I recently came across this DeepSpine paper and realized its approach is highly relevant to the problem statement in this competition: multi-input, multi-task, multi-class classification. I’ve started replicating the work and have reached the halfway mark.
\nBig thanks to the following people whose work served as valuable references and helped me build upon it:
\n@abhinavsuri @anoukstein @vaillant @hengck23
Notebook 1 Developed a segmentation model using U-Net with a pretrained ResNet34 backbone to localize vertebrae, later leveraging them to calculate the centroids of intervertebral discs.
Notebook 2 The trained model was then used to prepare the dataset as described in the DeepSpine paper. The output consists of sagittal crops for each level along with corresponding axial slices for each patient (study ID). ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5048844%2F14caaa02fb50c901e7baf7562415e84b%2FScreenshot%202024-08-18%20201541.png?generation=1723992373714825&alt=media\")
Next Steps:
\nI plan to use these datasets to classify multi-tasks for each level, as demonstrated below:
\n ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5048844%2F011413630166acabc6e540e20439b95c%2FScreenshot%202024-08-18%20202136.png?generation=1723992749055894&alt=media\")
I’d love to get feedback on the notebooks. If anyone is interested in collaborating, especially those with expertise in PyTorch, feel free to reach out!
", "messages": [ { "id": "2963284", "postDate": "08/18/2024 14:58:37", "content": "I recently came across this DeepSpine paper and realized its approach is highly relevant to the problem statement in this competition: multi-input, multi-task, multi-class classification. I’ve started replicating the work and have reached the halfway mark.
\nBig thanks to the following people whose work served as valuable references and helped me build upon it:
\n@abhinavsuri @anoukstein @vaillant @hengck23
Notebook 1 Developed a segmentation model using U-Net with a pretrained ResNet34 backbone to localize vertebrae, later leveraging them to calculate the centroids of intervertebral discs.
Notebook 2 The trained model was then used to prepare the dataset as described in the DeepSpine paper. The output consists of sagittal crops for each level along with corresponding axial slices for each patient (study ID).
Next Steps:
\nI plan to use these datasets to classify multi-tasks for each level, as demonstrated below:
\n
I’d love to get feedback on the notebooks. If anyone is interested in collaborating, especially those with expertise in PyTorch, feel free to reach out!
", "rawMarkdown": "I recently came across this [DeepSpine](https://arxiv.org/abs/1807.10215) paper and realized its approach is highly relevant to the problem statement in this competition: multi-input, multi-task, multi-class classification. I’ve started replicating the work and have reached the halfway mark.\n\nBig thanks to the following people whose work served as valuable references and helped me build upon it:\n@abhinavsuri @anoukstein @vaillant @hengck23\n\n1. [Notebook 1](https://www.kaggle.com/code/rahulnakka/lumbar-vertebrae-segmentation-using-spider-datset) Developed a segmentation model using U-Net with a pretrained ResNet34 backbone to localize vertebrae, later leveraging them to calculate the centroids of intervertebral discs.\n\n2. [Notebook 2](https://www.kaggle.com/code/rahulnakka/deepspine-custom-dataset-preparation-process/notebook) The trained model was then used to prepare the dataset as described in the DeepSpine paper. The output consists of sagittal crops for each level along with corresponding axial slices for each patient (study ID). \n\n\nNext Steps:\nI plan to use these datasets to classify multi-tasks for each level, as demonstrated below:\n \n\nI’d love to get feedback on the notebooks. If anyone is interested in collaborating, especially those with expertise in PyTorch, feel free to reach out!", "votes": null }, { "id": "2968684", "postDate": "08/24/2024 06:24:24", "content": "This is a 2018 paper. I think the framework and methods used are relatively old…
", "rawMarkdown": "This is a 2018 paper. I think the framework and methods used are relatively old...", "votes": null }, { "id": "2977305", "postDate": "09/02/2024 17:54:45", "content": "Hey! I would like to try the same. I'll let you know on my progress. Please do share developments if you can.
", "rawMarkdown": "Hey! I would like to try the same. I'll let you know on my progress. Please do share developments if you can.", "votes": null }, { "id": "2978185", "postDate": "09/03/2024 15:38:41", "content": "sure!! I will share my results soon
", "rawMarkdown": "sure!! I will share my results soon", "votes": null } ], "comments": [ { "id": 2968684, "author_name": "machengyuan", "author_url": "", "post_date": "08/24/2024 06:24:24", "content": "This is a 2018 paper. I think the framework and methods used are relatively old…
", "votes": null, "replies": [] }, { "id": 2977305, "author_name": "arindamroy23", "author_url": "", "post_date": "09/02/2024 17:54:45", "content": "Hey! I would like to try the same. I'll let you know on my progress. Please do share developments if you can.
", "votes": null, "replies": [ { "id": 2978185, "author_name": "rahulnakka", "author_url": "", "post_date": "09/03/2024 15:38:41", "content": "sure!! I will share my results soon
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2963284": "I recently came across this [DeepSpine](https://arxiv.org/abs/1807.10215) paper and realized its approach is highly relevant to the problem statement in this competition: multi-input, multi-task, multi-class classification. I’ve started replicating the work and have reached the halfway mark.\n\nBig thanks to the following people whose work served as valuable references and helped me build upon it:\n@abhinavsuri @anoukstein @vaillant @hengck23\n\n1. [Notebook 1](https://www.kaggle.com/code/rahulnakka/lumbar-vertebrae-segmentation-using-spider-datset) Developed a segmentation model using U-Net with a pretrained ResNet34 backbone to localize vertebrae, later leveraging them to calculate the centroids of intervertebral discs.\n\n2. [Notebook 2](https://www.kaggle.com/code/rahulnakka/deepspine-custom-dataset-preparation-process/notebook) The trained model was then used to prepare the dataset as described in the DeepSpine paper. The output consists of sagittal crops for each level along with corresponding axial slices for each patient (study ID). \n\n\nNext Steps:\nI plan to use these datasets to classify multi-tasks for each level, as demonstrated below:\n \n\nI’d love to get feedback on the notebooks. If anyone is interested in collaborating, especially those with expertise in PyTorch, feel free to reach out!", "2968684": "This is a 2018 paper. I think the framework and methods used are relatively old...", "2977305": "Hey! I would like to try the same. I'll let you know on my progress. Please do share developments if you can.", "2978185": "sure!! I will share my results soon" }, "source": "meta" }