Transfer learning framework for low-dose CT reconstruction based on marginal distribution adaptation in multiscale.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography (CT) dose reduction is crucial for minimizing cancer risks.
• Low-dose CT (LDCT) images suffer from significant noise degradation.
• Developing effective LDCT reconstruction methods is vital.

Purpose of the Study:

• To address the limitations of supervised learning in LDCT reconstruction due to the difficulty of obtaining paired normal-dose CT (NDCT) and LDCT datasets.
• To develop an unsupervised deep learning framework for LDCT reconstruction that does not require paired data.
• To enable accurate LDCT reconstruction without relying on labeled datasets.

Main Methods:

• Proposed an unsupervised learning framework, marginal distribution adaptation in multiscale (MDAM), for LDCT reconstruction.
• MDAM utilizes an identity mapping approach with dimensionality reduction and image reconstruction from low-dimensional features.
• Employed multiscale feature extraction, wavelet decomposition, and domain adaptation learning to align feature distributions between NDCT and LDCT data.

Main Results:

• The unsupervised MDAM framework achieved performance comparable to or exceeding state-of-the-art supervised methods.
• MDAM demonstrated excellent noise suppression, structural preservation, and lesion detection capabilities.
• Quantitative metrics and subjective scores validated the high image quality of MDAM reconstructions.

Conclusions:

• The MDAM framework accurately reconstructs normal-dose CT images from low-dose CT data without requiring paired labels.
• MDAM offers a viable unsupervised alternative to supervised methods for LDCT reconstruction.
• The framework is well-suited for clinical applications due to its unsupervised nature and high performance.