What is Transfer Learning
\nTransfer learning is a technique where a model is pre-trained on one task and then later can be re-purposed for a new task with a little additional training.
Why not ImageNet
\nImageNet is a dataset of millions of images. Most standard libraries such as torchvision, timm, etc. have a pretrained parameter that will allow you to use transfer learning with a model pre-trained on ImageNet. However, ImageNet may not be ideal for medical image classification because ImageNet includes more everyday items such as vehicles, instruments, and animals. Thus, finding models pre-trained on different datasets including images such as x-rays, body parts, etc may be for fitting for a medical task such as this.
Nonetheless, ImageNet is actually still a very good option and can produce very powerful models. It is also much simpler since it is more common, so you may have more time to work on training your model for this specific task leading to better results.
\nRelated Research Papers:
\nTransfer learning for medical image classification: a literature review
\nRethinking Transfer Learning for Medical Image Classification
Model Architecture
\nAccording to this article, when classifying images in the breast anatomical region (as in this challenge), the most commonly used model architecture is Inception. Pre-trained models such as Inception-v3 are very common and can be found in libraries such as torchvision and timm. Perhaps look for an inception model that has been pretrained on a medical task. Nonetheless, this is just a suggestion and the best model architecture should be decided through trials and cross validation on this particular task. You may also want to note that the article was written pre-image transformers era, so other more recent transformer model architectures may be stronger.
MedNet
\nIf you are looking for a pre-trained medical model, you should consider looking at MedNet. MedNet is basically a deep CNN pre-trained on a bunch of medical images. I haven't looked at it too closely yet, but below is the research paper and a github repo with code.
\nMedNet Research Paper
\nMedNet Github
Previous Competitions
\nEven though this competition is about breast cancer detection, other medical image competitions have been held by the same host (Radiological Society of North America). In these previous competitions, many people make their models' weights public at the end, so you may be able to use their weights as a pre-trained starting point for this competition. For reference see previous RSNA competition winning solutions by @radek1
What is Transfer Learning
\nTransfer learning is a technique where a model is pre-trained on one task and then later can be re-purposed for a new task with a little additional training.
Why not ImageNet
\nImageNet is a dataset of millions of images. Most standard libraries such as torchvision, timm, etc. have a pretrained parameter that will allow you to use transfer learning with a model pre-trained on ImageNet. However, ImageNet may not be ideal for medical image classification because ImageNet includes more everyday items such as vehicles, instruments, and animals. Thus, finding models pre-trained on different datasets including images such as x-rays, body parts, etc may be for fitting for a medical task such as this.
Nonetheless, ImageNet is actually still a very good option and can produce very powerful models. It is also much simpler since it is more common, so you may have more time to work on training your model for this specific task leading to better results.
\nRelated Research Papers:
\nTransfer learning for medical image classification: a literature review
\nRethinking Transfer Learning for Medical Image Classification
Model Architecture
\nAccording to this article, when classifying images in the breast anatomical region (as in this challenge), the most commonly used model architecture is Inception. Pre-trained models such as Inception-v3 are very common and can be found in libraries such as torchvision and timm. Perhaps look for an inception model that has been pretrained on a medical task. Nonetheless, this is just a suggestion and the best model architecture should be decided through trials and cross validation on this particular task. You may also want to note that the article was written pre-image transformers era, so other more recent transformer model architectures may be stronger.
MedNet
\nIf you are looking for a pre-trained medical model, you should consider looking at MedNet. MedNet is basically a deep CNN pre-trained on a bunch of medical images. I haven't looked at it too closely yet, but below is the research paper and a github repo with code.
\nMedNet Research Paper
\nMedNet Github
Previous Competitions
\nEven though this competition is about breast cancer detection, other medical image competitions have been held by the same host (Radiological Society of North America). In these previous competitions, many people make their models' weights public at the end, so you may be able to use their weights as a pre-trained starting point for this competition. For reference see previous RSNA competition winning solutions by @radek1
\n\nAccording to this article, when classifying images in the breast anatomical region (as in this challenge), the most commonly used model architecture is Inception.
