Statistical shape modeling of the diaphragm for application to Rb-82 cardiac PET-CT studies

It is important when motion-correcting Rb-82 cardiac PET-CT scans that diaphragm motion is accounted for, to avoid attenuation-correction artifacts. In the absence of a gated CT, a model of the diaphragm could assist in identifying the diaphragm position in noisy PET images as a step towards performing respiratory-matched attenuation-correction. To test this, a shape model was constructed from a training set of 10 gated CT datasets, in which the diaphragm was segmented. Principal Component Analysis was performed on corresponding landmarks from all surfaces to extract modes of variation in shape and motion between patients. Fitting the model to a segmented surface was then achieved by weighting each mode to minimize the sum of squared differences between the fitted and original surfaces: this was carried out for datasets used in its construction and previously unseen datasets, using a leave-one-out approach. It was found that 95% of training data variation was described in only 5 modes, indicating that 5 parameters need to be fitted in order to fully describe the diaphragm over the respiratory cycle. Model success was measured in terms of the residual differences after fitting and was found to be 3.8 ± 1.0 mm per landmark for the 10 leave-one-out models. Since the slice thickness in the PET data is 3.3 mm, it is likely that this level of error is tolerable in this application. Furthermore, the overall diaphragm shape was reproduced well in the presence of these errors, further indicating the validity of this approach. These results demonstrate the potential of this technique in benefiting the prediction of the time-varying diaphragm position. This could therefore provide a valuable technique in determining diaphragm motion in cardiac PET-CT studies.