{ "id": 375039, "title": "❗View feature is more important than you think❗", "url": "/competitions/rsna-breast-cancer-detection/discussion/375039", "author_name": "Francisco Javier Gallego", "post_date": "2022-12-30T01:23:15.022000", "votes": 7, "comment_count": 2, "views": 0, "content": "

Recently, @hengck23 posted a discussion talking about the fact that ADMANI dataset[1] is part of this competition's data (see discussion https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333#2076911, and one notebook explaining the details of some papers that show some techniques for getting good results with the ADMANI dataset. Let's examine better the image attached to the explanation:

\n

\"\"

\n

You can observe that images are clasified in main and auxiliary images. Let's now observe an example of an image for each of the different types of view we're given.

\n

\"\"

\n

Actually CC and MLO correspond to the main and auxiliary types, respectively. To find out it by yourselves please head to the following articles:

\n\n

Therefore, regarding everything mentioned by @hengck23, it could be time-worthy to spend some time playing around with these two types when training our models.

\n", "messages": [ { "id": 2080230, "postDate": "2022-12-30T01:23:15.023Z", "content": "

Recently, @hengck23 posted a discussion talking about the fact that ADMANI dataset[1] is part of this competition's data (see discussion https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333#2076911, and one notebook explaining the details of some papers that show some techniques for getting good results with the ADMANI dataset. Let's examine better the image attached to the explanation:

\n

\"\"

\n

You can observe that images are clasified in main and auxiliary images. Let's now observe an example of an image for each of the different types of view we're given.

\n

\"\"

\n

Actually CC and MLO correspond to the main and auxiliary types, respectively. To find out it by yourselves please head to the following articles:

\n\n

Therefore, regarding everything mentioned by @hengck23, it could be time-worthy to spend some time playing around with these two types when training our models.

\n", "rawMarkdown": "Recently, [@hengck23](https://www.kaggle.com/hengck23) posted a discussion talking about the fact that ADMANI dataset[1] is part of this competition's data (see discussion https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333#2076911, and one [notebook](https://www.kaggle.com/code/hengck23/mvccl-model-for-admani-dataset) explaining the details of some papers that show some techniques for getting good results with the ADMANI dataset. Let's examine better the image attached to the explanation: \n\n![](https://i.ibb.co/4S9BcxG/Selection-318.png)\n\nYou can observe that images are clasified in **main** and **auxiliary** images. Let's now observe an example of an image for each of the different types of view we're given. \n\n![](https://www.kaggleusercontent.com/kf/115052136/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..9kK8GDo8WDLo0ZucjVdgWg.Dk4tAe9fBSuZf4Th_axXk0s0yN4ynW2zgKb8nZpI03ND6ZqdUCu3kdnNA9wU1JHW--J75wqUTtw7y4WmIQIS8_tizi6JczJjZfhwnFxa5S07L4oWSiKcNAgwTj_GjslHFbrXe9wBbVUNl4OgTmlciGNDFJbm_Nz4ghPyH-RXS_4ouxULikdiV-b9Wb0_EIYueTrSwWVLy46ruolRI-trOgnFUY6qmocrPLSPpnxX3tZwuztaxbyMkDLV_s8LQAYn32nBtWieWQugfayY_0jGTLNfc9zoCYWKF5Rt30nOApIE9akEfLoRjKB_tOtvwi9h-cdC_oUi_PqkgqpR0qDaEqxPOilqNH6rhCA0eAUot-0PwgZIA24m7cPlDhcbgfwrNmk0F6g9gJj7A5DTNgf32cVCHLn27_m1x2WZ8Iee7RUVRfOe3SXIMdUPiBuOP9C6bXvafXWFfYvTy2bUDWO5Bosfx2aL80_2oeN239bIAT3j-D8hjk9OofjW4qXfUtXVlrJp1a0wQMC5KjlErzgZmdmMRpmuu6QteL8veRDw-zBb3eRWu2Gka5S-fNM0uSvRJJ1Iic33YNj0e9b7mSw-CHwZyHT1aAP7RCfmdaIINlwYN5VNApKxjLrWomU9hqPe8uv5ShRaP2PKn_0lDjmlM5GG509TCxi3k-hkbSR89rs.xCYORk2-mqWlxsR4_LVbOg/__results___files/__results___23_0.png)\n\nActually **CC and MLO correspond to the main and auxiliary types**, respectively. To find out it by yourselves please head to the following articles:\n\n* [Article 1](https://radiopaedia.org/articles/craniocaudal-view)\n* [Article 2](https://radiopaedia.org/articles/mediolateral-oblique-view?lang=us)\n\nTherefore, regarding everything mentioned by [@hengck23](https://www.kaggle.com/hengck23), it could be time-worthy to spend some time playing around with these two types when training our models. \n\n* For more detail please refer to: https://www.kaggle.com/code/javigallego/rsna-complete-eda", "votes": 7 }, { "id": 2081161, "postDate": "2022-12-30T21:35:43.897Z", "content": "

very interesting observation, @javigallego! Thanks for starting the discussion on this! πŸ™Œ

", "rawMarkdown": "very interesting observation, @javigallego! Thanks for starting the discussion on this! πŸ™Œ ", "votes": 1 }, { "id": 2080295, "postDate": "2022-12-30T04:15:33.657Z", "content": "

to see how it works:

\n
    \n
  1. train a single view model and get CAM activation maps
    2. \n
  2. train a dual view model. you can get CAM maps or attention maps.
    3. \n
\n

comparison of the two activation maps will tell you if it works or not.
\nif the cross attention is correct you should see how region in one view affects the region in another.

