Attention-based Deep Multiple Instance Learning could be applied in a wide range of medical imaging applications. Supported by the project "Deep Learning for Survival Prediction"@UTA-SMILE, I wrote the Keras version of ICML 2018 paper "Attention-based Deep Multiple Instance Learning" (https://arxiv.org/pdf/1802.04712.pdf) in this repo to share the solution for Keras users.
The official Pytorch implementation can be found here. I built it with Keras using Tensorflow backend. I wrote attention layers described in the paper and did experiments in colon images with 10-fold cross validation. I got the very close average accuracy described in the paper and visualization results can be seen as below. Parts of codes are from https://github.com/yanyongluan/MINNs.
When train the model, we only use the image-level label (0 or 1 to see if it is a cancer image). The attention layer can provide an interpretation of the decision by presenting only a small subset of positive patches.
- Colon cancer dataset [Data]
- Processed patches [Google Drive]
I put my processed data here and you can also set up according to the paper. If you have any problem, please feel free to contact me.
Applications
Several applications can be found recently. I will summarize them in the following and the first one is our recent work.
- Jiawen Yao, Xinliang Zhu, et al. "Whole slide images based cancer survival prediction using attention guided deep multiple instance learning networks", Medical Image Analysis, 101789, 2020.[PDF] [Code]
Other important work used multiple-instance learning include
- Ole-Johan Skrede et al., โDeep learning for prediction of colorectal cancer outcome: a discovery and validation study", the Lancet, 2020.
- Used Noisy-AND MIL pooling to get WSI-level representation.
If you have any questions about this code, I am happy to answer your issues or emails (to [email protected]).
I plan to review recent work using Deep MIL techniques in medical imaging and Your suggestions are very welcome !
The work conducted by Jiawen Yao was funded by Grants from the UTA-SMILE Lab.