The following codebase supplements the paper "Learning to Detect Histopathologic Cancer using Attention".

We conduct experiments to investigate the following research questions:

1. Can adding attention to standard vision models improve classification performance?
2. To what extent does attention add to the explainability of the model?

The experiments involve an ablation study on standard vision models, on augmentation techniques, as well as training the final InceptionV3 model with Attention blocks.

To train the final model, run the following command:

python3 train.py --name attn_inception --model attn_inception

The results of the ablation study are summarised below:

To conduct the experiments, modify the get_transforms() function in utilify_fn.py and set the train transform to transforms.ToTensor(). Note that the ablation models were trained with un-augmented data.

python3 train.py --name mobilenet --model mobilenet

python3 train.py --name efficientnet --model efficientnet

python3 train.py --name inception --model inception

python3 train.py --name densenet --model densenet

python3 train.py --name resnet34 --model resnet34

python3 train.py --name convnext --model convnext

The results of the ablation study are summarised below:

The final augmentation policy can be found in utilify_fn.py, and is listed below:

transform = transforms.Compose([
    transforms.RandomHorizontalFlip(),
    transforms.RandomVerticalFlip(),
    transforms.RandomRotation(180),
    transforms.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1),
    transforms.GaussianBlur(3),
    transforms.ToTensor(),
])

To conduct the experiments, modify the get_transforms() function in utilify_fn.py and set the train transform based on these conditions. The experiment can then be started by executing the following command:

python3 train.py --name aug --epochs 1 --model inception

transform = transforms.ToTensor()

transform = transforms.Compose([
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor()
])

transform = transforms.Compose([
    transforms.RandomVerticalFlip(),
    transforms.ToTensor()
])

We test the following angles of rotation: [90, 180, 270]. We can replace i in the below command with each of the aforementioned parameters.

transform = transforms.Compose([
    transforms.RandomRotation(i),
    transforms.ToTensor()
])

We test the following values for brightness, contrast, saturation and hue: [0.1, 0.2, 0.3, 0.4, 0.5]. We can replace i in the below command with each of the aforementioned parameters.

transform = transforms.Compose([
    transforms.ColorJitter(brightness=i, contrast=i, saturation=i, hue=i),
    transforms.ToTensor()
])

We test the following values for the kernel size: [3, 5, 7]. We can replace i in the below command with each of the aforementioned parameters.

transform = transforms.Compose([
    transforms.GaussianBlur(i),
    transforms.ToTensor()
])

To calculate the confusion matrix and classification report of all models in the log directory, run the following command.

python3 metrics.py

torch
torchvision
numpy
pandas
opencv-python
seaborn
matplotlib
scikit-learn