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

• Ge Wang  |  Hongming Shan  |  Ling Sun  |  Mannudeep K. Kalra  |  Qingsong Yang  |  Uwe Kruger  |  Wenxiang Cong  |  Yi Zhang

• Fudan University China

Since CT-associated x-ray radiation poses health risks to patients, low-dose computed tomography (CT) has gotten a lot of attention in the medical imaging sector. Deep-learning-based algorithms, especially convolutional neural network (CNN) and generative adversarial network (GAN) architectures, have recently shown promising results in low-dose CT denoising. For low-dose CT denoising, this model introduces a Conveying Pathbased Convolutional Encoder-decoder (CPCE) network in 2D and 3D configurations within the GAN setting. The idea that an initial 3D CPCE denoising model can be obtained directly by extending a trained 2D CNN, which is then fine-tuned to integrate 3D spatial information from adjacent slices, is a novel feature of this technique. When compared to a training from scratch, the 3D network converges faster and achieves better denoising efficiency using transfer learning from 2D to 3D. It is shown that the 3D CPCE denoising model has a better performance in suppressing image noise and preserving subtle structures when compared to recently published work based on the simulated Mayo dataset and the real Massachusetts General Hospital (MGH) dataset.

Input variables : CT Image
Output Variables : Denoised CT Image

Visit Model : github.com

Additional links : arxiv.org