Foveated Model Observers to predict human performance in 3D images

Greater benefits of deep learning-based computer-aided detection systems for finding small signals in 3D volumetric medical images

Purpose

Radiologists are tasked with visually scrutinizing large amounts of data produced by 3D volumetric imaging modalities. Small signals can go unnoticed during the 3D search because they are hard to detect in the visual periphery. Recent advances in machine learning and computer vision have led to effective computer-aided detection (CADe) support systems with the potential to mitigate perceptual errors.

Approach

Sixteen nonexpert observers searched through digital breast tomosynthesis (DBT) phantoms and single cross-sectional slices of the DBT phantoms. The 3D/2D searches occurred with and without a convolutional neural network (CNN)-based CADe support system. The model provided observers with bounding boxes superimposed on the image stimuli while they looked for a small microcalcification signal and a large mass signal. Eye gaze positions were recorded and correlated with changes in the area under the ROC curve (AUC).

Results

The CNN-CADe improved the 3D search for the small microcalcification signal ( Δ   AUC = 0.098 , p = 0.0002 ) and the 2D search for the large mass signal ( Δ   AUC = 0.076 , p = 0.002 ). The CNN-CADe benefit in 3D for the small signal was markedly greater than in 2D ( Δ Δ   AUC = 0.066 , p = 0.035 ). Analysis of individual differences suggests that those who explored the least with eye movements benefited the most from the CNN-CADe ( r = - 0.528 , p = 0.036 ). However, for the large signal, the 2D benefit was not significantly greater than the 3D benefit ( Δ Δ   AUC = 0.033 , p = 0.133 ).

Conclusion

The CNN-CADe brings unique performance benefits to the 3D (versus 2D) search of small signals by reducing errors caused by the underexploration of the volumetric data.

A 2D Synthesized Image Improves the 3D Search for Foveated Visual Systems

Current medical imaging increasingly relies on 3D volumetric data making it difficult for radiologists to thoroughly search all regions of the volume. In some applications (e.g., Digital Breast Tomosynthesis), the volumetric data is typically paired with a synthesized 2D image (2D-S) generated from the corresponding 3D volume. We investigate how this image pairing affects the search for spatially large and small signals. Observers searched for these signals in 3D volumes, 2D-S images, and while viewing both. We hypothesize that lower spatial acuity in the observers' visual periphery hinders the search for the small signals in the 3D images. However, the inclusion of the 2D-S guides eye movements to suspicious locations, improving the observer's ability to find the signals in 3D. Behavioral results show that the 2D-S, used as an adjunct to the volumetric data, improves the localization and detection of the small (but not large) signal compared to 3D alone. There is a concomitant reduction in search errors as well. To understand this process at a computational level, we implement a Foveated Search Model (FSM) that executes human eye movements and then processes points in the image with varying spatial detail based on their eccentricity from fixations. The FSM predicts human performance for both signals and captures the reduction in search errors when the 2D-S supplements the 3D search. Our experimental and modeling results delineate the utility of 2D-S in 3D search-reduce the detrimental impact of low-resolution peripheral processing by guiding attention to regions of interest, effectively reducing errors.

Deep-learning model observer for a low-contrast hepatic metastases localization task in computed tomography

PURPOSE

Conventional model observers (MO) in CT are often limited to a uniform background or varying background that is random and can be modeled in an analytical form. It is unclear if these conventional MOs can be readily generalized to predict human observer performance in clinical CT tasks that involve realistic anatomical background. Deep-learning-based model observers (DL-MO) have recently been developed, but have not been validated for challenging low contrast diagnostic tasks in abdominal CT. We consequently sought to validate a DL-MO for a low-contrast hepatic metastases localization task.

METHODS

We adapted our recently developed DL-MO framework for the liver metastases localization task. Our previously-validated projection-domain lesion-/noise-insertion techniques were used to synthesize realistic positive and low-dose abdominal CT exams, using the archived patient projection data. Ten experimental conditions were generated, which involved different lesion sizes/contrasts, radiation dose levels, and image reconstruction types. Each condition included 100 trials generated from a patient cohort of 7 cases. Each trial was presented as liver image patches (160×160×5 voxels). The DL-MO performance was calculated for each condition and was compared with human observer performance, which was obtained by three sub-specialized radiologists in an observer study. The performance of DL-MO and radiologists was gauged by the area under localization receiver-operating-characteristic curves. The generalization performance of the DL-MO was estimated with the repeated twofold cross-validation method over the same set of trials used in the human observer study. A multi-slice Channelized Hoteling Observers (CHO) was compared with the DL-MO across the same experimental conditions.

RESULTS

The performance of DL-MO was highly correlated to that of radiologists (Pearson's correlation coefficient: 0.987; 95% CI: [0.942, 0.997]). The performance level of DL-MO was comparable to that of the grouped radiologists, that is, the mean performance difference was -3.3%. The CHO performance was poorer than the grouped radiologist performance, before internal noise could be added. The correlation between CHO and radiologists was weaker (Pearson's correlation coefficient: 0.812, and 95% CI: [0.378, 0.955]), and the corresponding performance bias (-29.5%) was statistically significant.

