Breath-hold CBCT-to-CT synthesis using an unsupervised artifact disentanglement network with Mamba for breast cancer
Zhiqun Wang1, Yiwen Zhang2, Xiangyin Meng1
1Department of Radiation Oncology, Chinese Academy of Medical Sciences, Peking Union Medical College, Beijing, China.
Background:
Accurate and up-to-date anatomical information is critical for effective treatment planning in breast cancer adaptive radiotherapy. Cone-beam computed tomography facilitates real-time plan optimization but lacks sufficient electron density accuracy for direct clinical application. To address this limitation, we propose a novel unsupervised deep learning framework that integrates the Mamba architecture with an artifact disentanglement network to form the Artifact Disentanglement Network-Mamba model.
Methods:
This study proposes an unsupervised deep learning framework, ADN-Mamba, integrating an Artifact Disentanglement Network (ADN) with the structured state-space model Mamba for high-precision sCT synthesis from breath-hold CBCT (BH-CBCT). The model uses three encoders (CBCT content, CT content, artifact) and two generators to disentangle anatomical features from artifacts in CBCT. Mamba enhances the ability of the model to capture long-range dependencies, improving representation of complex anatomical structures.
Results:
The Artifact Disentanglement Network-Mamba model achieved a mean absolute error of 54.97 HU within the body. The mean absolute percent errors of synthetic and real CT images in the soft tissue (-150 HU to 150 HU) and bone (200 HU to 1500 HU) regions were 46.26% and 30.98%, respectively. The gamma pass rate of the calculated dose on sCT compared with that on pCT is 97.74% under the 2%/2 mm criterion. The proposed model outperforms six other state-of-the-art methods in terms of image quality, dose accuracy, and radiomic feature consistency. By overcoming challenges such as registration errors and the absence of paired cone-beam computed tomography-computed tomography datasets, the proposed framework demonstrated superior performance in terms of anatomical fidelity and dose calculation accuracy.
Conclusions:
ADN-Mamba enables precise BH-CBCT-to-CT synthesis via unsupervised artifact disentanglement and Mamba's long-range modeling, demonstrating superior performance in image quality, dose calculation accuracy, and radiomic consistency. This framework provides a reliable tool for online dose calculation and target delineation in breast ART. Future work will focus on extending the model to 3D data and multicenter validation.
