Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2023 May 1.
Published in final edited form as: Proc SPIE Int Soc Opt Eng. 2023 Apr 3;12467:124670U. doi: 10.1117/12.2654323

Optimizing data acquisition in undersampled magnetic resonance imaging (MRI) using two alternative forced choice (2-AFC) and search tasks

Tavianne M Kemp 1, Tetsuya A Kawakita 1, Rehan Mehta 1, Angel R Pineda 1
PMCID: PMC10150592  NIHMSID: NIHMS1895125  PMID: 37131343

Abstract

Undersampling in the frequency domain (k-space) in MRI accelerates the data acquisition. Typically, a fraction of the low frequencies is fully collected and the rest are equally undersampled. We used a fixed 1D undersampling factor of 5x where 20% of the k-space lines are collected but varied the fraction of the low k-space frequencies that are fully sampled. We used a range of fully acquired low k-space frequencies from 0% where the primary artifact is aliasing to 20% where the primary artifact is blurring in the undersampling direction. Small lesions were placed in the coil k-space data for fluid-attenuated inversion recovery (FLAIR) brain images from the fastMRI database. The images were reconstructed using a multi-coil SENSE reconstruction with no regularization. We conducted a human observer two-alternative forced choice (2-AFC) study with a signal known exactly and a search task with variable backgrounds for each of the acquisitions. We found that for the 2-AFC task, the average human observer did better with more of the low frequencies being fully sampled. For the search task, we found that after an initial improvement from having none of the low frequencies fully sampled to just 2.5%, the performance remained fairly constant. We found that the performance in the two tasks had a different relationship to the acquired data. We also found that the search task was more consistent with common practice in MRI where a range of frequencies between 5% and 10% of the low frequencies are fully sampled.

Keywords: Task-based assessment, 2-AFC, search, MRI, undersampling

INTRODUCTION

Magnetic resonance imaging provides a wide range of contrast mechanisms with diagnostic information but suffers from slow acquisition times. One of the methods to accelerate the data acquisition is by undersampling in the frequency (kspace) domain.1,2 There multiple sampling patterns which can be undersampled including cartesian, spiral, radial and random sampling.1,3 In clinical practice, most of the acquisitions are cartesian and the undersampling done with parallel imaging is uniform with a certain percentage of the low frequencies being fully sampled. This is the situation that we consider in this work. For a specific task and undersampling rate, how much of the lower frequencies should be fully sampled? How does the fraction of fully sampled low frequencies required change with task?

Previous work has looked at reconstruction of parallel imaging data in MRI using a task-based approach and compared it with perceptual difference models.4 That worked focused on the reconstruction rather than the data acquired.

We approach this question from a task-based perspective.5,6 Using this perspective, we need to specify the ensemble of images, the observer and the task. In particular, the fraction of fully sampled low frequencies depends on the frequency content of the images, the frequency content of the signal being detected in a detection task and whether the observer is a human or a machine observer. In this work, we consider two types of detection tasks, one where the location of the signal is known and one where it needs to be localized.

Task-based optimization often assumes a rank-ordering between the performance of a simple task with the performance in more complex tasks. In this work, along with evaluating the data acquisition in MRI, we also compare the performance of a location known task6 with a search task.7 In particular, knowing the location of the signal may decrease the effect of some of the aliasing artifacts from undersampling.

METHODS

2.1. Undersampled acquisition and multi-coil reconstruction in MRI

We consider 1-D undersampling of FLAIR images for a fixed undersampling factor of 5x. While this level of acceleration is beyond what is done with a 1-D acceleration, it provides a reasonable range of values for the fraction of the low k-space frequencies that are fully sampled. The data is acquired by fully sampling a fraction of the low k-space frequencies and uniformly undersampling the higher frequencies. The images were obtained from the NYU fastMRI Initiative database.8 As such, NYU fastMRI investigators provided data but did not participate in analysis or writing of this report. The images were reconstructed using a multi-coil SENSE reconstruction2 with the coil sensitivities estimated using the sum of squares method which leads to real estimates of the signal. The reconstruction of the images was done using the Berkeley Advanced Reconstruction Toolbox (BART) toolbox.9 Sample acquisitions and corresponding images are shown in Figure 1.

Figure 1.

Figure 1.

Open in a new tab

Sample acquisitions in k-space with images. The top row represents sample k-space acquisitions at different percentages of the low frequencies being fully sampled. In every acquisition, the total number of k-space lines is constant (20% of the fully sampled data for a 5X undersampling). The bottom row has the corresponding images for the same slice. All of the slices have the signal in the location shown in the 20% image. The distribution of frequencies acquired in k-space determines the type of dominant artifacts in the images. On the zero percent image, the aliasing artifact in the vertical direction is dominant. In the 20% image which only collects the lowest 20% of the frequencies in the fully acquired image, the blurring artifact is dominant.

