\n[1] TransMIL: Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification
\n
![\"https://i.ibb.co/M53dHF2/Selection-092.png\"](\"https://i.ibb.co/M53dHF2/Selection-092.png\")
\npretrain model (trained using SSL) for replacing resnet50:
reference:
\n[1] TransMIL: Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification
\n
\npretrain model (trained using SSL) for replacing resnet50:
reference:
\n[1] TransMIL: Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification
\n
\npretrain model (trained using SSL) for replacing resnet50:
Awesome, I am the author of TransMIL. Thank you for following our work~
", "rawMarkdown": "Awesome, I am the author of TransMIL. Thank you for following our work~", "votes": 15 }, { "id": 1864328, "postDate": "2022-07-21T00:48:57.940Z", "content": "there are many good videos from standford medical AI.
\nthis one is about transformer MIL for WSI
\nMedAI #36: Weakly supervised tumor detection in whole slide image analysis | Bin Li
\nhttps://www.youtube.com/watch?v=ZPe94q8wxPQ
I have been wrapping my head around the HIPT model and I really think it's the best model for this competition, but I don't think it's viable since it is a huge model.
\nFrom this notebook it seems that a single ViT with small images such as used, uses about 9GB. It could be the images loaded to memory but it is still a 21 Million param model.
\nHas anyone tried this out?
", "rawMarkdown": "I have been wrapping my head around the HIPT model and I really think it's the best model for this competition, but I don't think it's viable since it is a huge model.\n\nFrom [this notebook](https://colab.research.google.com/github/keras-team/keras-io/blob/master/examples/vision/ipynb/image_classification_with_vision_transformer.ipynb) it seems that a single ViT with small images such as used, uses about 9GB. It could be the images loaded to memory but it is still a 21 Million param model. \n\n Has anyone tried this out?", "votes": 1 }, { "id": 1897358, "postDate": "2022-08-13T17:17:09.317Z", "content": "Amazing work!!
", "rawMarkdown": "Amazing work!!" }, { "id": 1861348, "postDate": "2022-07-19T01:30:41.303Z", "content": "Good collection of relevant information @hengck23
", "rawMarkdown": "Good collection of relevant information @hengck23" }, { "id": 1861137, "postDate": "2022-07-18T19:40:03.540Z", "content": "Great content! @hengck23
", "rawMarkdown": "Great content! @hengck23 " } ], "comments": [ { "id": 1861891, "author_name": "Marcus Yang", "author_url": "", "post_date": "2022-07-19T10:23:34.827000", "content": "Awesome, I am the author of TransMIL. Thank you for following our work~
", "votes": 15, "replies": [] }, { "id": 1864328, "author_name": "hengck23", "author_url": "", "post_date": "2022-07-21T00:48:57.940000", "content": "there are many good videos from standford medical AI.
\nthis one is about transformer MIL for WSI
\nMedAI #36: Weakly supervised tumor detection in whole slide image analysis | Bin Li
\nhttps://www.youtube.com/watch?v=ZPe94q8wxPQ
I have been wrapping my head around the HIPT model and I really think it's the best model for this competition, but I don't think it's viable since it is a huge model.
\nFrom this notebook it seems that a single ViT with small images such as used, uses about 9GB. It could be the images loaded to memory but it is still a 21 Million param model.
\nHas anyone tried this out?
", "votes": 1, "replies": [] }, { "id": 1897358, "author_name": "Yimi Machado", "author_url": "", "post_date": "2022-08-13T17:17:09.317000", "content": "Amazing work!!
", "votes": 0, "replies": [] }, { "id": 1861348, "author_name": "Henry (Shang-Rou) Hsieh", "author_url": "", "post_date": "2022-07-19T01:30:41.303000", "content": "Good collection of relevant information @hengck23
", "votes": 0, "replies": [] }, { "id": 1861137, "author_name": "Apurba Pandey", "author_url": "", "post_date": "2022-07-18T19:40:03.540000", "content": "Great content! @hengck23
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1860259": "\nreference:\n[1] TransMIL: Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification\n\n\npretrain model (trained using SSL) for replacing resnet50:\n- https://github.com/mahmoodlab/HIPT\ncvpr 2022: Scaling Vision Transformers to Gigapixel Images via Hierarchical Self-Supervised Learning\n\n\n\nrelated:\n\n[2] Feature Re-calibration based Multiple Instance Learning for Whole Slide Image Classification\n\n[3] Multimodal Co-Attention Transformer for Survival Prediction in Gigapixel Whole Slide Images\n(formulation of set transformer for MIL)\n\n\nbackground reading:\n[4] Characterization of the ‘White’ Appearing Clots that Cause Acute Ischemic Stroke\nhttps://www.sciencedirect.com/science/article/pii/S1052305721005322", "1861891": "Awesome, I am the author of TransMIL. Thank you for following our work~", "1864328": "there are many good videos from standford medical AI.\n\nthis one is about transformer MIL for WSI\nMedAI #36: Weakly supervised tumor detection in whole slide image analysis | Bin Li\nhttps://www.youtube.com/watch?v=ZPe94q8wxPQ", "1862928": "I have been wrapping my head around the HIPT model and I really think it's the best model for this competition, but I don't think it's viable since it is a huge model.\n\nFrom [this notebook](https://colab.research.google.com/github/keras-team/keras-io/blob/master/examples/vision/ipynb/image_classification_with_vision_transformer.ipynb) it seems that a single ViT with small images such as used, uses about 9GB. It could be the images loaded to memory but it is still a 21 Million param model. \n\n Has anyone tried this out?", "1897358": "Amazing work!!", "1861348": "Good collection of relevant information @hengck23", "1861137": "Great content! @hengck23 " } }