RetCCL (This paper is under review)
A better and stronger pre-trained model was built for various histopathological image applications This model outperforms ImageNet pre-trained features by a large margin. We release our best model and invite researchers to test it on your computational pathology tasks.
Hardware
- 128GB of RAM
- 32*Nvidia V100 32G GPUs
Preparations
1.Download all TCGA 32000 WSIs.
2.Download all PAIP 2,457 WSIs. So, there will be about 15,000,000 images(~100T). It costs us $400,000 to advance the progress of digital pathology.
Pre-trained models for histopathological image tasks
This pre-train model is here
1.Classification through search
It is the most obvious and direct way to evaluate the distinctive power of the provided features.
| TissueNet | ||||
|---|---|---|---|---|
| Acc@1 | Acc@3 | Acc@5 | mMV@5 | |
| ImageNet | 50.35 | 77.65 | 87.68 | 46.15 |
| CCL (ours) | 67.09 | 87.81 | 93.4 | 70.1 |
| UniToPatho | ||||
|---|---|---|---|---|
| Acc@1 | Acc@3 | Acc@5 | mMV@5 | |
| ImageNet | 58.17 | 82.89 | 89.45 | 59.01 |
| CCL (ours) | 66.55 | 84.32 | 90.31 | 68.35 |
2.Multiple Instance Learning for Whole Slide Image Classification
This task is currently based on ImageNet pretrained features, which can also verify the superiority of our feature extractor.
| TCGA-NSCLC | ||
|---|---|---|
| Accuracy | AUC | |
| ABMIL | 0.7719 | 0.8656 |
| MIL-RNN | 0.8619 | 0.9107 |
| DSMIL | 0.8058 | 0.8925 |
| TransMIL | 0.8835 | 0.9603 |
| CLAM | 0.8422 | 0.9377 |
| CLAM+CCL (ours) | 0.911 | 0.967 |
3.Classification based on features using SVM
This task follows KimiaNet
| Colorectal cancer dataset | |
|---|---|
| Accuracy | |
| Combined features | 87.40 |
| Fine-tuned VGG-19 | 86.19 |
| Ensemble of CNNs | 92.83 |
| KamiaNet | 96.80 |
| CCL (ours) | 98.40 |
If you want to compute the features.
python get_feature.py
If you want to fine-tune model.
python resnet_lincls.py
Citation
Please use below to cite this paper if you find our work useful in your research.
