Traditional Inverse Synthetic Aperture Radar (ISAR) imaging algorithms frequently struggle with low resolution and poor noise suppression when operating under sparse aperture (SA) conditions, primarily due to the inherent non-uniformity and incompleteness of the collected data. In recent years, unfolding networks have emerged as a promising solution, establishing a systematic connection between conventional iterative algorithms and deep learning to achieve fast and efficient ISAR imaging. However, a common limitation of these unfolding networks is their reliance on a fixed number of layers for parameter optimization, which inherently limits their adaptability across diverse scenarios. To address these limitations, we propose an ISAR imaging unfolding network, termed DPSCS-Net, which is based on dynamic path selection and compressed sensing (CS). Specifically, DPSCS-Net unfolds a CS optimization algorithm into a deep network while simultaneously integrating phase error estimation within the CS framework. This integration significantly enhances cross-range resolution. To overcome the fixed-layer constraint of traditional unfolding networks, DPSCS-Net innovatively incorporates a dynamic path selector. This selector generates a one-hot vector based on the input data, where elements (0 or 1) dictate the selection of a fast and appropriate processing path. This dynamic execution mechanism effectively mitigates the unnecessary computational burden associated with conventional deep unfolding networks (DUNs), where all sampled data must traverse every processing stage. Both simulated and measured experimental results demonstrate that our DPSCS-Net exhibits notable flexibility and delivers superior performance with dynamic adaptability. This allows for an optimal balance that meets the primary demands of practical ISAR imaging applications. The network consistently shows strong robustness across various scenarios, coupled with superior performance and focusing accuracy.