Expert-attention guided deep learning for medical images

Get Started

Pip install the PYPI distro:

pip install expert-informed-dl

Here's an example of how to use the trained model for inference (with subimages)

Check out eidl/examples/subimage_example.py for a simple example of how to use the trained model for inference on subimages.

from eidl.utils.model_utils import get_subimage_model

subimage_handler = get_subimage_model()
subimage_handler.compute_perceptual_attention('9025_OD_2021_widefield_report', is_plot_results=True, discard_ratio=0.1)

If you want to use the rollouts/gradcams in a user interface, you may consider precomputing them, as it can be slow to compute them on the fly.

from eidl.utils.model_utils import get_subimage_model

subimage_handler = get_subimage_model(precompute='vit')

# or

subimage_handler = get_subimage_model(precompute='resnet')

# or

subimage_handler = get_subimage_model(precompute=['vit', 'resnet'])

If you don't want to use subimages:

Check out eidl/examples/example.py for a simple example of how to use the trained model for inference.

When forwarding image through the network, use the argument collapse_attention_matrix=True to get the attention matrix to get the attention matrix averaged across all heads and keys for each query token.

y_pred, attention_matrix = model(image_data, collapse_attention_matrix=False)

Train model locally

Install requirements.txt

Download Pytorch matching with a CUDA version matching your GPU from here.

Run train.py

For example, if you have 32 * 32 patches, the attention matrix will be of size (32 * 32 + 1) 1025. Plus one for the classificaiton token. If you set collapse_attention_matrix=False, the attention matrix will be uncollapsed. The resulting attention matrix will be of shape (n_batch, n_heads, n_queries, n_keys). For example, if you have 32 * 32 patches, one image and one head, the attention matrix will be of shape (1, 1, 1025, 1025).

Troubleshoot

If get model functions raises the following error:

RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.

You will need to install the correct version of Pytorch matching with a CUDA version matching your GPU from here. This is because all the models are trained on GPU.