A novel algorithm to derive robust internal respiratory signal for 4D CT and 4D MRI Cheukkai Becket Hui*, Zhifei Wen, Yelin Suh, Bjorn Stemkens, R.H.N. Tijssen, C.A.T van den Berg, Ken-Pin Hwang, Daniel Robertson, Tinsu Pan, Prajnan Das, Christopher H Crane, and Sam Beddar *Department of radiological physics University of Virginia Sep 2, 2015 MAC AAPM YIS
4D imaging Time-resolved images at different respiratory phases during a respiratory cycle Study tumor/normal tissue motion due to respiration Assign adequate margin for radiotherapy to compensate for target motion
External surrogates External surrogates track respiratory phases during acquisition RPM system www.varian.com Respiratory bellows learnmir.org
External surrogates Time consuming to set up Not always provide accurate respiratory pattern Example of artifactual RPM signal
Motivation Develop an algorithm to derive a robust internal respiratory (IR) signal for 4D images IR signal derived from IR surrogates, information from acquired data
Approach 1. Use multiple IR surrogates to derive IR signal Offset the impact of few potentially errant signals
Fourier-transform signal Spatial frequencies in Fourier-transform space as motion surrogates Magnitude k encodes the weight of the spatial frequency Phase φ(k) encodes the spatial shift of the spatial frequency Change in magnitude Intensity change F 1 st off-center frequency t Change in phase t Position change Time evolution of k (0,1) 4D CT data Fourier transformed data
Approach 1. Use multiple IR surrogates to derive IR signal Offset the impact of few potentially errant signals 2. Use clustering approach to select similar potential signals and filter dissimilar signals Remove noise and wrong information
Clustering algorithm IR surrogates independent from each other, and track different motion related changes Similarities in fluctuation patterns reflect true motion
Example 4D CT k (0,0) k (1,0,0) A-P φ(k (1,0,0) ) A-P k (0,0,1) S-I Clustered signals averaged to derive final IR signal φ(k (0,0,1) ) S-I V B V lung ρ lung
4D CT study Retrospective study 80 patients Derive IR signals from multi-slice 4D CT data IR signal match RPM signal in 72 patients Normalized cross correlation (NCC) 0.5 IR signal not entirely match RPM signal in 8 patients NCC < 0.5
Matching IR & RPM signal end inspiration RPM and IR-sorted images similar for matching signals IR signal and RPM signal of a 4D CT acquisition Radiation on IR signal RPM signal
Non-matching IR & RPM signal radiation on IR signal RPM signal end inspirations Respiratory phases Coronal view of RPMsorted images (clinical) Coronal view of IR-sorted images
4D MRI Data collected from 5 healthy volunteers Dynamic 2D images in sagittal view IR signal match bellows signal in 3 data sets IR signal not entirely match bellows signal in 2 sets
Matching IR-bellows signal IR signal IR signal and external bellows signal of an image slice
Non-matching IR-bellows signal IR signal IR signal and external bellows signal of an image slice
Summary Novel method to derive robust IR signal for 4D imaging Applicable to 4D CT and 4D MRI Correct for erroneous signal from external respiratory surrogate Potential to replace external surrogates and provide better workflow
Thank You Coronal and axial views of a 4D MR data set
Schematics of algorithm (CT)
IR surrogates usage rates (CT)
Diff in peak times (CT) Percentile plot (2%, 25%, 50%, 75% and 98%) of differences between the end inspiration times determined with the RPM signal and the IR signal at different anatomical regions in the cohort of matching 4D CT respiratory signals.
Comparison w/ IR surrogates (CT) NCC between potential signals from individual IR surrogates and RPM signal from the matching cohort
Non-matching IR & RPM signal radiation on IRS RPM signal end inspirations Respiratory phases Coronal view of RPM-sorted images (clinical) Coronal view of IR-sorted images
Image mismatched artifacts (CT) Mismatched artifacts quantified by the correlation metric (CM) Difference in CM between RPM- and IR-sorted images: CM = CM RPM CM IR Positive CM indicates less artifacts in IR-sorted images Baseline CM = 1.9 10-4 ± 3.7 10-4 in cohort of matching signal CM = 5.8 10-4 ± 5.7 10-4 in cohort of nonmatching signal (p-value = 0.0003)
Fourier signals (MR) The six Fourier elements used for internal respiratory surrogates and the movements of their corresponding spatial frequencies.
IR surrogate usage rates (MR)
Diff in peak times (MR) Difference in end inspiration times from IR signal and from the bellows signal (Δt ins ) versus its corresponding normalized cross correlation between the IR signal and the bellows signal (NCC IR-bel ). The data came from a cohort of matching IR and bellows signal.