2.2. Two-alternative forced choice experiments

The human observers had to choose between two images where one of the two contained the signal. The signal was presented to the observer along with the two options. The images had varying backgrounds but the signal was always at the center. This is a signal-known-exactly (SKE) task with anatomical variability. An example 2-AFC trial is shown in Figure 2. Four observers carried out 200 2-AFC trials for each experimental condition. For each observer, the images used in each 2-AFC trial were randomly chosen from 200 backgrounds with the signal and 200 background images. Each observer saw a different set of images.10,11

Figure 2.

Figure 2.

Open in a new tab

Sample 2AFC trial for 5% of the low frequencies being fully acquired with the signal in the left image.

2.3. Search Experiments

The human observers had to choose the location of the signal in a specified region of the image7 identified with a square. The signal was presented to the observer along with the search region. If the identified location was within five pixels of the true location, the signal was considered successfully located. The images with the signals had varying backgrounds. An example trial in the search task is shown in Figure 3. Four observers carried out 200 search trials for each experimental condition. These images were the same background images as were used in the 2AFC trial. Since we used a total of 200 slices, all observer saw the same 200 images. For each image, the location of the signal was randomly chosen.

Figure 3.

Figure 3.

Open in a new tab

Sample search trial for 2.5% of the low frequencies being fully acquired with signal identified by the arrow.

2.4. Common Experimental Conditions

All observers repeated the initial condition as training until the performance was stabilized. The experimental conditions were blinded to the observers and were done in the following order (5%, 20%, 0%, 2.5%, 7.5%, 10%, 15%). For the 2-AFC study, the signal amplitude was chosen so that the mean percent correct for the observers would provide a range of values over the experimental conditions. Because of this, the signal is subtle. In the search task, the amplitude of the signal was twice as in the 2-AFC study. All observers used a Barco MDRC 2321 monitor (in a dark room) with a resolution of 0.294 mm and were approximately 50 cm from the screen but a head restraint was not used. The observers received feedback on whether they had identified the signal correctly after each trial and took breaks between experimental conditions to avoid fatigue.

3. RESULTS AND DISCUSSION

In previous work,12–14 we studied the 2-AFC task performance for a fixed undersampling pattern with 5% of the low frequencies being fully acquired and 3.48 undersampling in MRI. In that situation, task performance remained constant for a wide range of regularization parameters for both human and ideal linear observers. Here we vary the undersampling pattern to explore the effect of which frequencies are acquired on task performance. Figure 4 shows the results of the 2-AFC study. We see that the average human percent correct consistently improves with more of the low frequencies being fully acquired. For this task, blurring seems to affect task performance less than aliasing. An interesting aspect is that having less aliasing by changing the acquisition has more of an impact on task performance than by using a regularizer.12–14

Figure 4.

Figure 4.

Open in a new tab

The average percent correct in the 2AFC trials consistently improved with a larger fraction of the low k-space frequencies being collected.

The results for the search task (Figure 5) show a different pattern than for the 2-AFC task. After an increase in average performance when going from a primarily aliased image (0% of the low frequencies fully acquired) to just 2.5% of the low frequencies fully acquired, the average performance is somewhat constant. In the search task, it seems that only the most aliased condition leads to a difference in average performance. There is a range of aliasing and blurring that has similar task performance.

Figure 5.

Figure 5.

Open in a new tab

The average percent correct in the 2AFC trials seemed to improve from the predominantly aliased image (0%) to having even a small percentage of the low frequencies fully sampled (2.5%). The performance on average remained fairly constant after that.

Future work will include a multireader multicase analysis of the variance of the results10,11 . The search results have a pattern that is more consistent with common practice in MRI (fully sampling 5–10% of the low frequencies) and lead to different choices than the 2-AFC task. One possible explanation for the low-resolution optimum in the 2-AFC task with variable backgrounds could have been than the high frequency content of the signal was too low (the signal was too big). Figure 6 shows that the signal has frequency content beyond what is acquired in the 20% acquisition. Future work includes exploring the frequency content of the backgrounds and signals as the acquisitions vary.

Figure 6.

Figure 6.

Open in a new tab

The k-space of the signal contains information beyond the 20% acquired in the lowest frequencies. The image shows the magnitude of the k-space of the signal with lines that show the boundaries of the regions acquired by the primarily blurred image (20% of the low frequencies fully acquired).

