In CLUES we will use and further develop two model observers that are currently of interest in the scientific community to predict human detection performance. The models have shown to be useful and successful in predicting human performance in medical images. Two model observers will be investigated in this project: the non-prewhitening matched filter with eye filter will be used (NPWE) and the channelized-Hotelling observer (CHO).
- The NPWE model observer uses statistical information about the signal and background but also takes into account the human visual sensitivity to different spatial frequencies. From the statistical information, the model computes a discrimination index d′ that is related to detection probability and that can be converted into a proportion correct (PC) decisions or area under the ROC curve (Az). The PC and Az are parameters that can be used to compare model observer performance directly with radiologist performance in a detection task. The figure gives more detailed information on this model observer type.
- The channelized model observer (CHO model observer) uses a set of channels to reflect the response of neurons in the primary visual cortex. Similar to the NPWE model observer, a discrimination index can be derived from the test statistics. The CHO observer involves a training procedure to first learn the characteristics of the image noise and eventually the signal in the image data. This is called decorrelation or pre-whitening of noise which is achieved by determination of co-variance matrices of the image data. The NPWE model observer on the other hand does not pre-whiten the noise and only uses knowledge of the signal based on pre-defined signal templates (no training is involved).
 |
The figure represents a simplified flowchart of a classic 2D NPWE model observer. A dataset of small images – i.e. extracted regions from actual images - containing a low contrast object is compared to a dataset containing only background regions (no object present). To obtain a measure of detectability, the model observer uses a predefined template that expresses the theoretical object image with respect to an ideal imaging system. The template is filtered to account for the system blur of the actual imaging system and to account for the human visual response (eye filter). The signal template is both correlated with the image dataset containing the object and the image dataset containing only background (both also filtered by the eye filter). From the correlation results one can compute a discrimination index d′. Intuitively it can be understood that strong correlation of template and object images suggests clear visibility of an object (higher d’).
|
