MFDRNet: A Self-Supervised Framework for Sparse-View XACT Image Reconstruction via Multi-View Fusion and Artifact Disentanglement

X-ray-induced Acoustic Computed Tomography (XACT) offers a promising approach for real-time radiotherapy dose monitoring by mapping absorbed radiation dose through thermoelastic acoustic responses. However, its clinical application remains limited by extremely sparse signal acquisition, which causes severe image degradation. To address this challenge, we propose MFDRNet, a self-supervised reconstruction framework that integrates multi-view fusion and artifact disentanglement to recover high-quality XACT images from sparsely sampled data. The network employs a multi-view masking strategy to simulate diverse sparse sampling configurations from a single acquisition, a Multi-Scale Attention-based Multi-View Fusion (MSA-MVF) module to integrate structural information across views, and an Artifact Disentanglement Block (ADB) to explicitly separate anatomical content from sampling-induced artifacts. Trained with image-domain consistency constraints, MFDRNet effectively enhances reconstruction fidelity without requiring paired ground-truth data. Experiments on phantom, animal, and clinical human datasets demonstrate that MFDRNet consistently outperforms conventional and deep learning methods under severe sparsity , achieving 29.37 dB PSNR and 0.9580 SSIM on phantom data and maintaining structural realism across modalities. Each 256×256 reconstruction requires approximately 0.24 seconds on an NVIDIA RTX 3090Ti GPU, indicating near real-time feasibility. Although this study primarily evaluates image quality, future work will focus on validating dosimetric accuracy to support clinical translation in precision radiotherapy.