CONCLUSION

The presented study demonstrated the potential of using the DL-MO for image quality assessment in patient abdominal CT tasks.

Medical image quality metrics for foveated model observers

Purpose: A recently proposed model observer mimics the foveated nature of the human visual system by processing the entire image with varying spatial detail, executing eye movements, and scrolling through slices. The model can predict how human search performance changes with signal type and modality (2D versus 3D), yet its implementation is computationally expensive and time-consuming. Here, we evaluate various image quality metrics using extensions of the classic index of detectability expression and assess foveated model observers for search tasks. Approach: We evaluated foveated extensions of a channelized Hotelling and nonprewhitening matched filter model with an eye filter. The proposed methods involve calculating a model index of detectability ( d ' ) for each retinal eccentricity and combining these with a weighting function into a single detectability metric. We assessed different versions of the weighting function that varied in the required measurements of the human observers' search (no measurements, eye movement patterns, size of the image, and median search times). Results: We show that the index of detectability across eccentricities weighted using the eye movement patterns of observers best predicted human performance in 2D versus 3D search performance for a small microcalcification-like signal and a larger mass-like. The metric with a weighting function based on median search times was the second best predicting human results. Conclusions: The findings provide a set of model observer tools to evaluate image quality in the early stages of imaging system evaluation or design without implementing the more computationally complex foveated search model.

Foveated Model Observers for Visual Search in 3D Medical Images

Model observers have a long history of success in predicting human observer performance in clinically-relevant detection tasks. New 3D image modalities provide more signal information but vastly increase the search space to be scrutinized. Here, we compared standard linear model observers (ideal observers, non-pre-whitening matched filter with eye filter, and various versions of Channelized Hotelling models) to human performance searching in 3D 1/f2.8 filtered noise images and assessed its relationship to the more traditional location known exactly detection tasks and 2D search. We investigated two different signal types that vary in their detectability away from the point of fixation (visual periphery). We show that the influence of 3D search on human performance interacts with the signal's detectability in the visual periphery. Detection performance for signals difficult to detect in the visual periphery deteriorates greatly in 3D search but not in 3D location known exactly and 2D search. Standard model observers do not predict the interaction between 3D search and signal type. A proposed extension of the Channelized Hotelling model (foveated search model) that processes the image with reduced spatial detail away from the point of fixation, explores the image through eye movements, and scrolls across slices can successfully predict the interaction observed in humans and also the types of errors in 3D search. Together, the findings highlight the need for foveated model observers for image quality evaluation with 3D search.

Search of low-contrast liver lesions in abdominal CT: the importance of scrolling behavior

Purpose: Visual search using volumetric images is becoming the standard in medical imaging. However, we do not fully understand how eye movement strategies mediate diagnostic performance. A recent study on computed tomography (CT) images showed that the search strategies of radiologists could be classified based on saccade amplitudes and cross-quadrant eye movements [eye movement index (EMI)] into two categories: drillers and scanners. Approach: We investigate how the number of times a radiologist scrolls in a given direction during analysis of the images (number of courses) could add a supplementary variable to use to characterize search strategies. We used a set of 15 normal liver CT images in which we inserted 1 to 5 hypodense metastases of two different signal contrast amplitudes. Twenty radiologists were asked to search for the metastases while their eye-gaze was recorded by an eye-tracker device (EyeLink1000, SR Research Ltd., Mississauga, Ontario, Canada). Results: We found that categorizing radiologists based on the number of courses (rather than EMI) could better predict differences in decision times, percentage of image covered, and search error rates. Radiologists with a larger number of courses covered more volume in more time, found more metastases, and made fewer search errors than those with a lower number of courses. Our results suggest that the traditional definition of drillers and scanners could be expanded to include scrolling behavior. Drillers could be defined as scrolling back and forth through the image stack, each time exploring a different area on each image (low EMI and high number of courses). Scanners could be defined as scrolling progressively through the stack of images and focusing on different areas within each image slice (high EMI and low number of courses). Conclusions: Together, our results further enhance the understanding of how radiologists investigate three-dimensional volumes and may improve how to teach effective reading strategies to radiology residents.

