This technical document assembles and presents a collection of notes
on how one should calculate sensitivity of measures such as model
Specificity and Generalisability, as well as label overlap. These
measures are used, in this particular context, to evaluate non-rigid
registration. Since they are separable families of methods, they ought
to be compared somehow; the eventual comparison must be compact and
easily-comprehended too. We have chosen the notion of sensitivity
as potentially the most valid way of reasoning about these two methods
and comparing them with little bias, if any at all. Derivation of
error bars is included throughout and is also the primary point of
focus of this document, which is intended to provide better understanding
of the current methods.
We begin with a brief introduction to the problem in question and
then describe the variables that we have available. A derivation of
sensitivity is then described in some depth and the corresponding
errors too. Lastly, aggregation of figures that we handle as to make
them more compact, is described somewhat briefly. Bound to that description
is the formulation that we use to propagate the error bars, which
altogether shape our final figure of interest.
PDF version of this document
