680043.1363' has random labels for C5 and C6 on the first slice. There are likely other cases. I'm trying to find a simple way to denoise these.\n ", "1960707": "You could get rid of single pixel lines and other small features with something like this,\nwhere img is a 512 x 512 array of segmentation values (0...)\n```\nimg2 = img.copy\nimg2[img2 > 0] = 1 # make the image 0 or 1 only\n#filter with 3x3 array of 1's \nimg2=scipy.signal.convolve2d(img2.astype(np.float32), np.ones((3,3)), mode='same', boundary='symm')\n#remove anything that is 4 or less\nimg[img2 < 5] = 0 # remove single pixel lines and other small features\n\n```\nFor example, this would remove the middle 1 from a 3x3 section that looked like:\n```\n010\n110\n010\n```\nbut would not remove the middle of:\n```\n010\n111\n010\n```\nAs a side note, there are values in the segmentation file that go higher than 7 (I forget the exact maximum), so watch for 8s and 9s!", "1960078": "I've found the same. It's a lot more than 4 cases though. I'd estimate about 40% of the segmentations have small errors in them. \n\n[![8884.png](https://i.postimg.cc/6pPX8vfh/8884.png)](https://postimg.cc/QFcPPVPB)\n\n[![mask2.png](https://i.postimg.cc/CxrFbtSk/mask2.png)](https://postimg.cc/xXHD26k1)" } }