A neural network adaptive approach to decomposition of patient stereo-infrared tracking data for motion using asymmetric median filters during cardiac SPECT imaging

Patient motion during cardiac SPECT imaging can cause diagnostic imaging artifacts. We have implemented a neural network (NN) approach to decompose monitored patient motion data, gathered during cardiac SPECT imaging, using the Polaris stereo-IR real-time motion-tracking system with stated 0.35 mm accuracy and 0.2 mm repeatability. In our earlier work we obtained excellent correlation (R2>0.998) of axial and vertical target position as measured by our SPECT system and Polaris. We also correlate with volunteer respiration data from a pneumatic bellows similarly wrapped around the abdomen. Herein, we show the successful decomposition of Polaris motion data with our NN approach into X (vertical), Y (lateral), and Z (axial) rigid body motion (RBM) and respiratory motion (RM). Tests have been done with volunteers (RBM and RM), phantoms (RBM only since phantoms don\´t breathe) using SPECT table motions, and simulated "signal known exactly" data (RBM and RM) with known RBM steps and RM frequencies, A phantom test with 19 RBM steps, motion between each SPECT projection, while monitoring radioactive pt sources was decomposed into RBM and RM, and used to correct RBM motion in SPECT images with excellent results. Our NN provides for setting RBM threshold limits (e.g. 0.4 mm, 1 mm, 1.4 mm, 2 mm etc.) to establish classification criteria. The "PRE", "FORWARD (Post)" and "GAPPED" asymmetrical median filter window widths (#scans) or periods (#scans*scan rate) utilized in the NN logic are initially set to a typical "fixed" average respiration period width (e.g. 77) determined from the raw data for the 1st NN run. This decomposes the data into RM and RBM. The RM is then fed into an FFT algorithm producing a scan by scan respiratory period output file to be used during a 2nd NN run in "variable" mode, where it can "adapt" to patient respiratory rate changes. Decomposition has been easily demonstrated at a threshold level of 1 mm RBM within RM of 40 mm, thus S/N level of 1/40=0.02- - 5 or N/S=40. Implementation was in the UNIX environment with Interactive Data Language (IDL)