Lecture 5: Registration and Fusion Flashcards
5.1 Explain the need for Image Registration and Fusion
In medical imaging, sets of data acquired by scanning the same subject at different times (or modalities) will be in different coordinate systems.
Image registration is the process of transforming the different sets of data into one coordinate system, which is necessary to compare/integrate the medical data obtained from different times, modalities or subjects.
- Eg 2 brain MRI scans taken at different times, one to plan the operation beforehand, and one taken during the operation
Image fusion then combines the information from multiple images and provides more comprehensive info than using individual images
- Image fusion = registration + combination into a single representation
- Eg CT + PET
5.2 What are some of the advantages of multi-sensor data registration and fusion?
- Improved system performance
- Improved detection, tracking and identification
- Improved situation assessment and awareness
- Improved robustness
- Extended spatial and temporal coverage
5.3 What are some of the applications of registration and fusion?
- Cancer staging
- Biopsy planning
- Radiotherapy treatment planning
- Assessment of treatment response
- Image-guided surgery (IGS)
- Growth monitoring
5.4 What is the goal of Monomodal Image Registration?
To register serial scans from the same imaging technique to temporal studies to evaluate disease progression, assess treatment response, detect changes and monitor growth and development.
Eg in multiple sclerosis lesion development monitoring
5.5 What is the goal of Multi-modal Image Registration?
Combine complementary info from different imaging modalities and to correlate anatomical structure with functional info, eg CT-PET, CT-MRI, MRI-PET, etc.
5.6 Describe the Framework of Medical Image Registration
Involves a number of key components: transformation, interpolation, criterion and optimisation.
Registration Optimisation - uses an optimisation algorithm as a searching strategy.
Registration Interpolation - required when an image needs transformations. Method types include nearest neighbour, linear, cubic, cubic spline, sinc function, etc. The more complex the interpolation methods, the more surrounding points concerned and the slower the registration speed
Registration Transformations - to relate the pixels of the moving/study/floating image to the corresponding pixels of a fixed/reference/target image - this transfers the moving image to the coordinate system of the fixed image. (see 5.7)
5.7 Describe some of the types of Registration Transformation methods.
Rigid transformation - preserves lengths and angle measures, often used to correct translation and rotation displacements. (see 5.8)
Affine transformation - maps parallel lines to parallel lines, can be used to correct skewing distortion introduced by eg a tilted gantry in CT
Deformable/elastic/nonlinear transformation - more complex, corrects dramatic deformations caused by changes of tissue structures, differences of volume and shape of organs
5.8 When might and might not Rigid Transformation be appropriate? What may be useful in the latter case?
Eg in brain image registration. More appropriate in intra-subject registration. Fails in:
- Matching atlas to patients
- Inter-subject registration
- Presence of deformations (tumour, craniotomy)
Elastic, nonlinear deformation (transformation; deformable registration) can be used to solve the above^
Non-rigid registration often required to accommodate anatomical variability across individuals. Can be represented as a dense deformation field using displacement vectors associated with homologous structures (relying on extracted features - points, curves, surfaces and requires some interpolation).
This transformation is mostly relevant in the neighbourhood of the homologous features.
5.9 Describe Feature-based Registration
Uses points/landmarks, contours and surfaces
Procedure involves pre-processing (segmentation; extraction of features), registration and verification
Main advantage: transformation can be formulated in an analytic form with efficient computational schemes
Slight disadvantage: pre-processing step of extracting the features is needed - success is dependent on the results of this
See the following:
- Landmarks 5.10
- Spline 5.11
- Fiducial Markers 5.12
- Surface Registration 5.13
5.10 Describe Feature-Based Registration: Landmarks, and the qualities that define a good Landmark
Point-based registration involves identifying corresponding points in the images to be registered, registering the points, and inferring the image transformation.
Corresponding points are called homologous landmarks to emphasise that they should represent the same feature in the different images.
Landmarks should:
- Have uniquely defined position
- Carry substantial image info
- Be well-suited for user-interaction
- Facilitate efficient registration
- Have salient, prominent characteristics
- Be scattered over the image
Once the corresponding landmarks have been decided, Iterative Closest Point (ICP) or Thin-plate Spline technique can be used to register the images
5.11 Describe Feature-Based Registration: Spline. What are the benefits of this technique?
Spline is referred to as a long flexible strip of metal which takes a least bent shape along the spline. This bent spline can be used to model the surface of deformed objects
Benefits:
- Can produce a smooth spline interpolation
- Has high computation speed
- Can correct local elastic deformations by mapping the study landmark points to the corresponding points in the reference image
5.12 Describe Feature-Based Registration: Fiducial Markers
An object/mark placed in the field of view of an imaging system which appears in the image produced, for use as a point of reference or measure.
- Non-invasive - mould, frame, dental adapter, skin markers
- Invasive - screw markers, stereotactic frames
Often easily automated, since the marker objects are designed to be well visible and detectable in the images involved.
Widely used in IGS because of no need of complex optimisation and computation of registration parameters
5.13 Describe Feature-Based Registration: Surface Registration
Hat and Head method:
- Hat surface is a sin surface from low-resolution PET scan
- Head surface is a stick of skin contour from high-resolution CT/MRI scans
- Two segmented surfaces visualised in 3D computer graphic system and aligned by minimising the mean square distance between them
Con: algorithm prone to finding wrong solution and failing when surfaces show symmetries to rotation
5.14 Discuss Intensity-based Registration; main methods and limitations.
Registration transformation determined by iteratively optimising a certain similarity measure calculated from pixel values. Similarity measures include:
- Minimising the intensity difference
- Correlation techniques
- Variance of Intensity Ratio (VIR)
- Information theoretic techniques (Maximisation of Mutual Info)
Simplest scheme involves (1), including Sum of Squared Differences (SSD) and the Sum of Absolute Differences (SAD), which exhibit a minimum in the case of perfect matching.
- CON: SSD and SAD are sensitive to intensity changes, limiting their application to monomodality image registration
Correlation Coefficient (2) - for multimodal image registration
- Cross-correlation - rigid motion correction of SPECT cardiac images.
- CON: Correlation methods require a linear dependence between intensity of the images. Because geometric deformations of image modalities are unlikely to be linear, this limits reliable registration results.
5.15 What are some of the Applications of Registration and Fusion to Biomedical Images? Compare Mono- and Multi-modalities, and Intra- and Inter-subject examples.
See attached table