4. CONCLUSION

Previous work has studied task-based evaluation of undersampling in MRI using model observers12–15 but to our knowledge, this is the first study that optimizes the task performance for various undersampling strategies in MRI using human observers with a 2-AFC or a search task.

Optimizing the task performance for various undersampling strategies in MRI lead to different conclusions if using a 2-AFC task versus a search task. The 2-AFC task leads to a conclusion that the choice with least resolution (and least aliasing) optimizes task performance even with varying backgrounds. The search task has performance that remains fairly constant after the largest amount of aliasing is removed.

ACKNOWLEDGMENTS

We acknowledge support from NIH grant R15-EB029172, the Manhattan College Faculty Development Grant and the Kakos Center for Scientific Computing. The authors also thank Dr. Krishna S. Nayak, Dr. Sajan G. Lingala and Dr. Craig K. Abbey for their helpful insights.

REFERENCES

  • [1].Lustig M, Donoho D, and Pauly JM, “Sparse mri: The application of compressed sensing for rapid mr imaging,” Magn. Reson. Med 58, 1182–1195 (2007). [DOI] [PubMed] [Google Scholar]
  • [2].Pruessmann KP, Weiger M, Scheidegger MB, and Boesiger P, “Sense: Sensitivity encoding for fast mri,” Magnetic Resonance in Medicine 42(5), 952–962 (1999). [PubMed] [Google Scholar]
  • [3].Bernstein MA, King KF, and Zhou XJ, [Handbook of MRI Pulse Sequences], Elsevier Academic Press, London, UK: (2004). [Google Scholar]
  • [4].Jiang Y, Huo D, and Wilson DL, “Methods for quantitative image quality evaluation of mri parallel reconstructions: detection and perceptual difference model,” Magnetic Resonance Imaging 25(5), 712–721 (2007). [DOI] [PubMed] [Google Scholar]
  • [5].Barrett HH and Myers KJ, [Foundations of Image Science], John Wiley and Sons, Hoboken, NJ: (2004). [Google Scholar]
  • [6].Abbey CK and Barrett HH, “Human and model-observer performance in ramp-spectrum noise: effects of regularization and object variability,” Journal of the Optical Society of America A 18(3), 473–488 (2001). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [7].Gifford HC, Liang Z, and Das M, “Visual-search observers for assessing tomographic x-ray image quality,” Medical Physics 43(3), 1563–1575 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [8].Knoll F and Zbontar J, “fastmri: A publicly available raw k-space and dicom dataset of knee images for accelerated mr image reconstruction using machine learning,” Radiology: Artificial Intelligence 2(1) (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [9].Uecker M, Ong F, Tamir JI, Bahri D, Virtue P, Cheng JY, Zhang T, and Lustig M, “Berkeley advanced reconstruction toolbox,” Proc. Intl. Soc. Mag. Reson. Med 23, 2486 (2015). [Google Scholar]
  • [10].Chen W, Gong Q, and Gallas BD, “Paired split-plot designs of multireader multicase studies,” Journal of Medical Imaging 5(3), 031410 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [11].Gallas BD, Pennello GA, and Myers KJ, “Multireader multicase variance analysis for binary data,” J. Opt. Soc. Am. A 24(12), B70–B80 (2007). [DOI] [PubMed] [Google Scholar]
  • [12].O’Neill AG, Valdez EL, Lingala SG, and Pineda AR, “Modeling human observer detection in undersampled magnetic resonance imaging (MRI),” in [Medical Imaging 2021: Image Perception, Observer Performance, and Technology Assessment], Samuelson FW and Taylor-Phillips S, eds., 11599, 85 – 90, International Society for Optics and Photonics, SPIE; (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [13].Pineda AR, Miedema H, Lingala SG, and Nayak KS, “Optimizing constrained reconstruction in magnetic resonance imaging for signal detection,” Physics in Medicine and Biology 66(14), 145014 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [14].O’Neill AG, Lingala SG, and Pineda AR, “Predicting human detection performance in magnetic resonance imaging (MRI) with total variation and wavelet sparsity regularizers,” in [Medical Imaging 2022: Image Perception, Observer Performance, and Technology Assessment], Mello-Thoms CR and Taylor-Phillips S, eds., 12035, 1203511, International Society for Optics and Photonics, SPIE; (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [15].Chen Y, Lou Y, Kupinski MA, and Anastasio MA, “Task-based data-acquisition optimization for sparse image reconstruction systems,” in [Medical Imaging 2017: Image Perception, Observer Performance, and Technology Assessment], Kupinski MA and Nishikawa RM, eds., 10136, 101360Z, International Society for Optics and Photonics, SPIE; (2017). [Google Scholar]

RESOURCES