Interventional X-ray quality measure based on a psychovisual detectability model Asli Kumcu, Benhur Ortiz-Jaramillo, Ljiljana Platisa, Bart Goossens, Wilfried Philips iminds-telin-ipi, Ghent University, Belgium CHO in Multi-Slice Images
Outline Interventional X-ray quality measure based on a psychovisual detectability model Background Design of interventional X-ray quality measure Results Conclusion & Future work 2
Clinical purpose of interventional X-ray Blockage of artery Stent opens artery Angiography procedure [1] [2] [3] [1] http://www.upmcphysicianresources.com/transradial/ [2] http://www.texasheart.org/ [3] http://www.aviva.co.uk 3
Interventional X-ray dose Clinical goal: Reduce dose to patient/staff (increases noise, affects contrast) Keep sufficient image quality State of the art: Assess dose to detector and use pre-programmed curves to modify X-ray output [4] Goal of this work: Assess perceived task-based image quality per acquisition (patient / anatomy / view) in real-time [4] AJ Gislason, et al., Allura Xper Cardiac System Implementation of Automatic Dose Rate Control, Philips Technical report, 2011 4
Interventional X-ray quality measure Task: Visibility (detectability) of vessels Metric estimates: Detection probability Quality Figure of merit: Ratio of # pixels with partial detectability to # all detectable pixels Quality FOM: 86% Clinical images acquired on Philips Allura with 100% dose and 50% dose with denoising 5
Detection probability Aim for dose which results in image parameters estimated to have 99.5% detectability P(det) =ƒ(contrast ratio, noise, background intensity) Probability of detecting object P(det) (%) 100 50 0 Quality too low (increase dose) X Parameter (e.g. contrast ratio, CR) Quality too high (reduce dose) Target: minimum contrast ratio (lowest dose) resulting in 99.5% detectability 6
Design of measure Interventional sequence acquisition Estimate image quality attributes Estimate detectability of clinical targets Psychovisual target detectability model Acquisition Dose feedback loop Quality model Quality target 7
Psychovisual target detectability model human experiments 1 up / 1 down staircase procedure Target Noise σ Noise types Local Background (cd/m 2 ) S loan σ=0 Static noise 59 L etters σ 1 =0.019 156 σ 2 =0.087 Dynamic noise 256 348 8
Psychovisual target detectability model results [5] Detectability reduced in higher noise and darker backgrounds Local background luminance (L LB ) σ 1, L LB = 59 cd/m 2 σ 2, L LB = 254 cd/m 2 [5] A. Kumcu, et al., Effects of static and dynamic image noise and background luminance on letter contrast threshold, QoMEX 2015 9
Estimate image quality attributes Interventional sequence Luminance domain Detectability Contrast ratio Noise (σ) Background luminance Psychovisual target detectability model 10
Estimate image quality attributes Contrast & background intensity Weber contrast computed from mean foreground and background intensity using local content informationbased contrast ratio [6] or shearlet-based [7] contrast ratio Noise variance Spatial noise estimator, extension of [8]: incorporates noise model which takes into account relationship between pixel intensity and noise [6] B. Ortiz, et al, Computing contrast ratio in medical images using local content information, MIPS XVI conference 2015 [7] B. Goossens, et al., "Efficient Design of a Low Redundant Discrete Shearlet Transform, " in Proc. 2009 International Workshop on Local and Non-Local Approximation in Image Processing (LNLA2009), August 19-21, 2009, Tuusula, Finland, p. 112-124. [8] V. Zlokolica, et al, "Noise estimation for video processing based on spatial-temporal gradient histograms," IEEE Signal Processing Letters, 2006, 13, 337-340. 11
Results interventional neurology (DSA) 100% dose 50% dose + denoising Frame Contrast ratio Frame Contrast ratio Noise Detectability Noise Detectability Quality FOM: 82.5% 83.7% Human scores from VGA experiment: 100% for both sequences 12
Results interventional cardiology 100% dose 50% dose + denoising Frame Contrast ratio Frame Contrast ratio Noise Detectability Noise Detectability Quality FOM: 86% 87% 13
Needed contrast DECREASE (%) Needed contrast DECREASE (%) needed needed Needed contrast INCREASE (%) Needed contrast INCREASE (%) Results alternative quality FOM Contrast too high: contrast decrease needed Contrast too low: contrast increase needed High dose sequence Lower dose sequence 14
Limitations Signal model 1 (complex) frequency Consider evaluating additional signal frequencies with vessel-like objects or characterize entire CSF White noise Consider extending psychovisual experiments to complex backgrounds 2 observers Follow-up psychovisual study planned with additional observers and parameters 15
Conclusion & Future work Task-based measure for real-time quality assessment in interventional X-ray Currently index is pixel-based go to object-based index in the future Include target motion in model Extended comparison to existing vision models for dynamic noise Extended validation with observers: effect of dose 16
Acknowledgments This work was supported by the Eniac PANORAMA project www.panorama-project.eu Thanks to project partners Philips and University of Leeds, and cardiologists at UZGent 17
Thank you! Questions? Interventional X-ray quality measure based on a psychovisual detectability model Asli Kumcu, Benhur Ortiz-Jaramillo, Ljiljana Platisa, Bart Goossens, Wilfried Philips iminds-telin-ipi, Ghent University, Belgium CHO in Multi-Slice Images