Convolutional Neural Network Model Observers Discount Signal-like Anatomical Structures During Search in Virtual Digital Breast Tomosynthesis Phantoms

Model observers are computational tools to evaluate and optimize task-based medical image quality. Linear model observers, such as the Channelized Hotelling Observer (CHO), predict human accuracy in detection tasks with a few possible signal locations in clinical phantoms or real anatomic backgrounds. In recent years, Convolutional Neural Networks (CNNs) have been proposed as a new type of model observer. What is not well understood is what CNNs add over the more common linear model observer approaches. We compare the CHO and CNN detection accuracy to the radiologist's accuracy in searching for two types of signals (mass and microcalcification) embedded in 2D/3D breast tomosynthesis phantoms (DBT). We show that the CHO model's accuracy is comparable to the CNN's performance for a location-known-exactly detection task. However, for the search task with 2D/3D DBT phantoms, the CHO's detection accuracy was significantly lower than the CNN accuracy. A comparison to the radiologist's accuracy showed that the CNN but not the CHO could match or exceed the radiologist's accuracy in the 2D microcalcification and 3D mass search conditions. An analysis of the eye position showed that radiologists fixated more often and longer at the locations corresponding to CNN false positives. Most CHO false positives were the phantom's normal anatomy and were not fixated by radiologists. In conclusion, we show that CNNs can be used as an anthropomorphic model observer for the search task for which traditional linear model observers fail due to their inability to discount false positives arising from the anatomical backgrounds.

On anthropomorphic decision making in a model observer

By analyzing human readers' performance in detecting small round lesions in simulated digital breast tomosynthesis background in a location known exactly scenario, we have developed a model observer that is a better predictor of human performance with different levels of background complexity (i.e., anatomical and quantum noise). Our analysis indicates that human observers perform a lesion detection task by combining a number of sub-decisions, each an indicator of the presence of a lesion in the image stack. This is in contrast to a channelized Hotelling observer, where the detection task is conducted holistically by thresholding a single decision variable, made from an optimally weighted linear combination of channels. However, it seems that the sub-par performance of human readers compared to the CHO cannot be fully explained by their reliance on sub-decisions, or perhaps we do not consider a sufficient number of sub-decisions. To bridge the gap between the performances of human readers and the model observer based upon sub-decisions, we use an additive noise model, the power of which is modulated with the level of background complexity. The proposed model observer better predicts the fast drop in human detection performance with background complexity.

It is hard to see a needle in a haystack: Modeling contrast masking effect in a numerical observer

Within the framework of a virtual clinical trial for breast imaging, we aim to develop numerical observers that follow the same detection performance trends as those of a typical human observer. In our prior work, we showed that by including spatiotemporal contrast sensitivity function (stCSF) of human visual system (HVS) in a multi-slice channelized Hotelling observer (msCHO), we can correctly predict trends of a typical human observer performance with the viewing parameters of browsing speed, viewing distance and contrast. In this work we further improve our numerical observer by modeling contrast masking. After stCSF, contrast masking is the second most prominent property of HVS and it refers to the fact that the presence of one signal affects the visibility threshold for another signal. Our results indicate that the improved numerical observer better predicts changes in detection performance with background complexity.

Development and evaluation of a 3D model observer with nonlinear spatiotemporal contrast sensitivity

We investigate improvements to our 3D model observer with the goal of better matching human observer performance as a function of viewing distance, effective contrast, maximum luminance, and browsing speed. Two nonlinear methods of applying the human contrast sensitivity function (CSF) to a 3D model observer are proposed, namely the Probability Map (PM) and Monte Carlo (MC) methods. In the PM method, the visibility probability for each frequency component of the image stack, p, is calculated taking into account Barten's spatiotemporal CSF, the component modulation, and the human psychometric function. The probability p is considered to be equal to the perceived amplitude of the frequency component and thus can be used by a traditional model observer (e.g., LG-msCHO) in the space-time domain. In the MC method, each component is randomly kept with probability p or discarded with 1-p. The amplitude of the retained components is normalized to unity. The methods were tested using DBT stacks of an anthropomorphic breast phantom processed in a comprehensive simulation pipeline. Our experiments indicate that both the PM and MC methods yield results that match human observer performance better than the linear filtering method as a function of viewing distance, effective contrast, maximum luminance, and browsing speed.

Integration of spatio-temporal contrast sensitivity with a multi-slice channelized Hotelling observer

Barten's model of spatio-temporal contrast sensitivity function of human visual system is embedded in a multi-slice channelized Hotelling observer. This is done by 3D filtering of the stack of images with the spatio-temporal contrast sensitivity function and feeding the result (i.e., the perceived image stack) to the multi-slice channelized Hotelling observer. The proposed procedure of considering spatio-temporal contrast sensitivity function is generic in the sense that it can be used with observers other than multi-slice channelized Hotelling observer. Detection performance of the new observer in digital breast tomosynthesis is measured in a variety of browsing speeds, at two spatial sampling rates, using computer simulations. Our results show a peak in detection performance in mid browsing speeds. We compare our results to those of a human observer study reported earlier (I. Diaz et al. SPIE MI 2011). The effects of display luminance, contrast and spatial sampling rate, with and without considering foveal vision, are also studied. Reported simulations are conducted with real digital breast tomosynthesis image stacks, as well as stacks from an anthropomorphic software breast phantom (P. Bakic et al. Med Phys. 2011). Lesion cases are simulated by inserting single micro-calcifications or masses. Limitations of our methods and ways to improve them are discussed.