Model Observers in Medical Imaging Research

He, Xin; Park, Subok

doi:10.7150/thno.5138

Cited by 101 publications

(105 citation statements)

References 80 publications

(111 reference statements)

Supporting

Mentioning

102

Contrasting

Order By: Relevance

“…From this evaluation, the first eight LG channels with width σ set to 187 µm (2.2 pixels of 85 µm were selected because that setting resulted in the highest free-search ROC (FROC) area under the curve (AUC) determined by the weighted JAFROC method 40 . The efficient LG channels were selected over the anthropomorphic difference-of-Gaussian or Gabor channels 6,8,41 , because the search step included in the model observer (see section II C 3) to account for the uncertain target locations in the phantom, will decrease the initial performance. Therefore it seemed more prudent to start from the higher scoring LG channels, where this drop in performance could be tuned to match human performance, rather than start from the anthropomorphic channels, where this further drop in performance could end up lower than the actual human performance.…”

Section: Templates and Channelsmentioning

confidence: 99%

“…Further optimization within the constraints of existing hardware could therefore potentially increase the clinical performance of these devices, but this process requires relevant quality metrics and efficient methods to evaluate them. A good candidate is the detection performance of simple geometric shapes in a structured background, because for this task human performance can be estimated by model observers [4][5][6][7][8] .…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Design of a model observer to evaluate calcification detectability in breast tomosynthesis and application to smoothing prior optimization

et al. 2016

View full text Add to dashboard Cite

Purpose: In this work we design and validate a model observer that can detect groups of microcalcifications in a four alternative forced choice (4-AFC) experiment and use it to optimize a smoothing prior for detectability of microcalcifications.Methods: A channelized Hotelling observer (CHO) with eight Laguerre-Gauss channels was designed to detect groups of five microcalcifications in a background of acrylic spheres by adding the CHO log-likelihood ratios calculated at the expected locations of the five calcifications.This model observer is then applied to optimize the detectability of the microcalcifications as a function of the smoothing prior. We examine the quadratic and total variation (TV) priors, and a combination of both. A selection of these reconstructions was then evaluated by human observers to validate the correct working of the model observer.Results: We found a clear maximum for the detectability of microcalcification when using the total variation prior with weight βT V = 35. Detectability only varied over a small range for the quadratic and combined quadratic-TV priors when weight βQ of the quadratic prior was changed by two orders of magnitude.Spearman correlation with human observers was good except for the highest value of β for the quadratic and TV priors. Excluding those, we found ρ = 0.93 when comparing detection fractions, and ρ = 0.86 for the fitted detection threshold diameter.Conclusions: We successfully designed a model observer that was able to predict human performance over a large range of settings of the smoothing prior, except for the highest values of β which were outside the useful range for good image quality.Since detectability only depends weakly on the strength of the combined prior, it is not possible to pick an optimal smoothness based only on this criterion. On the other hand, such choice can now be made based on other criteria without worrying about calcification detectability.

show abstract

Section: Templates and Channelsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Design of a model observer to evaluate calcification detectability in breast tomosynthesis and application to smoothing prior optimization

et al. 2016

View full text Add to dashboard Cite

show abstract

“…This facilitates exploration of large numbers of experimental conditions that can outstrip what is practicable with limited expert resources, in terms of time and observer numbers usually available for such studies. Human and model observers have been used as sophisticated tools for determining image quality, in this context referring to the ability of an image to deliver useable content for a particular task [2]. One such approach is the use of n-alternative forced choice studies where the observer must decide which of n images contains a particular target.…”

Section: Introductionmentioning

confidence: 99%

“…The principal advantage of a model observer (MO) is the ability to reliably, and tirelessly, undertake evaluations of sets of images without recourse to the variables introduced by using groups of human observers. This approach is routed in statistical decision theory and linear discriminant analysis and has been an active area of research for many years [2,3,5].…”

Section: Introductionmentioning

confidence: 99%

A deep learning model observer for use in alterative forced choice virtual clinical trials

Alnowami

Mills

Awis

et al. 2018

Medical Imaging 2018: Image Perception, Observer Performance, and Technology Assessment

