Pain is an unpleasant sensory and emotional experience involving multi-level neural processing, with a highly complex neural activity pattern. Recent advancements in non-invasive brain functional imaging techniques have enhanced our understanding of the neural mechanisms underlying pain processing in humans at the whole-brain level. Functional magnetic resonance imaging (fMRI), in particular, plays an important role due to its high spatial resolution and has driven significant advancements in this field. This review focused on fMRI studies of pain in humans. We first summarized research that explored brain responses to pain and showing that pain processing involves neural activities across multiple brain regions, constituting the pain matrix, which includes the somatosensory cortex, thalamus, insula, anterior cingulate cortex, and other areas. However, modulating the activity of a single brain region has limited effects on pain experiences, suggesting that pain processing entails communications among multiple brain regions. Thus, we reviewed research investigating interactions between brain regions, finding that multiple neural pathways spanning the whole brain are involved in pain processing. Beyond interactions between pairs of regions, understanding how these interactions construct a pain-related network is crucial for fully comprehending the neural representation of pain. Two main approaches are used to describe neural networks across the whole brain. The first one is theory-driven, such as graph theory. Using this method, researchers explored how network properties evolve during pain processing and identified a tightly connected network that emerges during pain, encompassing the somatosensory, salience, and fronto-parietal networks, forming a pain-related super-system. As pain is modulated or diminishes, this system becomes less connected. The second approach relies on data-driven methods, such as methods based on independent component analysis or principal component analysis, and machine learning. These methods are not constrained by pre-defined brain networks. Advancements in machine learning have provided valuable insights, enabling researchers to develop pain biomarkers with promising clinical potential. Theory-driven and data-driven approaches provide complementary insights into our understanding of the neural mechanisms of pain. In recent years, two rapidly advancing and promising techniques have further enhanced the precision and comprehensiveness of pain neural network. One is ultra-high-field magnetic resonance imaging, and the other is simultaneous brain-spinal imaging. Ultra-high-field magnetic resonance imaging has overcome previous spatial resolution limitations in fMRI. In subcortical regions, it helps distinguish neural activities of different nuclei. In cortical regions, high resolution enables the differentiation of neural activities across cortical layers, thereby providing a more in-depth and detailed understanding of the neural mechanisms of pain. Simultaneous brain-spinal imaging technology enables the exploration of brain-spinal networks involved in pain processing, making it possible to construct a comprehensive neural network representation of pain throughout the entire central nervous system. Based on current findings, we suggested that in the clinical treatment of pain using neuromodulation techniques, the selection of stimulation targets could be guided by the pain neural network. Targeting hubs within the pain network could significantly impact the network and may efficiently influence pain experiences. Finally, we discussed the limitations of current research on the neural representation of pain and proposed future directions, including exploring pain-specific representation, systematically comparing experimental and clinical pain, and examining individualized neural representations.