\n
I skimmed the article quickly, from my understanding: one thing to note-the article compiles learning from the \"pre-image transformer eras \"
", "rawMarkdown": "> According to this article, when classifying images in the breast anatomical region (as in this challenge), the most commonly used model architecture is Inception.\n\nI skimmed the article quickly, from my understanding: one thing to note-the article compiles learning from the \"pre-image transformer eras \"", "votes": 1, "replies": [ { "id": 2050706, "postDate": "2022-11-30T23:45:50.193Z", "content": "@init27 Thanks for pointing this out. I updated my post with a note to address this.
", "rawMarkdown": "@init27 Thanks for pointing this out. I updated my post with a note to address this. " } ] } ], "comments": [ { "id": 2049324, "author_name": "Sanyam Bhutani", "author_url": "", "post_date": "2022-11-30T03:48:53.353000", "content": "\n\nAccording to this article, when classifying images in the breast anatomical region (as in this challenge), the most commonly used model architecture is Inception.
\n
I skimmed the article quickly, from my understanding: one thing to note-the article compiles learning from the \"pre-image transformer eras \"
", "votes": 1, "replies": [ { "id": 2050706, "author_name": "Ravi Shah", "author_url": "", "post_date": "2022-11-30T23:45:50.193000", "content": "@init27 Thanks for pointing this out. I updated my post with a note to address this.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2049145": "# Transfer Learning for Medical Images\n\n**What is Transfer Learning**\nTransfer learning is a technique where a model is pre-trained on one task and then later can be re-purposed for a new task with a little additional training. \n\n**Why not ImageNet**\nImageNet is a dataset of millions of images. Most standard libraries such as torchvision, timm, etc. have a pretrained parameter that will allow you to use transfer learning with a model pre-trained on ImageNet. However, ImageNet may not be ideal for medical image classification because ImageNet includes more everyday items such as vehicles, instruments, and animals. Thus, finding models pre-trained on different datasets including images such as x-rays, body parts, etc may be for fitting for a medical task such as this.\n\nNonetheless, ImageNet is actually still a very good option and can produce very powerful models. It is also much simpler since it is more common, so you may have more time to work on training your model for this specific task leading to better results. \n\n**Related Research Papers:**\n[Transfer learning for medical image classification: a literature review](https://pubmed.ncbi.nlm.nih.gov/35418051/)\n[Rethinking Transfer Learning for Medical Image Classification](https://arxiv.org/pdf/2106.05152.pdf)\n\n# Finding a Model\n\n**Model Architecture**\nAccording to [this article](https://learnopencv.com/transfer-learning-for-medical-images/), when classifying images in the breast anatomical region (as in this challenge), the most commonly used model architecture is Inception. Pre-trained models such as Inception-v3 are very common and can be found in libraries such as torchvision and timm. Perhaps look for an inception model that has been pretrained on a medical task. Nonetheless, this is just a suggestion and the best model architecture should be decided through trials and cross validation on this particular task. You may also want to note that the article was written pre-image transformers era, so other more recent transformer model architectures may be stronger. \n\n**MedNet**\nIf you are looking for a pre-trained medical model, you should consider looking at MedNet. MedNet is basically a deep CNN pre-trained on a bunch of medical images. I haven't looked at it too closely yet, but below is the research paper and a github repo with code.\n[MedNet Research Paper](https://arxiv.org/ftp/arxiv/papers/2110/2110.06512.pdf)\n[MedNet Github](https://github.com/Tencent/MedicalNet)\n\n**Previous Competitions**\nEven though this competition is about breast cancer detection, other medical image competitions have been held by the same host (Radiological Society of North America). In these previous competitions, many people make their models' weights public at the end, so you may be able to use their weights as a pre-trained starting point for this competition. For reference see [previous RSNA competition winning solutions](https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/369103) by @radek1 ", "2049324": "> According to this article, when classifying images in the breast anatomical region (as in this challenge), the most commonly used model architecture is Inception.\n\nI skimmed the article quickly, from my understanding: one thing to note-the article compiles learning from the \"pre-image transformer eras \"" } }