Medical lmaging has been moving from 2D modalities such as mammography to 3D modalities like Computed Tomography and Breast Tomosynthesis. These modalities have seen an increase in radiation dose to the patient and need to be optimized in an objective manner to maximize the image quality and minimize the negative effects. The way that we define image quality is related to the task that is to be performed with the image, and this entails a decision made by a radiologist when observing an image. The number of variables that can be tuned in these complex machines gives a very large number of possibilities that would be too costly and time consuming to optimize by means of human observer trials. For this reason, there are human observer surrogates that are able to be used in the optimization of medical imaging devices. These surrogates are known as model observers and they are algorithms that are able to make a detection or make a classification when given an image. They have been used in the optimization of 2D modalities with much success in the optimization of parameters such as image reconstruction, image compression, dose, etc. Because of the shift towards volumetric images, we are interested in extending the use of these algorithms to 3D images.
The detection process when viewing volumetric images involves some factors not present in the observation of 2D images simply because the number of images for one single acquisition is greatly increased. These factors include the scrolling speed and scrolling path as one moves through the slices, the effects of the peripheral vision when going from one slice to the next, the temporal persistence of the display screen, and the human observer vision. ln order to properly take into account these factors in model observers we need to understand how radiologists behave when searching for a signal in complex volumetric backgrounds. The purpose of this work was to extend existing model observer algorithms to more complex signais, and to measure the way in which human observers make complex volumetric detection tasks.
There were four main results. The first result was the extension of a model observer to detect 2D signais which are not radially symmetric. ln clinical practice, the signais that are searched for are not generally spheres or circles, and we were able to mathematically derive a mode! observer which generalized previously-proposed algorithms that made it able to detect a signal of any shape and still account for the background statistic. The second result was the measurement of the detectability of a signal as a function of vision eccentricity. This is a measurement that has been made before for simple signais, but never in clinical images. The third important measurement done was the quantification of the speed at which radiologists scroll through image slices and the manner in which they decided to explore volumetric images. We found that radiologists take a markedly different strategy than naïve observers and we were able to determine the distributions of scrolling speeds. The final result was the determination of a temporal response to the detection of a simple Gaussian signal in Gaussian noise. This measurement needed a large amount of time and a complex analysis. lt was, however, necessary to measure the temporal effects of human perception and to model them.
The results of this work culminated in a proposal for a simple 3D model observer that aims to take into account the different aspects of human perception that were measured. There are some aspects that still need to be integrated to fully mimic human observer detection. The studies performed lay the groundwork for the final aim, which is to be able to validate these model observers and use them to optimize and characterize 3D imaging systems in an objective and reproducible way.