![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4433335%2F40595811209e58920a30728f49d2e940%2Fplot.png?generation=1716383631811098&alt=media\")
Each patient in this competition has 2-6 scans with each scan ranging from 5 to 192 slices.
\nWhat would be some approaches to handle this greatly varying number of scans and slices?
\nAs this interesting competition just started any idea's and remarks are welcome!
\n| # Scans | \nCount | \n
|---|---|
| 2 | \n3 | \n
| 3 | \n1632 | \n
| 4 | \n309 | \n
| 5 | \n30 | \n
| 6 | \n1 | \n
Each patient in this competition has 2-6 scans with each scan ranging from 5 to 192 slices.
\nWhat would be some approaches to handle this greatly varying number of scans and slices?
\nAs this interesting competition just started any idea's and remarks are welcome!
\n| # Scans | \nCount | \n
|---|---|
| 2 | \n3 | \n
| 3 | \n1632 | \n
| 4 | \n309 | \n
| 5 | \n30 | \n
| 6 | \n1 | \n
Neuroradiologist here. I plan to write a post about this, but the gist is: the scans are either a sagittal or a transversal/axial view of the body (the sagittal being along the long axis and the axial (mostly) perpendicular plane to this axis).
\nThe whole overview of all the levels/lumbar spine are only on the sagittal planes present and the perpendicular (axial ones) slices and depict only some or rarely all the of the levels affected. Usually either not all levels, or by overlapping volumes the same level.
\nAssuming that all the sequences were aquired within the same aquisition, since the dicom images usually contain the position of the slices, one could co-register them to a common volume, and then see, where the x,y points of the axial images are situated on the sagittal counterparts… and use only the sagittal images for a training (omitting all the axial ones).
\nBy the inference, you would do this backwards… like identify the level and severity on the sagittal plane and then assign the axial images that are on that exact position of the volume, but let's solve this on another day =)
\nTthis is just scratching the surface, but let's dive deeper, ill post here the link when i'm ready.
", "rawMarkdown": "Neuroradiologist here. I plan to write a post about this, but the gist is: the scans are either a sagittal or a transversal/axial view of the body (the sagittal being along the long axis and the axial (mostly) perpendicular plane to this axis).\n\nThe whole overview of **all** the levels/lumbar spine are only on the sagittal planes present and the perpendicular (axial ones) slices and depict only some or rarely all the of the levels affected. Usually either not all levels, or by overlapping volumes the same level.\n\nAssuming that all the sequences were aquired within the same aquisition, since the dicom images usually contain the position of the slices, one could co-register them to a common volume, and then see, where the x,y points of the axial images are situated on the sagittal counterparts... and use only the sagittal images for a training (omitting all the axial ones).\n\nBy the inference, you would do this backwards... like identify the level and severity on the sagittal plane and then assign the axial images that are on that exact position of the volume, but let's solve this on another day =)\n\nTthis is just scratching the surface, but let's dive deeper, ill post here the link when i'm ready.", "votes": null }, { "id": "2831357", "postDate": "05/23/2024 16:25:38", "content": "I'm also new to this data format so I looked at old competitions and saw that most solutions were doing something along the lines of: input data (variable number of slices of shape (n, m)) -> input to 2d cnn like efficientnet to extract features from all slices (output shape: (num_slices, embedding_size)) -> finally input to sequence model like LSTM, GRU, or transformer encoder for classification.
", "rawMarkdown": "I'm also new to this data format so I looked at old competitions and saw that most solutions were doing something along the lines of: input data (variable number of slices of shape (n, m)) -> input to 2d cnn like efficientnet to extract features from all slices (output shape: (num_slices, embedding_size)) -> finally input to sequence model like LSTM, GRU, or transformer encoder for classification.", "votes": null } ], "comments": [ { "id": 2831181, "author_name": "sandorkonya", "author_url": "", "post_date": "05/23/2024 14:47:23", "content": "Neuroradiologist here. I plan to write a post about this, but the gist is: the scans are either a sagittal or a transversal/axial view of the body (the sagittal being along the long axis and the axial (mostly) perpendicular plane to this axis).
\nThe whole overview of all the levels/lumbar spine are only on the sagittal planes present and the perpendicular (axial ones) slices and depict only some or rarely all the of the levels affected. Usually either not all levels, or by overlapping volumes the same level.
\nAssuming that all the sequences were aquired within the same aquisition, since the dicom images usually contain the position of the slices, one could co-register them to a common volume, and then see, where the x,y points of the axial images are situated on the sagittal counterparts… and use only the sagittal images for a training (omitting all the axial ones).
\nBy the inference, you would do this backwards… like identify the level and severity on the sagittal plane and then assign the axial images that are on that exact position of the volume, but let's solve this on another day =)
\nTthis is just scratching the surface, but let's dive deeper, ill post here the link when i'm ready.
", "votes": null, "replies": [] }, { "id": 2831357, "author_name": "snehalverma10", "author_url": "", "post_date": "05/23/2024 16:25:38", "content": "I'm also new to this data format so I looked at old competitions and saw that most solutions were doing something along the lines of: input data (variable number of slices of shape (n, m)) -> input to 2d cnn like efficientnet to extract features from all slices (output shape: (num_slices, embedding_size)) -> finally input to sequence model like LSTM, GRU, or transformer encoder for classification.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2829196": "Each patient in this competition has 2-6 scans with each scan ranging from 5 to 192 slices.\n\nWhat would be some approaches to handle this greatly varying number of scans and slices?\n\nAs this interesting competition just started any idea's and remarks are welcome!\n\n| **# Scans** | **Count** |\n| --- | --- |\n| 2 |\t3 |\n| 3 | 1632 |\n| 4 | 309 |\n| 5 | 30 |\n| 6 | 1 |\n\n\n", "2831181": "Neuroradiologist here. I plan to write a post about this, but the gist is: the scans are either a sagittal or a transversal/axial view of the body (the sagittal being along the long axis and the axial (mostly) perpendicular plane to this axis).\n\nThe whole overview of **all** the levels/lumbar spine are only on the sagittal planes present and the perpendicular (axial ones) slices and depict only some or rarely all the of the levels affected. Usually either not all levels, or by overlapping volumes the same level.\n\nAssuming that all the sequences were aquired within the same aquisition, since the dicom images usually contain the position of the slices, one could co-register them to a common volume, and then see, where the x,y points of the axial images are situated on the sagittal counterparts... and use only the sagittal images for a training (omitting all the axial ones).\n\nBy the inference, you would do this backwards... like identify the level and severity on the sagittal plane and then assign the axial images that are on that exact position of the volume, but let's solve this on another day =)\n\nTthis is just scratching the surface, but let's dive deeper, ill post here the link when i'm ready.", "2831357": "I'm also new to this data format so I looked at old competitions and saw that most solutions were doing something along the lines of: input data (variable number of slices of shape (n, m)) -> input to 2d cnn like efficientnet to extract features from all slices (output shape: (num_slices, embedding_size)) -> finally input to sequence model like LSTM, GRU, or transformer encoder for classification." }, "source": "meta" }