Denoising of low-dose chest computed tomography images using a U-net based convolutional autoencoder and transfer...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Low-Dose Computed Tomography (LDCT) reduces radiation exposure but suffers from noise and artifacts.
• Deep learning (DL) shows promise for LDCT denoising but requires extensive training data.
• Limited clinical datasets pose a significant challenge for developing effective LDCT denoising algorithms.

Purpose of the Study:

• To propose a lightweight and versatile DL algorithm for chest LDCT denoising.
• To address the challenge of limited training data in LDCT denoising.
• To enhance image quality and diagnostic accuracy in LDCT scans.

Main Methods:

• Introduced a U-Net-based Convolutional Autoencoder (UNbCAE) for LDCT denoising.
• Employed a two-step training strategy: initial training on phantom images followed by transfer learning on clinical data.
• Validated the model on the LUNA16 dataset for quantitative and qualitative assessment.

Main Results:

• The UNbCAE model effectively reduced noise in chest LDCT scans while preserving anatomical structures.
• Image quality was enhanced, leading to improved detectability of pulmonary nodules.
• The proposed training strategy achieved performance comparable to or exceeding state-of-the-art denoising techniques, with an average noise reduction factor of (3.4 +/- 0.6).

Conclusions:

• The UNbCAE model offers high-quality LDCT denoising.
• The transfer learning scheme effectively mitigates dataset size limitations.
• The method demonstrates clinical applicability by minimizing dependence on simulated data.