My plan was to label the images with three possible conditions (sagitalt1/t2_model exmpl) ('Spinal Canal Stenosis', 'Right Neural Foraminal Narrowing', 'Left Neural Foraminal Narrowing') across five spinal levels ('L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1'). Each label includes a vector of six values: [confidence, severity_0, severity_1, severity_2, x, y].
\nI was thinking about creating a 3x5x6 tensor label to organize these labels, but I'm unsure about handling the scenario where one picture can have multiple labels. Additionally, I'm not clear on the appropriate dimensionality for the labels in this case. Am I moving in the right direction with this approach?
\nTo add, I'm considering training two backbones for localization and classification (one for sagital T1/T2 and another for axial T2). The idea is to use the last layer to create sequences of embeddings for some rnn architecture, but I'm not entirely sure how to implement this.
\nThis is my first competition in medical computer vision and working with \"3D\" data, so any insights or suggestions on how to effectively organize and implement this labeling scheme would be greatly appreciated!
", "messages": [ { "id": "2865651", "postDate": "06/10/2024 22:11:33", "content": "My plan was to label the images with three possible conditions (sagitalt1/t2_model exmpl) ('Spinal Canal Stenosis', 'Right Neural Foraminal Narrowing', 'Left Neural Foraminal Narrowing') across five spinal levels ('L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1'). Each label includes a vector of six values: [confidence, severity_0, severity_1, severity_2, x, y].
\nI was thinking about creating a 3x5x6 tensor label to organize these labels, but I'm unsure about handling the scenario where one picture can have multiple labels. Additionally, I'm not clear on the appropriate dimensionality for the labels in this case. Am I moving in the right direction with this approach?
\nTo add, I'm considering training two backbones for localization and classification (one for sagital T1/T2 and another for axial T2). The idea is to use the last layer to create sequences of embeddings for some rnn architecture, but I'm not entirely sure how to implement this.
\nThis is my first competition in medical computer vision and working with \"3D\" data, so any insights or suggestions on how to effectively organize and implement this labeling scheme would be greatly appreciated!
", "rawMarkdown": "My plan was to label the images with three possible conditions (sagitalt1/t2_model exmpl) ('Spinal Canal Stenosis', 'Right Neural Foraminal Narrowing', 'Left Neural Foraminal Narrowing') across five spinal levels ('L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1'). Each label includes a vector of six values: [confidence, severity_0, severity_1, severity_2, x, y].\n\nI was thinking about creating a 3x5x6 tensor label to organize these labels, but I'm unsure about handling the scenario where one picture can have multiple labels. Additionally, I'm not clear on the appropriate dimensionality for the labels in this case. Am I moving in the right direction with this approach?\n\nTo add, I'm considering training two backbones for localization and classification (one for sagital T1/T2 and another for axial T2). The idea is to use the last layer to create sequences of embeddings for some rnn architecture, but I'm not entirely sure how to implement this.\n\nThis is my first competition in medical computer vision and working with \"3D\" data, so any insights or suggestions on how to effectively organize and implement this labeling scheme would be greatly appreciated!", "votes": null }, { "id": "2874259", "postDate": "06/16/2024 06:23:46", "content": "I was thinking about creating a 3x5x6 tensor label to organize these labels, but I'm unsure about handling the scenario where one picture can have multiple labels
\nSay you have an image, named 'image1' which has two labels present in it. Let's say 'label1' and 'label2'. And we know this with the help of 'train_label_coordinates.csv'
\narr = []
\nNow, for the image1 and label1 add the 356 data to the array . for image1 and label2 add the 356 data to the array.
It will look something like
\nfor image in images:
\n----for label in image:
\n-------# combine label and image
\n-------# Add data to arr
Did you get it? Has this problem been solved?
", "rawMarkdown": "I was thinking about creating a 3x5x6 tensor label to organize these labels, but I'm unsure about handling the scenario where one picture can have multiple labels\n\nSay you have an image, named 'image1' which has two labels present in it. Let's say 'label1' and 'label2'. And we know this with the help of 'train_label_coordinates.csv'\narr = []\nNow, for the image1 and label1 add the 3*5*6 data to the array . for image1 and label2 add the 3*5*6 data to the array. \n\nIt will look something like\nfor image in images:\n----for label in image:\n-------# combine label and image \n-------# Add data to arr\n\nDid you get it? Has this problem been solved?", "votes": null }, { "id": "2874263", "postDate": "06/16/2024 06:28:28", "content": "Additionally, I'm not clear on the appropriate dimensionality for the labels in this case. Am I moving in the right direction with this approach?
