Samuel Richard1, Ehsan Samei. 1. Department of Radiology, Carl E. Ravin Advanced Imaging Laboratories, Duke University, Durham, North Carolina 27705, USA. samuel.richard@duke.edu
Abstract
PURPOSE: This work investigates a framework for modeling volumetric breast imaging to compare detection and estimation task performance and optimize quantitative breast imaging. METHODS: Volumetric reconstructions of a breast phantom, which incorporated electronic, quantum, and anatomical noise with embedded spherical lesions, were simulated over a range of acquisition angles varying from 4 degrees to 204 degrees with a constant total acquisition dose of 1.5 mGy. A maximum likelihood estimator was derived in terms of the noise power spectrum, which yielded figures of merit for quantitative imaging performance in terms of accuracy and precision. These metrics were computed for estimation of lesion area, volume, and location. Estimation task performance was optimized as a function of acquisition angle and compared to the performance of a more conventional lesion detection task. RESULTS: Results revealed tradeoffs between electronic, quantum, and anatomical noise. The detection of a 4 mm sphere was optimal at an acquisition angle of 84 degrees, where reconstructed images using a smaller acquisition angle exhibited increased anatomical noise and reconstructed images using a larger acquisition angle exhibited increased quantum and electronic noise. For all estimation tasks, accuracy was found to be fairly constant as a function acquisition angle indicating adequate system calibration, whereas a more significant dependence on acquisition angle was observed for precision performance. Precision for the 2D area estimation task was optimal at approximately 104 degrees, while precision of the 3D volume estimation task was optimal at larger angles (approximately 124 degrees). Precision for the localization task showed orientation dependence where localization was significantly inferior in the depth direction. Overall, precision for localization was optimal at larger angles (i.e., > 125 degrees) compared to the size estimation tasks. Results suggested that for quantitative imaging tasks, the acquisition angle should be larger than currently used in conventional breast tomosynthesis for lesion detection. CONCLUSIONS: Analysis of quantitative imaging performance using Fourier-based metrics highlights the difference between estimation and detection task in volumetric breast imaging and provides a meaningful framework for optimizing the performance of breast imaging systems for quantitative imaging applications.
PURPOSE: This work investigates a framework for modeling volumetric breast imaging to compare detection and estimation task performance and optimize quantitative breast imaging. METHODS: Volumetric reconstructions of a breast phantom, which incorporated electronic, quantum, and anatomical noise with embedded spherical lesions, were simulated over a range of acquisition angles varying from 4 degrees to 204 degrees with a constant total acquisition dose of 1.5 mGy. A maximum likelihood estimator was derived in terms of the noise power spectrum, which yielded figures of merit for quantitative imaging performance in terms of accuracy and precision. These metrics were computed for estimation of lesion area, volume, and location. Estimation task performance was optimized as a function of acquisition angle and compared to the performance of a more conventional lesion detection task. RESULTS: Results revealed tradeoffs between electronic, quantum, and anatomical noise. The detection of a 4 mm sphere was optimal at an acquisition angle of 84 degrees, where reconstructed images using a smaller acquisition angle exhibited increased anatomical noise and reconstructed images using a larger acquisition angle exhibited increased quantum and electronic noise. For all estimation tasks, accuracy was found to be fairly constant as a function acquisition angle indicating adequate system calibration, whereas a more significant dependence on acquisition angle was observed for precision performance. Precision for the 2D area estimation task was optimal at approximately 104 degrees, while precision of the 3D volume estimation task was optimal at larger angles (approximately 124 degrees). Precision for the localization task showed orientation dependence where localization was significantly inferior in the depth direction. Overall, precision for localization was optimal at larger angles (i.e., > 125 degrees) compared to the size estimation tasks. Results suggested that for quantitative imaging tasks, the acquisition angle should be larger than currently used in conventional breast tomosynthesis for lesion detection. CONCLUSIONS: Analysis of quantitative imaging performance using Fourier-based metrics highlights the difference between estimation and detection task in volumetric breast imaging and provides a meaningful framework for optimizing the performance of breast imaging systems for quantitative imaging applications.