Couple of years ago I came across this paper describing weights pre-trained on grayscale ImageNet (please see the Xie and Richmond paper for details).
\n\n\nXie, Y. and Richmond, D., “Pre-training on grayscale ImageNet improves medical image classification,” in [Proceedings of the European Conference on Computer Vision (ECCV) Workshops], 476–484, Springer (September 2019).
\n
I haven't been able to find models with such design and pre-training openly available, but it sounds that they could be equally good as their 3-channel versions, but a little bit lighter. Adjusting the first convolutional layer of the pre-trained network or having copies of the input images to make the input match the usual pre-trained model architecture provides only partial relief, right.
\nAny thoughts? Or resources to test this?
\nPS. It has been really great to follow the competition so far! Rarely there are so many people gathered in one place to talk about breast cancer detection, much like some dedicated conference or workshop, but for an extended amount of time. I have read 99% of the discussion here and enjoyed it very much. As many of you here know and many others have noticed, there are several types of breast cancer. This means large within class variation in the cancer class. Subcategories are difficult to learn as there are probably only few samples of each of those. What will be making the task somewhat more difficult are the similarities between non-cancerous findings and cancerous findings---this is possible. This might lead to conclude that some kind of normal detector would be good to have, but as we are talking about cancer here, it is good to concentrate---in real life---on learning the patterns which associate to cancer---for a competition this might be different.
Couple of years ago I came across this paper describing weights pre-trained on grayscale ImageNet (please see the Xie and Richmond paper for details).
\n\n\nXie, Y. and Richmond, D., “Pre-training on grayscale ImageNet improves medical image classification,” in [Proceedings of the European Conference on Computer Vision (ECCV) Workshops], 476–484, Springer (September 2019).
\n
I haven't been able to find models with such design and pre-training openly available, but it sounds that they could be equally good as their 3-channel versions, but a little bit lighter. Adjusting the first convolutional layer of the pre-trained network or having copies of the input images to make the input match the usual pre-trained model architecture provides only partial relief, right.
\nAny thoughts? Or resources to test this?
\nPS. It has been really great to follow the competition so far! Rarely there are so many people gathered in one place to talk about breast cancer detection, much like some dedicated conference or workshop, but for an extended amount of time. I have read 99% of the discussion here and enjoyed it very much. As many of you here know and many others have noticed, there are several types of breast cancer. This means large within class variation in the cancer class. Subcategories are difficult to learn as there are probably only few samples of each of those. What will be making the task somewhat more difficult are the similarities between non-cancerous findings and cancerous findings---this is possible. This might lead to conclude that some kind of normal detector would be good to have, but as we are talking about cancer here, it is good to concentrate---in real life---on learning the patterns which associate to cancer---for a competition this might be different.
Similar discussion here.
", "rawMarkdown": "Similar discussion [here](https://www.kaggle.com/questions-and-answers/199224)." } ], "comments": [ { "id": 2129941, "author_name": "Antti Isosalo", "author_url": "", "post_date": "2023-02-05T03:22:00.003000", "content": "Similar discussion here.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2129934": "Couple of years ago I came across this paper describing weights pre-trained on grayscale ImageNet (please see the Xie and Richmond paper for details).\n\n>Xie, Y. and Richmond, D., “Pre-training on grayscale ImageNet improves medical image classification,” in [Proceedings of the European Conference on Computer Vision (ECCV) Workshops], 476–484, Springer (September 2019).\n\nI haven't been able to find models with such design and pre-training openly available, but it sounds that they could be equally good as their 3-channel versions, but a little bit lighter. [Adjusting the first convolutional layer of the pre-trained network](https://www.kaggle.com/code/romanrybalko/pretrained-resnet-with-grayscale-images?scriptVersionId=106060681&cellId=16) or having copies of the input images to make the input match the usual pre-trained model architecture provides only partial relief, right.\n\nAny thoughts? Or resources to test this?\n\nPS. It has been really great to follow the competition so far! Rarely there are so many people gathered in one place to talk about breast cancer detection, much like some dedicated conference or workshop, but for an extended amount of time. I have read 99% of the discussion here and enjoyed it very much. As many of you here know and many others have noticed, there are several types of breast cancer. This means large within class variation in the `cancer` class. Subcategories are difficult to learn as there are probably only few samples of each of those. What will be making the task somewhat more difficult are the similarities between non-cancerous findings and cancerous findings---this is possible. This might lead to conclude that some kind of normal detector would be good to have, but as we are talking about cancer here, it is good to concentrate---in real life---on learning the patterns which associate to cancer---for a competition this might be different.", "2129941": "Similar discussion [here](https://www.kaggle.com/questions-and-answers/199224)." } }