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

Abstract Background and Purpose: X-ray-induced Acoustic Computed Tomography (XACT) holds significant promise for real-time radiotherapy dose verification by leveraging the linear relationship between acoustic pressure and deposited radiation dose. However, clinical implementation faces significant challenges from highly sparse data acquisition, constrained by equipment cost, patient safety, and acquisition time, resulting in severe reconstruction artifacts that compromise diagnostic utility. Methods: We introduce MFDRNet, a novel Multi-view Fusion and Disentangling Reconstruction Network that addresses sparse-view XACT reconstruction through three core components: (1) Multi-View Masking Modeling (MVM) that exploits structural redundancy across multiple sparse sampling perspectives, (2) Multi-Scale Attention Multi-View Fusion (MSA-MVF) that employs a tri-branch attention mechanism to integrate features, and (3) Artifact-Disentangling Block (ADB) that explicitly separates content from artifacts. The framework operates under a self-supervised learning paradigm, eliminating the dependency on ground truth annotations while achieving effective sparse signal reconstruction. Results: Extensive experiments on phantom, animal, and human datasets demonstrates substantial performance improvements over established methodolo-gies. Under extreme sparsity conditions (16-channel sampling representing 94% data reduction), MFDRNet achieves 29.37 dB PSNR and 0.9580 SSIM on phantom data, yields up to 400% PSNR improvement over traditional Time-Reversal 1 methods in clinical scenarios. Ablation studies confirm the additive contributions of all architectural components, with progressive performance enhancement across module additions culminating in a 10.92 dB PSNR gain over baseline approaches. Conclusion: MFDRNet successfully enables high-quality XACT image reconstruction under severe sparse sampling constraints while preserving critical anatomical structures and diagnostic contrast essential for radiotherapy dose verification. Its strong generalization across domains and demonstrated clinical viability position this framework as a viable solution for practical XACT implementation in precision radiotherapy guidance.