3-D Convolutional Encoder-Decoder Network for Low-Dose CT via Transfer Learning From a 2-D Trained Network.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Low-dose computed tomography (LDCT) reduces radiation exposure but degrades image quality.
• Deep learning, particularly CNNs and GANs, shows promise for LDCT denoising.
• Existing methods face challenges in balancing noise reduction and detail preservation.

Purpose of the Study:

• To develop and evaluate a novel 2D/3D deep learning network for LDCT denoising.
• To introduce a transfer learning approach from 2D to 3D for faster convergence and improved performance.
• To demonstrate the efficacy of the proposed CPCE network in suppressing noise and preserving structures.

Main Methods:

• Implementation of a 2D and 3D conveying path-based convolutional encoder-decoder (CPCE) network within a GAN framework.
• Utilizing transfer learning by extending a trained 2D CPCE model to a 3D configuration.
• Fine-tuning the 3D model to incorporate spatial information from adjacent slices for enhanced denoising.

Main Results:

• The 3D CPCE network, leveraging transfer learning, converges faster than training from scratch.
• The proposed model achieves superior denoising performance compared to existing methods on both simulated (Mayo) and real (MGH) datasets.
• The 3D CPCE model effectively suppresses image noise while preserving subtle anatomical structures.

Conclusions:

• The 3D CPCE network offers a significant advancement in LDCT denoising technology.
• Transfer learning from 2D to 3D provides an efficient pathway to high-performance 3D denoising models.
• This approach enhances diagnostic accuracy in low-dose CT imaging by improving image quality.