Our results show that when image sequences have strong frame-to-frame correlation, the savings from long-term memory prediction do not outweigh the cost of sending the additional time buffer displacements, since the previous frame is selected with a significant frequency when compared to the previous image at the same phase in the periodic cycle (in our case, the heart cycle). Additionally, when a strong frame-to-frame correlation is coupled with motion limited to a very small region of the image or of very small incremental change from frame to frame (as is the case in our cardiac sequences), the best performance at low bit rates is achieved from simple frame-difference encoding without the use of MCP. The encoding of any motion vectors in this case tends to consume more bits at low bit rates than saved by improving prediction.

The cardiac sequences studied are created from data that shares a significant percentage between successive frames. Strong correlation between successive frames could also occur when an image sequence has very small temporal distance between time samples (i.e., has a large frame rate.) As the temporal distance between frames is increased, the correlation between successive frames is decreased and the savings from long-term memory MCP will increase.

Although long-term memory MCP has shown significant gains with other image sequences, in the specific application of real-time cardiac MRI studies, we conlude that the possible gains achievable with long-term memory MCP and single-frame MCP are very slight over simple frame-difference compression. If the sequences are to be compressed in real-time, the relative computational simplicity and speed of frame-difference encoding would almost certainly outweigh the compression gains of MCP. Bit rates achievable are well within the range for real-time streaming across widely available channels such as DSL.