Synthetic CT image generation from CBCT: A Systematic Review

TL;DR

This systematic review assesses the generation of synthetic CT (sCT) images from cone-beam CT (CBCT) data for radiation therapy planning, focusing on deep learning (DL) methods from 2014–2024. It catalogs architectures such as U-Net/CNNs, GANs (including CycleGAN), Transformer-based networks (TransCBCT), and diffusion models (DDPM), and evaluates dosimetric validation metrics across 35 studies. Reported results show that sCT can achieve HU accuracy and image quality approaching planning CT (pCT) benchmarks, with MAE in the tens of HU, RMSE in tens/hundreds of HU, PSNR around 20–67 dB, and SSIM near 0.8–0.99, along with gamma-analysis-based dosimetric concordance in many cases. The review highlights ongoing challenges, including limited dataset sizes, lack of standardization in evaluation, FOV disparities, and integration into clinical workflows, while underscoring the potential of sCT-based DL methods to enable adaptive radiotherapy with reduced CT exposure and improved treatment precision.

Abstract

The generation of synthetic CT (sCT) images from cone-beam CT (CBCT) data using deep learning methodologies represents a significant advancement in radiation oncology. This systematic review, following PRISMA guidelines and using the PICO model, comprehensively evaluates the literature from 2014 to 2024 on the generation of sCT images for radiation therapy planning in oncology. A total of 35 relevant studies were identified and analyzed, revealing the prevalence of deep learning approaches in the generation of sCT. This review comprehensively covers synthetic CT generation based on CBCT and proton-based studies. Some of the commonly employed architectures explored are convolutional neural networks (CNNs), generative adversarial networks (GANs), transformers, and diffusion models. Evaluation metrics including mean absolute error (MAE), root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) consistently demonstrate the comparability of sCT images with gold-standard planning CTs (pCT), indicating their potential to improve treatment precision and patient outcomes. Challenges such as field-of-view (FOV) disparities and integration into clinical workflows are discussed, along with recommendations for future research and standardization efforts. In general, the findings underscore the promising role of sCT-based approaches in personalized treatment planning and adaptive radiation therapy, with potential implications for improved oncology treatment delivery and patient care.

Figures (14)

• Figure 1: Schematic diagram of U-Net for CBCT-to-CT image translation. The U-Net architecture features an encoder-decoder design with skip connections to preserve spatial details. The input CBCT images are compressed by the encoder and then reconstructed back to corresponding CT images. (CT and CBCT images taken from b48.)
• Figure 2: Generative Adversarial Network (GAN) framework of CBCT image synthesis. The architecture includes a generator (G), comprising as multilayer neural network, that generates CBCT images out of random noise. The discriminator (D) is a multilayer neural net, the input will be images, and its targets are binary answers, real or fake. (CBCT image taken from b48.)
• Figure 3: Cyclic GAN architecture for CBCT image synthesis. The architecture includes two generator modules ($G_A$ and $G_B$) to transfer between domains A and B, as well as discriminator modules $D_A$ and $D_B$. The Generators are trained to produce synthetic CBCT images from real ones. The discriminators judge the realism of generated images with a cycle consistency enforced to optimize the model. (CT and CBCT images taken from b48.)
• Figure 4: Transformer-Based Architecture for CBCT to CT image synthesis (TransCBCT). The network leverages self-attention mechanisms, including multi-head and masked multi-head attention, to capture long-range dependencies in CBCT images. Positional encoding is applied to input and output embeddings, allowing the model to synthesize CT images by processing and transforming CBCT inputs through a sequence of attention layers and feed-forward networks. (CT and CBCT images taken from b48.)
• Figure 5: Denoising Diffusion Probabilistic Model (DDPM) for CBCT-to-CT image synthesis. The model employs a diffusion-based framework where Gaussian noise is progressively added to the CT image $\mathbf{x}_{CT}$ over $T$ time steps. This results in intermediate noisy representations denoted as $\mathbf{x}_1, \mathbf{x}_2, \ldots, \mathbf{x}_T$. In the reverse process, the model iteratively estimates and removes noise using a neural network with parameters $\bm\mu$, conditioned on the CBCT image $\mathbf{x}_{CBCT}$. (CT and CBCT images taken from b48.)