View full text Add to dashboard Cite

Virtual clinical trials (VCTs) represent an alternative assessment paradigm that overcomes issues of dose, high cost and delay encountered in conventional clinical trials for breast cancer screening. However, to fully utilize the potential benefits of VCTs requires a machine-based observer that can rapidly and realistically process large numbers of experimental conditions. To address this, a Deep Learning Model Observer (DLMO) was developed and trained to identify lesion targets from normal tissue in small (200 x 200 pixel) image segments, as used in Alternative Forced Choice (AFC) studies. The proposed network consists of 5 convolutional layers with 2x2 kernels and ReLU (Rectified Linear Unit) activations, followed by max pooling with size equal to the size of the final feature maps and three dense layers. The class outputs weights from the final fully connected dense layer are used to consider sets of n images in an n-AFC paradigm to determine the image most likely to contain a target. To examine the DLMO performance on clinical data, a training set of 2814 normal and 2814 biopsy-confirmed malignant mass targets were used. This produced a sensitivity of 0.90 and a specificity of 0.92 when presented with a test data set of 800 previously unseen clinical images. To examine the DLMOs minimum detectable contrast, a second dataset of 630 simulated backgrounds and 630 images with simulated lesion and spherical targets (4mm and 6mm diameter), produced contrast thresholds equivalent to/better than human observer performance for spherical targets, and comparable (12 % difference) for lesion targets.

show abstract

“…An alternative is to use computational observers. [28][29][30][31] A widely used model in medical imaging is the channelized Hotelling observer (CHO). The CHO models the human observer using spatial frequency-selective filters, or channels, and optimal linear statistics to build a template for signal detection.…”

Section: Introductionmentioning

confidence: 99%

Computational and human observer image quality evaluation of low dose, knowledge‐based CT iterative reconstruction

et al. 2015

View full text Add to dashboard Cite

Purpose: Aims in this study are to (1) develop a computational model observer which reliably tracks the detectability of human observers in low dose computed tomography (CT) images reconstructed with knowledge-based iterative reconstruction (IMR™, Philips Healthcare) and filtered back projection (FBP) across a range of independent variables, (2) use the model to evaluate detectability trends across reconstructions and make predictions of human observer detectability, and (3) perform human observer studies based on model predictions to demonstrate applications of the model in CT imaging. Methods: Detectability (d ′ ) was evaluated in phantom studies across a range of conditions. Images were generated using a numerical CT simulator. Trained observers performed 4-alternative forced choice (4-AFC) experiments across dose (1.3, 2.7, 4.0 mGy), pin size (4, 6, 8 mm), contrast (0.3%, 0.5%, 1.0%), and reconstruction (FBP, IMR), at fixed display window. A five-channel Laguerre-Gauss channelized Hotelling observer (CHO) was developed with internal noise added to the decision variable and/or to channel outputs, creating six different internal noise models. Semianalytic internal noise computation was tested against Monte Carlo and used to accelerate internal noise parameter optimization. Model parameters were estimated from all experiments at once using maximum likelihood on the probability correct, P C . Akaike information criterion (AIC) was used to compare models of different orders. The best model was selected according to AIC and used to predict detectability in blended FBP-IMR images, analyze trends in IMR detectability improvements, and predict dose savings with IMR. Predicted dose savings were compared against 4-AFC study results using physical CT phantom images. Results: Detection in IMR was greater than FBP in all tested conditions. The CHO with internal noise proportional to channel output standard deviations, Model-k4, showed the best trade-off between fit and model complexity according to AIC c . With parameters fixed, the model reasonably predicted detectability of human observers in blended FBP-IMR images. Semianalytic internal noise computation gave results equivalent to Monte Carlo, greatly speeding parameter estimation. Using Model-k4, the authors found an average detectability improvement of 2.7 ± 0.4 times that of FBP. IMR showed greater improvements in detectability with larger signals and relatively consistent improvements across signal contrast and x-ray dose. In the phantom tested, Model-k4 predicted an 82% dose reduction compared to FBP, verified with physical CT scans at 80% reduced dose. Conclusions: IMR improves detectability over FBP and may enable significant dose reductions. A channelized Hotelling observer with internal noise proportional to channel output standard deviation agreed well with human observers across a wide range of variables, even across reconstructions with drastically different image characteristics. Utility of the model observer was demonstrated by predicting the effect of im...

show abstract

Model Observers in Medical Imaging Research

Cited by 101 publications

References 80 publications

Design of a model observer to evaluate calcification detectability in breast tomosynthesis and application to smoothing prior optimization

Design of a model observer to evaluate calcification detectability in breast tomosynthesis and application to smoothing prior optimization

A deep learning model observer for use in alterative forced choice virtual clinical trials

Computational and human observer image quality evaluation of low dose, knowledge‐based CT iterative reconstruction

Contact Info

Product

Resources

About