Research on obtaining pseudo CT images based on stacked generative adversarial network

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Image Synthesis
• Radiotherapy Planning and Simulation

Background:

• Computed tomography (CT) is crucial for radiotherapy planning, but ultrasound (US) imaging is more accessible.
• Synthesizing pseudo CT images from US images could bridge this gap, improving radiotherapy guidance.
• Investigating the feasibility of using stacked generative adversarial networks (sGAN) for this synthesis.

Purpose of the Study:

• To evaluate the efficacy of a stacked generative adversarial network (sGAN) in generating pseudo computed tomography (CT) images from ultrasound (US) images.
• To assess the accuracy and quality of synthesized pseudo CT images for potential use in radiotherapy.
• To compare the performance of the proposed sGAN method against existing techniques like Neural Style Transfer (NSF) and CycleGAN.

Main Methods:

• A two-stage sGAN model was developed, utilizing pre-radiotherapy US and CT images from 75 cervical cancer patients.
• The first stage generated low-resolution pseudo CT images from US images using a conditional GAN.
• The second stage employed a super-resolution GAN to enhance image quality, texture, and grayscale accuracy. Five-fold cross-validation and comparison with NSF and CycleGAN were performed. Dosimetric accuracy was verified using phantom experiments.

Main Results:

• The sGAN method achieved mean absolute error (MAE) values between 66.34±1.75 HU and 67.26±2.37 HU compared to real CT images.
• Statistically significant improvements in normalized mutual information (NMI), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR) were observed compared to NSF and CycleGAN.
• Dice similarity coefficient (DSC) indicated higher similarity of sGAN-generated pseudo CT images to ground truth CT for organs at risk. Phantom experiments confirmed similar dose distributions.

Conclusions:

• The sGAN method demonstrates superior performance in synthesizing accurate pseudo CT images from US data compared to NSF and CycleGAN.
• This technique offers a viable new approach for image guidance in radiotherapy, particularly for cervical cancer.
• The synthesized pseudo CT images maintain dosimetric accuracy, supporting their clinical utility.