Denoise DICOMs using pretrained models

Prerequisite

This example assumes you have the package ldct-benchmark installed. Please refer to Getting Started for instructions on how to do this.

The pretrained models provided as part of the model hub can be used to denoise any CT DICOM dataset. See here for a list of all available algorithms.

Let's use two of these models, RED-CNN1 and DU-GAN2 to denoise DICOM slices from the CT data of the NLM Visible Human Project3 which can be downloaded from NCI Imaging Data Commons.

Warning

This is an out-of-distribution setting as data of the Visible Human CT Dataset were acquired with a (31 year old) scanner and scan-protocols that are far from the training data distribution. The results of the models on such data should be interpreted with caution.

Start by importing some modules:

# Make sure that s5cmd is installed! (pip install s5cmd)
import os
import torch
from ldctbench.hub import Methods
from ldctbench.hub.utils import denoise_dicom

We'll download 10 slices of the female pelvis data to a new folder ./visible-human/orig:

# Create folders
folder = "./visible-human"
orig_data = os.path.join(folder, "orig")

if not os.path.exists(orig_data):
    os.makedirs(orig_data)

# Filenames of 10 pelvis slices
files = [
    "496788de-f0f0-41fd-b19a-6da82268fd0a.dcm",
    "a535613b-de28-4080-850a-f5647ee33c96.dcm",
    "9f7ef52e-c93d-430a-9038-970a47e95e3a.dcm",
    "0c7ac013-41e3-404f-9081-9e0cc18f4f67.dcm",
    "2591aad8-7673-4a12-98e0-8984dafa5175.dcm",
    "28494e9b-d274-4310-a0ed-15d4220e1dc1.dcm",
    "f5e41514-d30a-4cef-81ac-fce50b4743d8.dcm",
    "21292f8c-072c-4223-859a-1e70bbc87a42.dcm",
    "5a214c6b-6898-43c1-89f3-52c967dff39e.dcm",
    "cd90f914-2b13-4cd9-9119-976a3c5721c1.dcm",
]

# Download the data
for file in files:
    os.system(
        f's5cmd --no-sign-request --endpoint-url https://s3.amazonaws.com cp "s3://idc-open-data/b9cf8e7a-2505-4137-9ae3-f8d0cf756c13/{file}" visible-human/orig'
    )

The function ldctbench.hub.utils.denoise_dicom can be used to apply a pretrained model either to a single DICOM file or a folder containing multiple DICOM files. The processed DICOMs differ only in the PixelData, all other DICOM tags are identical to those in the source files. We'll use this function to apply the RED-CNN and DU-GAN models to all 10 slices we just downloaded:

# Apply RED-CNN and DU-GAN and store the processed DICOMs
# to ./visible-human/redcnn and ./visible-human/dugan
for method in [Methods.REDCNN, Methods.DUGAN]:
    denoise_dicom(
        dicom_path=orig_data,
        savedir=os.path.join(folder, method.value),
        method=method,
        device=torch.device("mps"),  # Use "mps" for Apple Silicon, "cuda" for NVIDIA GPUs or "cpu" for CPU
    )
The denoised DICOMs can be loaded with any DICOM viewer. Below we show a comparison of the original and denoised images:

Denoised DICOMs
Ten slices of the NLM Visible Human Project denoised using RED-CNN and DU-GAN. Original data is courtesy of the U.S. National Library of Medicine.

Here we find that both RED-CNN and DU-GAN reduce the noise in the images. Additionally, it can be observed that RED-CNN smooths the images more than DU-GAN does which can be attributed to the fact that DU-GAN is trained in an adversarial fashion, whereas RED-CNN is trained with a simple mean squared error loss.

  1. H. Chen, Y. Zhang, M. K. Kalra, F. Lin, Y. Chen, P. Liao, J. Zhou, and G. Wang, “Low-dose CT with a residual encoder-decoder convolutional neural network,” IEEE Transactions on Medical Imaging, vol. 36, no. 12, pp. 2524–2535, Dec. 2017 ↩

  2. Z. Huang, J. Zhang, Y. Zhang, and H. Shan, “DU-GAN: Generative adversarial networks with dual-domain U-Net-based discriminators for low-dose CT denoising,” IEEE Transactions on Instrumentation and Measurement, vol. 71, pp. 1–12, 2022. ↩

  3. M. J. Ackerman, "The Visible Human Project," in Proceedings of the IEEE, vol. 86, no. 3, pp. 504-511, March 1998, doi: 10.1109/5.662875. ↩