\nTo add, I'm considering training two backbones for localization and classification (one for sagital T1/T2 and another for axial T2). The idea is to use the last layer to create sequences of embeddings for some rnn architecture, but I'm not entirely sure how to implement this
\nTake help from the publicly available notebooks in the code section. People have tried very similar things to you.
Which notebooks specifically? @devsya
", "rawMarkdown": "Which notebooks specifically? @devsya", "votes": null } ], "comments": [ { "id": 2874259, "author_name": "devsya", "author_url": "", "post_date": "06/16/2024 06:23:46", "content": "I was thinking about creating a 3x5x6 tensor label to organize these labels, but I'm unsure about handling the scenario where one picture can have multiple labels
\nSay you have an image, named 'image1' which has two labels present in it. Let's say 'label1' and 'label2'. And we know this with the help of 'train_label_coordinates.csv'
\narr = []
\nNow, for the image1 and label1 add the 356 data to the array . for image1 and label2 add the 356 data to the array.
It will look something like
\nfor image in images:
\n----for label in image:
\n-------# combine label and image
\n-------# Add data to arr
Did you get it? Has this problem been solved?
", "votes": null, "replies": [] }, { "id": 2874263, "author_name": "devsya", "author_url": "", "post_date": "06/16/2024 06:28:28", "content": "Additionally, I'm not clear on the appropriate dimensionality for the labels in this case. Am I moving in the right direction with this approach?
\nTo add, I'm considering training two backbones for localization and classification (one for sagital T1/T2 and another for axial T2). The idea is to use the last layer to create sequences of embeddings for some rnn architecture, but I'm not entirely sure how to implement this
\nTake help from the publicly available notebooks in the code section. People have tried very similar things to you.
Which notebooks specifically? @devsya
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2865651": "My plan was to label the images with three possible conditions (sagitalt1/t2_model exmpl) ('Spinal Canal Stenosis', 'Right Neural Foraminal Narrowing', 'Left Neural Foraminal Narrowing') across five spinal levels ('L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1'). Each label includes a vector of six values: [confidence, severity_0, severity_1, severity_2, x, y].\n\nI was thinking about creating a 3x5x6 tensor label to organize these labels, but I'm unsure about handling the scenario where one picture can have multiple labels. Additionally, I'm not clear on the appropriate dimensionality for the labels in this case. Am I moving in the right direction with this approach?\n\nTo add, I'm considering training two backbones for localization and classification (one for sagital T1/T2 and another for axial T2). The idea is to use the last layer to create sequences of embeddings for some rnn architecture, but I'm not entirely sure how to implement this.\n\nThis is my first competition in medical computer vision and working with \"3D\" data, so any insights or suggestions on how to effectively organize and implement this labeling scheme would be greatly appreciated!", "2874259": "I was thinking about creating a 3x5x6 tensor label to organize these labels, but I'm unsure about handling the scenario where one picture can have multiple labels\n\nSay you have an image, named 'image1' which has two labels present in it. Let's say 'label1' and 'label2'. And we know this with the help of 'train_label_coordinates.csv'\narr = []\nNow, for the image1 and label1 add the 3*5*6 data to the array . for image1 and label2 add the 3*5*6 data to the array. \n\nIt will look something like\nfor image in images:\n----for label in image:\n-------# combine label and image \n-------# Add data to arr\n\nDid you get it? Has this problem been solved?", "2874263": "Additionally, I'm not clear on the appropriate dimensionality for the labels in this case. Am I moving in the right direction with this approach?\n\nTo add, I'm considering training two backbones for localization and classification (one for sagital T1/T2 and another for axial T2). The idea is to use the last layer to create sequences of embeddings for some rnn architecture, but I'm not entirely sure how to implement this\n\nTake help from the publicly available notebooks in the code section. People have tried very similar things to you.", "2875584": "Which notebooks specifically? @devsya" }, "source": "meta" }