This Article Abstract Full Text Full Text (PDF) Services Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Crum, W R Articles by Hill, D L G Search for Related Content PubMed PubMed Citation Articles by Crum, W R Articles by Hill, D L G

Non-rigid image registration: theory and practice

Division of Imaging Sciences, The Guy's, King's and St. Thomas' School of Medicine, London SE1 9RT, UK

View larger version (21K): [in a new window]   Figure 1. Schematic showing rigid and non-rigid registration. The source image is rotated, of a different size and contains different internal structure to the target. These differences are corrected by a series of steps with the global changes generally being determined before the local changes.

View larger version (76K): [in a new window]   Figure 2. A more unusual application is registering images of knee joints for the purpose of tracking changes in the thickness of cartilage. The knee is particularly difficult to image consistently in three dimensions on consecutive occasions due to the high degree of mobility around the joint regardless of any disease process. Non-rigid registration, in this case using B-spline based free-form deformations, can recover most of the differences between scans of the same subject acquired at different times.

View larger version (164K): [in a new window]   Figure 3. Non-rigid registration applied to intersubject brain matching. The top row shows two selected slices from a T1 weighted MR-volume of a normal subject. The middle row shows the same slices from a similar image of a different subject. The bottom row shows the result of using non-rigid (fluid) registration to match the second subject to the first. The major neuroanatomical features have been brought into good correspondence. A closer inspection shows that not all of the fine cortical structure has been matched successfully. This is a typical finding when comparing brain images across subjects due to population variation in the geometry of the cortical surface. Note, the left and right views are not of the same scale.

View larger version (121K): [in a new window]   Figure 4. Non-rigid registration applied to lesion detection in contrast-enhanced MR mammography. The subject is scanned at rest and then scanned repeatedly after introduction of a contrast agent. There is often rigid and non-rigid movement when the agent is given. The pre- and post-contrast images appear virtually identical but subtraction reveals many differences (top panel), most caused by motion. Rigid registration (middle panel) reduces the difference between scans significantly. Non-rigid registration using a hierarchical B-spline technique (bottom panel) removes virtually all the artefact associated with motion leaving clear evidence of an enhancing lesion.

View larger version (105K): [in a new window]   Figure 5. Non-rigid registration applied to myocardial segmentation. In this example the myocardium has been manually delineated on each slice of the end-diastolic phase to define a surface. The end-diastolic image volume has been registered to the end-systolic volume to delineate the myocardium at end systole. Technical details: the images are short axis electrocardiogram triggered SSFP SENSE factor 2 images from a healthy volunteer collected on a Philips Intera 1.5 T scanner (Philips Medical Systems, Best, The Netherlands) at Guy's Hospital, London, UK. 20 cardiac phases were acquired with each volume consisting of 12–14 contiguous slices, collected in blocks of three, during up to five breath-holds. Registration was performed with vtknreg (available free from www.image-registration.com), which uses free-form deformations modelled with B-splines.

View larger version (102K): [in a new window]   Figure 6. An example of 1.5 T versus 3 T MRI of the brain. It must be remembered that image acquisition is evolving along with image registration. With new imaging opportunities come new registration challenges. This pair of corresponding slices from rigidly registered brains acquired on two different scanners have exactly the same structure but appear subtly different, despite efforts to match the image acquisition schemes. The 3 T image has higher signal to noise ratio than the 1.5 T image but is also more prone to image artefacts, most obviously here in significant amounts of signal inhomogeneity across the brain (so-called "shading artefact"), and also flow artefacts from the carotid arteries. Registration algorithms driven by intensity information find it hard to differentiate between image differences caused by biological processes and those caused by details of the acquisition process. Image analysis studies that migrate from 1.5 T to 3 T scanners, or which involve aggregation of scans collected from scanners of different field strength are likely to have problems separating real effects from scanner-induced effects.