\n

google for \"cross view attention mammography\" for more

", "rawMarkdown": "to see how it works:\n1. train a single view model and get CAM activation maps\n2. train a dual view model. you can get CAM maps or attention maps.\n\n\ncomparison of the two activation maps will tell you if it works or not.\nif the cross attention is correct you should see how region in one view affects the region in another.\n\ngoogle for \"cross view attention mammography\" for more", "votes": 1 } ], "comments": [ { "id": 2081161, "author_name": "Radek Osmulski", "author_url": "", "post_date": "2022-12-30T21:35:43.897000", "content": "

very interesting observation, @javigallego! Thanks for starting the discussion on this! πŸ™Œ

", "votes": 1, "replies": [] }, { "id": 2080295, "author_name": "hengck23", "author_url": "", "post_date": "2022-12-30T04:15:33.657000", "content": "

to see how it works:

\n
    \n
  1. train a single view model and get CAM activation maps
    2. \n
  2. train a dual view model. you can get CAM maps or attention maps.
    3. \n
\n

comparison of the two activation maps will tell you if it works or not.
\nif the cross attention is correct you should see how region in one view affects the region in another.

\n

google for \"cross view attention mammography\" for more

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "2080230": "Recently, [@hengck23](https://www.kaggle.com/hengck23) posted a discussion talking about the fact that ADMANI dataset[1] is part of this competition's data (see discussion https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333#2076911, and one [notebook](https://www.kaggle.com/code/hengck23/mvccl-model-for-admani-dataset) explaining the details of some papers that show some techniques for getting good results with the ADMANI dataset. Let's examine better the image attached to the explanation: \n\n![](https://i.ibb.co/4S9BcxG/Selection-318.png)\n\nYou can observe that images are clasified in **main** and **auxiliary** images. Let's now observe an example of an image for each of the different types of view we're given. \n\n![](https://www.kaggleusercontent.com/kf/115052136/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..9kK8GDo8WDLo0ZucjVdgWg.Dk4tAe9fBSuZf4Th_axXk0s0yN4ynW2zgKb8nZpI03ND6ZqdUCu3kdnNA9wU1JHW--J75wqUTtw7y4WmIQIS8_tizi6JczJjZfhwnFxa5S07L4oWSiKcNAgwTj_GjslHFbrXe9wBbVUNl4OgTmlciGNDFJbm_Nz4ghPyH-RXS_4ouxULikdiV-b9Wb0_EIYueTrSwWVLy46ruolRI-trOgnFUY6qmocrPLSPpnxX3tZwuztaxbyMkDLV_s8LQAYn32nBtWieWQugfayY_0jGTLNfc9zoCYWKF5Rt30nOApIE9akEfLoRjKB_tOtvwi9h-cdC_oUi_PqkgqpR0qDaEqxPOilqNH6rhCA0eAUot-0PwgZIA24m7cPlDhcbgfwrNmk0F6g9gJj7A5DTNgf32cVCHLn27_m1x2WZ8Iee7RUVRfOe3SXIMdUPiBuOP9C6bXvafXWFfYvTy2bUDWO5Bosfx2aL80_2oeN239bIAT3j-D8hjk9OofjW4qXfUtXVlrJp1a0wQMC5KjlErzgZmdmMRpmuu6QteL8veRDw-zBb3eRWu2Gka5S-fNM0uSvRJJ1Iic33YNj0e9b7mSw-CHwZyHT1aAP7RCfmdaIINlwYN5VNApKxjLrWomU9hqPe8uv5ShRaP2PKn_0lDjmlM5GG509TCxi3k-hkbSR89rs.xCYORk2-mqWlxsR4_LVbOg/__results___files/__results___23_0.png)\n\nActually **CC and MLO correspond to the main and auxiliary types**, respectively. To find out it by yourselves please head to the following articles:\n\n* [Article 1](https://radiopaedia.org/articles/craniocaudal-view)\n* [Article 2](https://radiopaedia.org/articles/mediolateral-oblique-view?lang=us)\n\nTherefore, regarding everything mentioned by [@hengck23](https://www.kaggle.com/hengck23), it could be time-worthy to spend some time playing around with these two types when training our models. \n\n* For more detail please refer to: https://www.kaggle.com/code/javigallego/rsna-complete-eda", "2081161": "very interesting observation, @javigallego! Thanks for starting the discussion on this! πŸ™Œ ", "2080295": "to see how it works:\n1. train a single view model and get CAM activation maps\n2. train a dual view model. you can get CAM maps or attention maps.\n\n\ncomparison of the two activation maps will tell you if it works or not.\nif the cross attention is correct you should see how region in one view affects the region in another.\n\ngoogle for \"cross view attention mammography\" for more" } }