<p>In recent years, feature space–based image compressive sensing (ICS) has seen significant progress, largely driven by advances in deep learning. Among these approaches, deep unfolding networks (DUNs), which translate iterative optimization processes into end-to-end trainable neural architectures, have demonstrated impressive reconstruction performance. However, their practical deployment still faces several unresolved challenges. In particular, existing DUNs often suffer from reduced robustness under complex and noisy conditions, feature representation loss caused by insufficient multi-scale modeling, and limited computational efficiency. Many current methods adopt rigid serial structures, which restrict their ability to capture complementary information across scales. To overcome these limitations, we introduce MSJA-Net, a multi-path synergistic joint attention network embedded within the DUN framework. Our design integrates local convolutional operators, global channel attention mechanisms, and serial convolutional branches in a parallel manner, enabling more comprehensive feature extraction. These outputs are adaptively combined through a novel dual-gated multiplicative fusion strategy, which strengthens multi-scale feature representation while mitigating spatial information loss. In addition, a lightweight architectural design enhances efficiency and adaptability, making the method applicable across varied datasets and noise levels. Extensive experiments demonstrate that MSJA-Net consistently surpasses state-of-the-art methods across a variety of datasets, noise conditions, and sampling rates.</p>