Deep-learning-based model observer for a lung nodule detection task in computed tomography

Purpose: Task-based image quality assessment using model observers (MOs) is an effective approach to radiation dose and scanning protocol optimization in computed tomography (CT) imaging, once the correlation between MOs and radiologists can be established in well-defined clinically relevant tasks. Conventional MO studies were typically simplified to detection, classification, or localization tasks using tissue-mimicking phantoms, as traditional MOs cannot be readily used in complex anatomical background. However, anatomical variability can affect human diagnostic performance. Approach: To address this challenge, we developed a deep-learning-based MO (DL-MO) for localization tasks and validated in a lung nodule detection task, using previously validated projection-based lesion-/noise-insertion techniques. The DL-MO performance was compared with 4 radiologist readers over 12 experimental conditions, involving varying radiation dose levels, nodule sizes, nodule types, and reconstruction types. Each condition consisted of 100 trials (i.e., 30 images per trial) generated from a patient cohort of 50 cases. DL-MO was trained using small image volume-of-interests extracted across the entire volume of training cases. For each testing trial, the nodule searching of DL-MO was confined to a 3-mm thick volume to improve computational efficiency, and radiologist readers were tasked to review the entire volume. Results: A strong correlation between DL-MO and human readers was observed (Pearson's correlation coefficient: 0.980 with a 95% confidence interval of [0.924, 0.994]). The averaged performance bias between DL-MO and human readers was 0.57%. Conclusion: The experimental results indicated the potential of using the proposed DL-MO for diagnostic image quality assessment in realistic chest CT tasks.

Evaluation of Search Strategies for Microcalcifications and Masses in 3D Images

Medical imaging is quickly evolving towards 3D image modalities such as computed tomography (CT), magnetic resonance imaging (MRI) and digital breast tomosynthesis (DBT). These 3D image modalities add volumetric information but further increase the need for radiologists to search through the image data set. Although much is known about search strategies in 2D images less is known about the functional consequences of different 3D search strategies. We instructed readers to use two different search strategies: drillers had their eye movements restricted to a few regions while they quickly scrolled through the image stack, scanners explored through eye movements the 2D slices. We used real-time eye position monitoring to ensure observers followed the drilling or the scanning strategy while approximately preserving the percentage of the volumetric data covered by the useful field of view. We investigated search for two signals: a simulated microcalcification and a larger simulated mass. Results show an interaction between the search strategy and lesion type. In particular, scanning provided significantly better detectability for microcalcifications at the cost of 5 times more time to search while there was little change in the detectability for the larger simulated masses. Analyses of eye movements support the hypothesis that the effectiveness of a search strategy in 3D imaging arises from the interaction of the fixational sampling of visual information and the signals' visibility in the visual periphery.

Interactions of lesion detectability and size across single-slice DBT and 3D DBT

Three dimensional image modalities introduce a new paradigm for visual search requiring visual exploration of a larger search space than 2D imaging modalities. The large number of slices in the 3D volumes and the limited reading times make it difficult for radiologists to explore thoroughly by fixating with their high resolution fovea on all regions of each slice. Thus, for 3D images, observers must rely much more on their visual periphery (points away from fixation) to process image information. We previously found a dissociation in signal detectability between 2D and 3D search tasks for small signals in synthetic textures evaluated with non-radiologist trained observers. Here, we extend our evaluation to more clinically realistic backgrounds and radiologist observers. We studied the detectability of simulated microcalcifications (MCALC) and masses (MASS) in Digital Breast Tomosynthesis (DBT) utilizing virtual breast phantoms. We compared the lesion detectability of 8 radiologists during free search in 3D DBT and a 2D single-slice DBT (center slice of the 3D DBT). Our results show that the detectability of the microcalcification degrades significantly in 3D DBT with respect to the 2D single-slice DBT. On the other hand, the detectability for masses does not show this behavior and its detectability is not significantly different. The large deterioration of the 3D detectability of microcalcifications relative to masses may be related to the peripheral processing given the high number of cases in which the microcalcification was missed and the high number of search errors. Together, the results extend previous findings with synthetic textures and highlight how search in 3D images is distinct from 2D search as a consequence of the interaction between search strategies and the visibility of signals in the visual periphery.

The role of extra-foveal processing in 3D imaging

The field of medical image quality has relied on the assumption that metrics of image quality for simple visual detection tasks are a reliable proxy for the more clinically realistic visual search tasks. Rank order of signal detectability across conditions often generalizes from detection to search tasks. Here, we argue that search in 3D images represents a paradigm shift in medical imaging: radiologists typically cannot exhaustively scrutinize all regions of interest with the high acuity fovea requiring detection of signals with extra-foveal areas (visual periphery) of the human retina. We hypothesize that extra-foveal processing can alter the detectability of certain types of signals in medical images with important implications for search in 3D medical images. We compare visual search of two different types of signals in 2D vs. 3D images. We show that a small microcalcification-like signal is more highly detectable than a larger mass-like signal in 2D search, but its detectability largely decreases (relative to the larger signal) in the 3D search task. Utilizing measurements of observer detectability as a function retinal eccentricity and observer eye fixations we can predict the pattern of results in the 2D and 3D search studies. Our findings: 1) suggest that observer performance findings with 2D search might not always generalize to 3D search; 2) motivate the development of a new family of model observers that take into account the inhomogeneous visual processing across the retina (foveated model observers).