Lecture 7- Advanced Image Quality Flashcards
quantitative physical measurement
spatial resolution, noise, contrast, CNR, SNR, NEQ,DQE
phantom-based observer assessment human or algorithm
performance on some phantom base taste e.g. object dtectibility, low contrast resolution
patient based observer assessment
radiologist acceptance
quantitative measurement of diagnostic/detective value for a specific diagnostic task
ROC study to determine sensitivity and specificity
signal to noise ratio
measure of a signal strength relative to the background noise level In the X-ray imaging, signal provided by number of x-ray photons, subject attenuation, and imaging system. Noise also contributed by number of x-ray photons, subject attenuation, and by imaging system
what is the formula for SNR
SNR= Nbar/signma=suqareroot of Nbar
what does SNR depend on
number of photons
what is the noise equivalent quanta
the number of Poisson-distributed quanta that would produce the same SNR with an ideal detector at a given spatial frequency. NEQ is essentially a spatial frequency domain descriptor of SNR^2, assuming LSI system
what is the absolute measure of image quality:
of x-ray quanta that an image is worth at each spatial frequency
why is the task-based approaches for image quality evaluation
ultimate goal of imaging is to provide useful images for given diagnostic tasks. System optimization has to improve the diagnostic outcome.
individual physical properties of image quality
contrast, resolution, noise. do not characterize the overall image quality
what does the NEQ represent
absolute and overall image quality, but may not be able to predict or correlate with the diagnostic performance for a given task
what are the four components of a task based approach
task, objects and images, observer (decesion maker), figure of merit
what goes into classification and detection
the image is to be classified into one of the many possible alternatives, finite number of hypotheses like detection or classification of a tumor
what does into estimation or quantification
estimating one or more numerical parameters for the image, infinite number of hypotheses, typically involve a numerical algorithm, rather than a computation by a human. like tumor volume
what goes into hybrid estimation-classification
estimate one or more parameters that will then be used as input to subsequent decision making operations
what goes into the observer role, what is being decided
if an image is there or not. so presence or absence. can be human or algorithm
what are estimation tasks
accuracy, variance, MSE or ML estimator, cramer rao lower bound
what are classification/detection tasks
accuracy, sensitivity and specificity pairs, positive and negative predictive value, area under the curve ROC, SNR detectibility index and cost + utility
what is accuracy as a figure of merit
accuracy is the simplest FOM to quntiy the binary classification system from. Fraction of cases for which the decision is correct compared with the truth
what are the limits using accuracy for FOM
because accuracy is highly dependent on the prevalence of the underlying hypothesis.
what is the sensitivity and specificity and what does it mean and formula
sensitivity is the probability of the correct decision when presented with positive cases. Sp the probability of the correct devision when presented with negative cases.
Sn= TP/(TP+FN) Sp= TN/(TN+FP)
what are the PPV and NPV and the formulas
PPV= TP/(TP+FP)- the probability that the case is actually positive, when the observer says the case is positive.
NPV= TN/(TN+FN)- the probability that the case is actually negative when the observer says it is negative
draw a curve with lines representing true positive and true negative and threshold values
draw it
what is an ROC and what does it look like
draw and it is a receiver operative characteristic curve
what is the area under the ROC curve mean
it is a meaningful measure of overall performance for classification/detection tasks.
higher area under the curve means what
higher performance
what is the SNR
it is the signal to noise ratio and it is degree of overall for the two distributions of the two decision variable to determine the operability of the two classes determines the detectability. AUC is one measure of the overlap and another is SNR
what is the index of detectability
it is the SNR derived from the AUC. Normal assumption is appproriate for many experimental ROC from human data
what is the yes no experiment
it is where someone picks yes or no or positive vs. negative
what is the rating experiment scale
for each trial, the observer is asked to indicate the confidence level that the trial belongs to the two hypotheses. At leas five rating scales. this rating data can be used to to make an ROC by sweeping the decision threshold between rating scales
go through homework of how to turn data into a ROC
draw it
what are the issues with human observers in ROC studies
human oberver and ROC analysis can be considered as gold standard method to measure image quality for a specific visual task. ROC studies with human observers are time consuming and difficult to control. Tests of statical observation are hard to tell.
mathematical model observers mean what
algorithms to quantify objectively or to predict human observer performacnce for a given visual task. Ideal is bayesian observer
what is the ideal observer and likelihood ratio
any observe that uses the likelihood ratio as the decision variable is referred to as an ideal observer. points on ROC curve for ideal observer are generated by comparing A(g) to different thresholds
propertites of ideal observer
it maximizes the area under the Roc curve, maximizes true positive fraction at any false positive fraction. An observer than uses a monotonically trasnformed version of the linelihood ratio is still an ideal observer. The same post processing is used
what are the difficulties with an ideal observer
requires immense amounts of prior knowledge. Usually a nonlinearr function of the imaging data and it is difficult to compute except for textbook cases. Not obviously related to humans or other practical observers
