Although the automatic technique developed herein successfully detected the intracranial boundary in every test case, the additional analyses and tests listed below are desirable:
These tests have been initiated, but remain incomplete. Initial results, which are not be presented here, are favorable.
One algorithm aspect of the new intracranial boundary detection scheme also requires investigation. As shown in the previous chapter, the Segment Head process does not precisely segment the head. This imprecision, probably caused by high intensity speckle in the MR scans, proved to have no impact on the test results. However, improvement of the head segmentation is likely to render the intracranial boundary detection scheme more robust.
The last area of ongoing work worth mentioning is the integration of the processing steps. Currently, the intracranial boundary detection program consists of three igraphs. Combining these igraphs is a trivial pursuit that does not contribute to the novelty of the algorithm. However, to be completely useful, especially for batch processing, the program should be unified in a single igraph, and eventually compiled into a single executable.
The preliminary goal of reducing low frequency intensity variation in MRI scans has been met by the RF correction method developed herein. Still, some subjects of interest have yet to be considered:
The mask size of the low pass filter in the RF correction process was chosen experimentally to yield ``reasonable'' RF inhomogeneity profiles. Without a true MRI phantom, these profiles are impossible to verify. Therefore, the new method should eventually be verified using MRI data from a scanner with an available phantom.
Although outside the scope of this thesis, preliminary tests show that this new method of RF correction improves the MS lesion segmentation produced by Johnston's algorithm . However, the method necessarily introduces artifacts into the corrected images because an approximate low pass filter is used. Although these artifacts are invisible to the human observer in the corrected scans, they might inhibit automatic segmentation algorithms.
Further, the new method reduces low frequency intensity variation in the MR images regardless of its origin. Low frequency intensity variations that should appear in the corrected data are therefore reduced. This problem introduced no visible artifacts into the observed corrected images, but might have an impact on tissue segmentation algorithms.
The last question on the above list stems from an observation by Dawant et al.  and Zijdenbos . They observed that the intensity of MRI scans varies in the axial direction according to acquisition time rather than spatial location. The cause of this variation is unknown.
Although not mentioned earlier, axial slices are acquired in an interlaced fashion. Indeed, an intensity variation that might track slice interlacing is apparent in the T2-weighted z-dimension profiles of Figures 8.28, 8.31, and 8.34. Further investigation is required to verify slice acquisition order.
With a trivial modification, the newly developed RF correction technique could reduce an intensity variation in the axial direction due to slice interlacing. The modification would simply be to spatially reorder the slices to correspond with their acquisition sequence before homomorphic filtering, then to do the opposite after.