I have noticed that the highest-scoring notebooks in the competition have used pre-trained networks such as resnet, VGG16, and various 3D classifying models. These models, however, aren't trained on MRI images, models such as resnet and efficientnet were trained on the imagenet database. So how are they able to classify MRI images so well?
", "messages": [ { "id": 1402365, "postDate": "2021-07-28T06:01:51.707Z", "content": "I have noticed that the highest-scoring notebooks in the competition have used pre-trained networks such as resnet, VGG16, and various 3D classifying models. These models, however, aren't trained on MRI images, models such as resnet and efficientnet were trained on the imagenet database. So how are they able to classify MRI images so well?
", "rawMarkdown": "I have noticed that the highest-scoring notebooks in the competition have used pre-trained networks such as resnet, VGG16, and various 3D classifying models. These models, however, aren't trained on MRI images, models such as resnet and efficientnet were trained on the imagenet database. So how are they able to classify MRI images so well?", "votes": 7 }, { "id": 1411359, "postDate": "2021-08-02T18:39:40.210Z", "content": "I built a simple (much smaller than any pre-trained network) 3DCNN initialized from scratch and scored 0.65.
\nSo I would hesitate to say that pre-training is doing better than an uninitialized model.
\nJust my 2 cents though.
", "rawMarkdown": "I built a simple (much smaller than any pre-trained network) 3DCNN initialized from scratch and scored 0.65. \n\nSo I would hesitate to say that pre-training is doing better than an uninitialized model.\n\nJust my 2 cents though.", "votes": 4, "replies": [ { "id": 1486614, "postDate": "2021-08-23T05:27:19.940Z", "content": "@dschettler8845 I tried to do the same but the model training is not converging , may i request you to share some simplified version of your notebook. Currently i am using pytorch , with 2 blocks of 3d cnn(32 filters)->batchNormaliztion followed by 2 fully connected layers , no skip connections , what i got on public lb was 0.53
", "rawMarkdown": "@dschettler8845 I tried to do the same but the model training is not converging , may i request you to share some simplified version of your notebook. Currently i am using pytorch , with 2 blocks of 3d cnn(32 filters)->batchNormaliztion followed by 2 fully connected layers , no skip connections , what i got on public lb was 0.53" }, { "id": 1487365, "postDate": "2021-08-23T15:22:14.850Z", "content": "This is my notebook.
\nThese models (the saved models) score similar to what I mentioned (T2W specifically).
\nI hope this helps.
", "rawMarkdown": "This is [my notebook](https://www.kaggle.com/dschettler8845/eda-3d-baseline-rsna-glioma-radiogenomics). \n\nThese models (the saved models) score similar to what I mentioned (T2W specifically).\n\nI hope this helps." } ] }, { "id": 1402905, "postDate": "2021-07-28T16:19:39.770Z", "content": "I think we evaluate MRI images the same way we evaluate other images, e.g. the ones that resnet, VGG16, etc. are trained on. We look for areas of uniformity, contrast, shapes etc. We are genetically and by evolution conditioned to look at images that way. We are anthropocentric. I'm not sure this is always optimal.
", "rawMarkdown": "I think we evaluate MRI images the same way we evaluate other images, e.g. the ones that resnet, VGG16, etc. are trained on. We look for areas of uniformity, contrast, shapes etc. We are genetically and by evolution conditioned to look at images that way. We are anthropocentric. I'm not sure this is always optimal. \n\n", "votes": 4, "replies": [ { "id": 1403059, "postDate": "2021-07-28T18:41:02.140Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1403704, "postDate": "2021-07-29T09:48:12.650Z", "content": "Thanks, this makes sense. So the pre-trained neural networks have learned to identify patterns that transfer over to identify patterns in MRI images.
