184 IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 60, NO. 1, FEBRUARY 2013
Let the channel outputs for classes 1 and 2 b e denoted as
and , respectively. Throughout this work , we a ssum e
that the channel output vector follows a multiv ariate normal dis-
tribution under each class with equal covariance matrices, i.e.,
and .(Ifa random
vector
follows a m ultivariate normal distributio n with
mean
and covariance matrix ,wewrite .) In
this setting, the CHO template becomes
,andthe
CHO is op tim al among all observers that operate on the channel
output vector [2, p. 851].
The performance of an observer on a binary classification task
is fully characterized by the observer’s r eceiver operating char-
acteristic (ROC) cur ve, which plots true positive fraction (TPF)
as a function of false positiv e fraction (FPF) [2], [3 5]. One figure
of merit that is commonly used for ROC evaluations is the area
under the ROC curve, denoted as AU C. The AUC can be in-
terpreted as the average T PF, averaged uniformly over all FPF
values [35]. Alternatively, if the o nly pertinent FPF values are
in the range
, then the partial area under the ROC
curve, defined as
(1)
can be a useful figure of merit [35]. The pAUC may be inter-
preted as the TPF averaged over the FPF values between
and .
Let
and be the cumulati ve distribution function
(cdf) and the inverse cdf, respectively, for the standard normal
distribution,
. Under our distributional assumptions for
the channel outputs, the ROC curve for a CHO takes the special
form [35, Result 4.7, p. 82]
(2)
where SNR is the observer signal-to-noise ratio, defined as the
difference of class means for
divided by the pooled standard
deviation [2, p. 819]. In our setting, where the CHO rating
statistic,
, is norm ally distributed for each class, SNR is a
meaningful figure of merit for class separability [2, p. 819], an d
it can be written in the form [2, p. 967]
(3)
From (2), observe that the ROC curve is parameterized by
only SNR. Hence, TPF at fixed FPF, AUC, and pAUC are func-
tions of SNR. Moreover, these figures of merit are a ll strictly
increasing functions of SNR, i.e., they are related to each other
through one-to-one mappin gs. This fact directly results from
(2), which shows that TPF at fixed FP F is a str ictly increasing
function of SNR. Later, we will utilize this property to con stru ct
confidence intervals for TPF, AUC, and pAUC from confidence
intervals for SNR.
The functional dependence of AUC and p AUC on SNR that
was mentioned above is relatively simple. Namely, under our
assumptions, AUC takes the form [2, p. 819] [35, p. 84]
(4)
and pAUC can be written as
(5)
Note that for a CHO,
, and hence, .
III. SNR POIN T ESTIMATION
Here, we introduce our unbiased SNR point estimators, which
are a useful alternative to the unbiased
estimators given
in [29], since SNR, rather than
is often of interest. Gen -
erally, we use the same notation as in [29], with only small
changes that are clear from the text. In order to write general
expressions that include the possibility of zero images from one
class, we use the notational convention that a summation is zero
if its upper limi t is zero.
Suppose that we wish to estimate SNR for a CHO with
channels. That is, given independent, identically distributed
(i.i.d.) measurements of the class-1 channel output vector, de-
noted as
,and i.i.d. measurements of the
class-2 channel output vector, denoted as
,
we seek to estimate
.
A. Estimator Definitions
As in our previous pap er [29], we consider two estimatio n
scenarios:
(1) known
and with unknown
(2) known with unknown , ,and .
Both scenarios have their practical m erits. As discussed in the
introduction, there are cases where finding
is much easier
than finding
and . On the other hand, finding and
may sometimes be easier than finding directly, particularly
when the imaging p rocess includes strong nonlinearities.
To build our SNR estimator for scenario 1, we start by
defining the pooled sam ple covariance m atrix estimator
(6)
Next, the SNR point estimator for scenario 1 is defined to be
(7)
where
(8)
and
is the Euler Beta function. T he multiplicativ e con-
stant
is an original contribution of this work; as we will see
later, it makes the SNR estimator unbiased.