Semi-Supervised MVCT Enhancement Using Diffusion Model Refined With KVCT Priors

Abstract

Megavoltage computed tomography (MVCT) on the tomotherapy system has been widely used as a tomographic imaging modality for image-guided radiotherapy. However, the quality of MVCT images is often compromised by poor tissue contrast and significant noise. Conventional networks designed to enhance CT quality typically require the clean ground-truth images, which are not feasible for MVCT. In this study, we introduce a semi-supervised framework named Semi-Diff, which leverages the denoising diffusion probabilistic model and the prior information sourced from kilovoltage computed tomography (KVCT) to address challenges in MVCT enhancement. Specifically, employing a discriminative prior learning method, we first learn a mapping function to estimate MVCT noise and perform state matching. With this state matching dictionary, we then represent the MVCT image as a sample from an intermediate posterior distribution within the diffusion Markov chain, which enables the reverse conditional sampling process of the diffusion model to start directly from the noisy MVCT images. To fully explore the prior information from the plan KVCT images of the same patients, we introduce a novel diffusion base network called RefNet, whose dynamic feature aggregation module can extract and align the relevant features from reference KVCT image to enhance image restoration performance. Quantitative evaluations using simulated digital phantom data show that the proposed Semi-Diff model achieves the average FSIM score of 0.954, PSNR score of 33.22 dB, and RMSE value of 0.023, demonstrating improvements of approximately 2.16% in FSIM, 0.59% in PSNR, and a reduction of 3.58% in RMSE compared to the best-performing baseline method. Results from physical phantom and patient data further validate the model’s superior performance in noise suppression and structural preservation.