Synchrotron-based computed microtomography is an efficient tool for investigating pore-scale dynamics of multiphase flow in porous media. It provides temporal resolution down to several seconds per 3D image. However, the multiphase flow in porous media has a sporadic nature and happens at subsecond timescales, resulting in motion artifacts in tomographic reconstructions. We address this limitation by extracting information from tomographic projections that capture images of the hydrodynamic system with a resolution of tens of milliseconds. We propose using a neural network to detect pore-scale displacement events by analyzing the differences between successive projections. Our neural-network model was trained on a synthetic dataset representing low-contrast tomography data. The model was applied to synchrotron tomographic data of methane gas-hydrate formation within a crushed coal sample where cryogenic suction caused rapid sporadic fluid-flow events exhibiting dynamics reminiscent of Haines jumps. These events, with durations ranging from 100 ms to over 2 s, were separated by stabilization intervals of various lengths. We observed heterogeneity in the spatial distribution of the fast-flow events, which predominantly occurred near metal inclusions. The approach outperformed traditional segmentation methods, especially in conditions of low material contrast, low flow rate, or small pore sizes. Overall, the proposed approach serves as an effective tool for identifying the specific time intervals of pore-scale dynamics during extended tomographic imaging. The resulting data provides insights into local flow regimes as well as their characteristic temporal and spatial scales.