I am working on a deep learning model for detecting lung cancer from lung CR images. I know that instead of creating a new model from scratch, I can use pretrained models like InceptionV3 for faster training and/or better performance.
But since most of these models are trained on ImageNet data-set, would they prove to be useful for classifying medical image data-sets like lung CR images?
Also, would cutting off/freezing the final layers and training them with my data-set work in this scenario?
Edit: I found a model called as niftynet that is specifically for medical image analysis, but my main question here is whether these popular models could be successfully used for transfer learning of medical image data-sets?
Yes, there are many research that used transfer learning with fine-tuning to gain better result in medical related application including detection and segmentation.
See this article:
https://reader.elsevier.com/reader/sd/pii/S2095809918301887?token=CECF0ACE7A62918720199CB44137B9621DCD3C319535E67F4245557136D179C94790BD5E57175A31D0387E14EFD4F7A9
My answer is yes. Actually, most recently studies used pre-trained model for transfer learning, which could decrease a lot the training time and achieve a better performance. However, as you have mentioned, these pre-trained model were developed on natural images. I have to say that using them is a "no-choice" choice, as there is no a specific pre-trained model for specific medical image such as CT, MR. Hopefully my comment is helpful.
You can, but actually it may not help so much.
Those Image-Net pre-trained networks are mainly trained from natural images. Due to the huge difference between natural images and medical images such as CT/MRI, you have to fine-tune your network, which is known as the 'transfer learning'. Transfer learning may have very limited effect when you switch the dataset from one modality to anther. Note that: Transfer learning in this case is NO BETTER than training from scratch, as the networks learn very different high-level features in these two problems. Reference: 'Rethinking ImageNet Pre-training, Kaiming He'.
Conclusion: Yes, you can use pre-trained networks, but you have to fine-tune and the result is no better than training from scratch.
here is a bit of comparison
https://github.com/talhaanwarch/Eye_Disease_Classification_CNN
check this paper for the purpose of transfer learning with medical imaging tasks: Article Novel Transfer Learning Approach for Medical Imaging with Li...
- Badges
- Science topic