![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F505720556c2b2b46333341ee93fa7142%2FSelection_049.png?generation=1720739651889086&alt=media\")
Do you think that x, y coordinates are enough to guide the training task (segmentation/classification)?
", "messages": [ { "id": "2846369", "postDate": "05/31/2024 05:24:16", "content": "Do you think that x, y coordinates are enough to guide the training task (segmentation/classification)?
", "rawMarkdown": "Do you think that x, y coordinates are enough to guide the training task (segmentation/classification)?", "votes": null }, { "id": "2846873", "postDate": "05/31/2024 09:57:07", "content": "I think it's possible, but I've just finished the baseline and haven't had the chance to experiment yet.
", "rawMarkdown": "I think it's possible, but I've just finished the baseline and haven't had the chance to experiment yet.", "votes": null }, { "id": "2847592", "postDate": "05/31/2024 15:32:05", "content": "I think it would be better if it is an actual segmentation mask but it would suffice if we do some patching
", "rawMarkdown": "I think it would be better if it is an actual segmentation mask but it would suffice if we do some patching", "votes": null }, { "id": "2910537", "postDate": "07/07/2024 18:14:16", "content": "I also think that it's possible. SAM with point prompt might be worth trying.
", "rawMarkdown": "I also think that it's possible. SAM with point prompt might be worth trying.", "votes": null }, { "id": "2917961", "postDate": "07/11/2024 23:21:52", "content": "point annotations are common in medical imaging. i show 2 examples and you can google for more.
\n[1] paper: A scalable physician-level deep learning algorithm detects universal trauma on pelvic radiographs
\n
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F78f060e8c318ce108529cbf80395c160%2FSelection_050.png?generation=1720739667133993&alt=media\")
[2] paper: Lesion Segmentation and RECIST Diameter Prediction via Click-driven Attention and Dual-path Connection
\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4da1ececc1feaa7bb67449e2760ef975%2FSelection_047.png?generation=1720739787647623&alt=media\")
![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F9ab4cbba89b5f1139e90bdb9656b85b3%2FSelection_048.png?generation=1720739819424165&alt=media\")
how to use the point information
\nin summary, use the point supervision to guide your CAM or attention map. classification is given by pooling the feature (of CAM) or attention
", "rawMarkdown": "point annotations are common in medical imaging. i show 2 examples and you can google for more.\n\n[1] paper: A scalable physician-level deep learning algorithm detects universal trauma on pelvic radiographs\n\n\n\n\n\n---\n\n[2] paper: Lesion Segmentation and RECIST Diameter Prediction via Click-driven Attention and Dual-path Connection\n\n\n\n\n---\nhow to use the point information\n1. as supervision signal (i.e.your classification loss function that uses x,y)\n2. aux prediction and aux loss\n3. use image + (x,y) as input and train mopdel to produce heatmap. Then train a model to predict heatmap from image. finally train a classification model using input = image+heatmap.\n\n\nin summary, use the point supervision to guide your CAM or attention map. classification is given by pooling the feature (of CAM) or attention", "votes": null } ], "comments": [ { "id": 2846873, "author_name": "", "author_url": "", "post_date": "05/31/2024 09:57:07", "content": "I think it's possible, but I've just finished the baseline and haven't had the chance to experiment yet.
", "votes": null, "replies": [] }, { "id": 2847592, "author_name": "dantingzeng", "author_url": "", "post_date": "05/31/2024 15:32:05", "content": "I think it would be better if it is an actual segmentation mask but it would suffice if we do some patching
", "votes": null, "replies": [] }, { "id": 2910537, "author_name": "gunesevitan", "author_url": "", "post_date": "07/07/2024 18:14:16", "content": "I also think that it's possible. SAM with point prompt might be worth trying.
", "votes": null, "replies": [] }, { "id": 2917961, "author_name": "hengck23", "author_url": "", "post_date": "07/11/2024 23:21:52", "content": "point annotations are common in medical imaging. i show 2 examples and you can google for more.
\n[1] paper: A scalable physician-level deep learning algorithm detects universal trauma on pelvic radiographs
\n
[2] paper: Lesion Segmentation and RECIST Diameter Prediction via Click-driven Attention and Dual-path Connection
\n
how to use the point information
\nin summary, use the point supervision to guide your CAM or attention map. classification is given by pooling the feature (of CAM) or attention
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2846369": "Do you think that x, y coordinates are enough to guide the training task (segmentation/classification)?", "2846873": "I think it's possible, but I've just finished the baseline and haven't had the chance to experiment yet.", "2847592": "I think it would be better if it is an actual segmentation mask but it would suffice if we do some patching", "2910537": "I also think that it's possible. SAM with point prompt might be worth trying.", "2917961": "point annotations are common in medical imaging. i show 2 examples and you can google for more.\n\n[1] paper: A scalable physician-level deep learning algorithm detects universal trauma on pelvic radiographs\n\n\n\n\n\n---\n\n[2] paper: Lesion Segmentation and RECIST Diameter Prediction via Click-driven Attention and Dual-path Connection\n\n\n\n\n---\nhow to use the point information\n1. as supervision signal (i.e.your classification loss function that uses x,y)\n2. aux prediction and aux loss\n3. use image + (x,y) as input and train mopdel to produce heatmap. Then train a model to predict heatmap from image. finally train a classification model using input = image+heatmap.\n\n\nin summary, use the point supervision to guide your CAM or attention map. classification is given by pooling the feature (of CAM) or attention" }, "source": "meta" }