Support training the classification network on histology datasets. Explore multiple tricks based on stain variations.

Code Organizations

Run main.py to train the models.

Tune all the hyper-parameters in config.yaml.

• train_root: Path to the training set.
• test_root: Path to the test set.
• output_path: Path to the output. Output files will be exported to a folder created in output_path started with the date, hence no worry for overriding.

Dataset

Datasets can be downloadee use preprocess/download.py.

• Normalized-v1: Stain normalized with template NORM-AAAWMSFI.tif (from training set).
• Normalized-v2: Stain normalized with template STR-AAEILWWE.tif (from training set).
• Normalized-v3: Stain normalized with template NORM-TCGA-AASSYQPA.tif (from test set).

To Run the Experinments

Generally, we don't depend on any specific libraries.

If datasets are not on you device, use download.py to download the zip files and unzip in the terminal.

1. Firstly, change the hyperparameters in config.yaml, e.g., train_root pointed to the training set, test_root to the validation set, output_path to the output path where loggings and checkpoints are saved.
2. To train the model, simpily run

python main.py

Methods

1. LabPreNorm: Learnable normalization parameters (i.e., channel mean and channel std) of the template in LAB color space, and use the Reinhard's normalization method.
2. LabEMAPreNorm: Use EMA to update the normalization template. Hyper-parameter: lambda. When lambda=0, degenerates to vanilla Reinhard's normalization method; when lambda=1, degenerates to a speical case of LabRandNorm.
3. LabRandNorm: Randomly select template in each mini-batch, and use the Reinhard's normalization method.

Results

ResNet-18	w/o Pretrain	w/ Pretrain
w/o Norm	64.958	58.788
w/ Norm v1	78.914	78.106
w/ Norm v3	89.624	89.262
w/ RandNorm	88.454
w/ PreNorm	92.549
w/ EMAPreNorm (lambd=0)	91.504