", "rawMarkdown": "Thanks, this makes sense. So the pre-trained neural networks have learned to identify patterns that transfer over to identify patterns in MRI images." } ] }, { "id": 1408217, "postDate": "2021-08-02T11:27:35.193Z", "content": "This is the power of transfer learning
", "rawMarkdown": "This is the power of transfer learning " }, { "id": 1406010, "postDate": "2021-07-31T12:11:05.960Z", "rawMarkdown": "", "votes": -1, "isDeleted": true }, { "id": 1402602, "postDate": "2021-07-28T10:51:02.243Z", "rawMarkdown": "", "votes": -1, "isDeleted": true } ], "comments": [ { "id": 1411359, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-08-02T18:39:40.210000", "content": "I built a simple (much smaller than any pre-trained network) 3DCNN initialized from scratch and scored 0.65.
\nSo I would hesitate to say that pre-training is doing better than an uninitialized model.
\nJust my 2 cents though.
", "votes": 4, "replies": [ { "id": 1486614, "author_name": "AdityaVikramSingh", "author_url": "", "post_date": "2021-08-23T05:27:19.940000", "content": "@dschettler8845 I tried to do the same but the model training is not converging , may i request you to share some simplified version of your notebook. Currently i am using pytorch , with 2 blocks of 3d cnn(32 filters)->batchNormaliztion followed by 2 fully connected layers , no skip connections , what i got on public lb was 0.53
", "votes": 0, "replies": [] }, { "id": 1487365, "author_name": "Darien Schettler", "author_url": "", "post_date": "2021-08-23T15:22:14.850000", "content": "This is my notebook.
\nThese models (the saved models) score similar to what I mentioned (T2W specifically).
\nI hope this helps.
", "votes": 0, "replies": [] } ] }, { "id": 1402905, "author_name": "Philippe De Wilde", "author_url": "", "post_date": "2021-07-28T16:19:39.770000", "content": "I think we evaluate MRI images the same way we evaluate other images, e.g. the ones that resnet, VGG16, etc. are trained on. We look for areas of uniformity, contrast, shapes etc. We are genetically and by evolution conditioned to look at images that way. We are anthropocentric. I'm not sure this is always optimal.
", "votes": 4, "replies": [ { "id": 1403059, "author_name": "", "author_url": "", "post_date": "2021-07-28T18:41:02.140000", "content": "", "votes": 0, "replies": [] }, { "id": 1403704, "author_name": "Keegan Fernandes", "author_url": "", "post_date": "2021-07-29T09:48:12.650000", "content": "Thanks, this makes sense. So the pre-trained neural networks have learned to identify patterns that transfer over to identify patterns in MRI images.
", "votes": 0, "replies": [] } ] }, { "id": 1408217, "author_name": "Haw Keat", "author_url": "", "post_date": "2021-08-02T11:27:35.193000", "content": "This is the power of transfer learning
", "votes": 0, "replies": [] }, { "id": 1406010, "author_name": "", "author_url": "", "post_date": "2021-07-31T12:11:05.960000", "content": "", "votes": -1, "replies": [] }, { "id": 1402602, "author_name": "", "author_url": "", "post_date": "2021-07-28T10:51:02.243000", "content": "", "votes": -1, "replies": [] } ], "raw_markdown_by_id": { "1402365": "I have noticed that the highest-scoring notebooks in the competition have used pre-trained networks such as resnet, VGG16, and various 3D classifying models. These models, however, aren't trained on MRI images, models such as resnet and efficientnet were trained on the imagenet database. So how are they able to classify MRI images so well?", "1411359": "I built a simple (much smaller than any pre-trained network) 3DCNN initialized from scratch and scored 0.65. \n\nSo I would hesitate to say that pre-training is doing better than an uninitialized model.\n\nJust my 2 cents though.", "1402905": "I think we evaluate MRI images the same way we evaluate other images, e.g. the ones that resnet, VGG16, etc. are trained on. We look for areas of uniformity, contrast, shapes etc. We are genetically and by evolution conditioned to look at images that way. We are anthropocentric. I'm not sure this is always optimal. \n\n", "1408217": "This is the power of transfer learning ", "1406010": "", "1402602": "" } }