Dose reduction and image enhancement in micro-CT using deep learning | JoVE Visualize

Area of Science:

Medical Imaging
Preclinical Research
Artificial Intelligence in Radiology

Background:

Micro-computed tomography (CT) is vital for high-resolution rodent imaging in preclinical research, enabling non-invasive assessment of disease and therapy.
Achieving human-equivalent resolution in rodents necessitates higher doses, raising concerns about cumulative radiation effects in longitudinal studies.
Dose reduction is critical under ALARA principles, but low-dose CT yields noisy images that impair diagnostic performance.

Purpose of the Study:

To investigate the efficacy of convolutional neural networks (CNNs) for denoising low-dose micro-CT images in preclinical settings.
To develop and evaluate novel CNN frameworks utilizing realistic noise models for training, matching noisy low-dose images to clearer high-dose counterparts.
To address the limited research on deep learning for preclinical CT denoising compared to clinical applications.

Main Methods:

Acquisition of low and high dose ex vivo micro-CT scans from 38 mice.
Training of two CNN models (2D and 3D U-Net architectures) using mean absolute error loss.
Comparative analysis against spatial filtering and iterative reconstruction methods, validated with phantom data and observer studies for image quality and dose reduction estimation.

Main Results:

Both 2D and 3D CNN models demonstrated superior noise suppression, structural preservation, and contrast enhancement compared to traditional methods.
An observer study with medical imaging experts ranked the 2D CNN approach as the best performing denoising method.
Quantitative measurements and a second observer study suggested a 2-4x dose reduction potential, with the 2D CNN achieving an estimated factor of approximately 3.2.

Conclusions:

Deep learning, specifically CNNs, shows significant potential for improving micro-CT image quality in low-dose acquisition scenarios.
This advancement offers a promising strategy to mitigate cumulative radiation effects in preclinical longitudinal studies.
The developed denoising techniques can facilitate higher quality imaging while adhering to dose reduction principles in animal research.