CBCT-IDDNet: a three-dimensional Res-UNet based image domain denoising network for clinical dose cone-beam computed tomography-winner of the International Conference on Acoustics, Speech, and Signal Processing-2024 challenge.

Background: While deep learning techniques have gained increasing prominence in cone-beam computed tomography (CBCT) reconstruction, the denoising of clinical dose CBCT images remains relatively underexplored. This study aimed to propose and evaluate a CBCT image domain denoising network, named CBCT-IDDNet, which won first place in the clinical dose task of the three-dimensional (3D) CBCT Challenge at the 2024 International Conference on Acoustics, Speech, and Signal Processing (ICASSP). Methods: The proposed CBCT-IDDNet model was based on a 3D Res-UNet architecture and consisted of two components: an encoder and a decoder. The CBCT-IDDNet was trained on data from 800 patients, validated on 100 patients, and tested on 110 patients. On the validation set, the CBCT-IDDNet model was evaluated against five state-of-the-art (SOTA) models using quantitative metrics, including mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM). On the test set, the model was compared with six competing models submitted by other participants. Results: On the validation set, the CBCT-IDDNet model achieved averaged MSE, PSNR, and SSIM values of 0.00089, 38.30 dB, and 0.9486, respectively, outperforming all five SOTA models. On the test set, the model achieved an average MSE of 0.00077, prevailing over all six competing models. Conclusions: The proposed CBCT-IDDNet model improves CBCT image quality by effectively reducing noise, making it a valuable tool for clinical applications.