Image Processing Techniques in Emission Tomography - Correction Techniques in Emission Tomography - page 146

(a) Original image (b) LR image (c) HR image

FIGURE 7.15: Artificial super-resolution example. (a) Input image. (b) The
LR images (only one shown) generated from the original image. (c) The HR
image computed from the LR images.

7.7 Validation
When new image processing methods for emission tomography are de-
veloped one question automatically arises: How good is the method? The
word \good" is ambiguous|it can mean the general accuracy, the accuracy
compared to other methods, the computational complexity, or the clinical
usefulness. Validation can be understood as an attempt to answer the above
question.
Validation is essential in many fields of emission tomography. Registra-
tion methods (see Section 7.4), denoising techniques (see Section 7.2), partial
volume correction (see Section 7.5), and reconstruction methods need to be
validated, to name just a few.
In general, every application has its own case-specific validation methods.
However, two universal ways of validation will be discussed in the following.
Firstly, error measures are introduced in Section 7.7.1, as validation can of-
ten be understood as a pixel or voxel-based comparison of images. Secondly,
phantoms providing ground-truth information are discussed in Section 7.7.2.
Another aspect of validation is a direct performance comparison with other
popular methods. When bringing a method into clinical practice this is a
mandatory step. Methods for training and testing, and useful statistical tests
are discussed in detail in [7] and are not part of this section.

7.7.1 Intensity-based measures
A common situation in medical image processing is that images need to be
judged or compared based on their intensity values. The similarity measures