Model Hub

We provide pretrained models that can be loaded via

from ldctbench.hub import load_model
net = load_model("<method>")

where <method> is one of the method names provided in this table (e.g., cnn10, redcnn, wganvgg, ...) or a member of ldctbench.hub.Methods (e.g., ldctbench.hub.Methods.CNN10). To apply it, make sure that CT images are stored with an offset of 1024, i.e. air has a value of ~24.

import numpy as np
from ldctbench.hub import load_model, Methods
from ldctbench.utils import preprocess, denormalize

method = Methods.RESNET # method="resnet" also works
# Setup model
net = load_model(method)
# Define image
x = # ... some numpy array of shape [1, 512, 512] you wish to denoise
# Preprocess and normalize input
x_t = preprocess(x, method=method)
# Apply network
y_hat = net(x_t)
# Denormalize
y_hat = denormalize(y_hat, method=method)

We provide a comprehensive example on how to denoise DICOM data using the pretrained models in this example.