Cadiac C-Am CT SNR Enhancement by Combining Multiple Retospectively Motion Coected FDK-Like Reconstuctions M. Pümme 1, L. Wigstöm 2,3, R. Fahig 2, G. Lauitsch 4, J. Honegge 1 1 Institute of Patten Recognition, FA Univesity Elangen-Nuembeg, Gemany 2 Depatment of Radiology, Stanfod Univesity, USA 3 Cente fo Medical Image Science and Visualization, Linköping Univesity, Sweden 4 Siemens AG, Medical Solutions, Fochheim, Gemany Email: puemme@infomatik.uni-elangen.de Abstact. Cadiac C-am CT is a pomising technique that enables 3D cadiac image acquisition and eal-time fluooscopy on the same system. Retospective ECG gating techniques have aleady been adapted fom clinical cadiac CT that allow 3D econstuction using etospectively gated pojection images of a multi-sweep C-am CT scan accoding to the desied cadiac phase. Howeve, it is known that etospective gating of pojection data does not povide an optimal signal-to-noise-atio (SNR) since the measued pojection data is only patially consideed duing the econstuction. In this wok we intoduce a new econstuction technique fo cadiac C-am CT that povides inceased SNR by including additional coected and esampled filteed back-pojections (FBP) fom tempoal windows outside of the tageted econstuction phase. We take advantage of seveal motion coected FDK-like econstuctions of the subject to incease SNR. In the pesented esults, using in vivo data fom an animal model, the SNR could be inceased by appoximately 30 pecent. 1 Intoduction The combination of eal-time pojection imaging with 3D imaging modalities in the inteventional suite is becoming moe impotant as pocedues incease in complexity. One such combination, x-ay fluooscopy with cadiac C-am CT, is unde development. Retospectively ECG gated FDK [1] econstuctions (RG- FDK) povide pomising image quality fo cadiac C-am CT as shown by Lauitsch et al. [2]. Image pesentation duing a pocedue often equies new imaging algoithms fo segmentation and image egistation, and such algoithms have stingent need fo high image quality including signal-to-noise atio (SNR) and contast-to-noise atio (CNR). In cadiac C-am CT, the etospective selection of pojection images whose ECG time is closest to the cadiac phase that is desied fo econstuction consides only 1 N s of the measued pojection data of an e.g. N s 4s multi-sweep scan, which does not maximize the potential SNR
223 of the 3D econstuctions. In this wok we intoduce a technique that allows use of all pojection images fom a multi-sweep scan and theefoe can povide enhanced SNR in the econstucted volume. A simila appoach fo espiatoy motion coection and SNR enhancement was intoduced by Li et al. [3]. The appoach we intoduce is a tade-off between the spatial esolution povided by a etospectively ECG gated FDK econstuction using only a selected subset of all acquied pojection images and a econstuction that uses all measued pojection data (and theefoe has good SNR) but may exhibit educed spatial esolution due to appoximations applied duing the tempoally dependent spatial motion coection. 2 Methods Let P be the set that contains all acquied pojection images of a multi-sweep scan, then P t is a subset of P that povides the etospectively selected pojection images of cadiac phase t (defined in pecent between subsequent R-peaks); fo each pojection angle one pojection image is selected that is closest to the cadiac phase t. Pβ t denotes the β-th pojection image whee the index β is odeed accoding to an inceasing pojection angle of a shot-scan pojection data set P t. The cadiac phase of a pojection image is denoted by τ(pβ t ). The effective cadiac phase (ECP) of a econstuction using the set of images P t is τ E (t) = 1 P t P t τ(p t β) (1) As mentioned above we have a tade-off between spatial esolution and SNR enhancement due to an appoximate motion coection scheme. A RG-FDK volume of the desied cadiac phase t is econstucted to povide a baseline of spatial esolution whee the pojection images ae etospectively selected such that the obseved ECP τ E (t) is closest to t. This volume is denoted by ft RG (x) and x is a 3D gid of the econstucted volume intensities. It can be seen as a fist sample fom multiple volume econstuctions that ae late combined to enhance SNR. Howeve, duing the econstuction of ft RG (x) only 1 N s of the pojection data is used. To make use of the emaining unused pojection data, we apply etospective motion coection. β=1 2.1 Retospective motion coection As shown by Pümme et al. [4] inceased tempoal esolution can be achieved by computing a 4D motion vecto field U(t) (MVF) of the subjects individual heat motion using image egistation as intoduced by Modesitzki [5]. The 3D MVF u t := U t (t) descibes the elative 3D defomation of each voxel between a selected efeence cadiac phase t (RCP) that is desied fo a econstuction with enhanced SNR, and cadiac phase t. The MVF is combined with an FDKlike algoithm [4] that allows a tempoally dependent spatial waping of the
224 filteed-backpojections P t β i (x). Since we apply a voxel-diven back-pojection the FBPs ae defined on the 3D gid x. The integal in the FDK-like algoithm [1], that we call FDK-4D, ove all discete β i can be witten as the sum f t (x) = P t i=1 P t β i (x u τ(p t βi )) (2) of the tempoally dependent spatially waped FBPs accoding to the MVF. This allows the use of additional pojection images duing the econstuction that would, due to thei cadiac phase, othewise intoduce motion atifacts. 2.2 Multiple volume econstuction Since we coect fo motion we can ceate seveal e.g. N m abitay subsets P l of pojection images fo a shot-scan, that ae not etospectively gated. Fo each of the N m subsets a motion coected volume f t l (x) accoding to (2) is econstucted using FDK-4D. The selection stategy of the pojection data implies that each measued pojection image of a multi-sweep scan is contained in at least one of the subsets P l. The multiple volume econstuctions ae then voxel-wise combined using a 1D Gaussian window G µ σ x(i) whee the mean µx := ft RG (x) is defined by the voxel intensity of the RG-FDK econstuction f RG, i is the intensity and standad deviation σ. The SNR enhanced econstuction f SNR using the poposed algoithm, that we call Multiple-Volume-FDK (MV-FDK), is then given by ( Nm ) ft SNR 1 (x) = G µ σ x( 1 + Nm f G µ σ x( f t l (x)) f t l (x) + ft RG (x) (3) t l (x)) l=1 3 Results l=1 To investigate impovement of image quality using the MV-FDK algoithm, a seies of ten RG-FDK econstuctions (using 191 etospective selected pojections fo the econstuction) was pefomed. These initial econstuctions wee used to compute the MVF elative to the cadiac phase t = 80. Using the FDK-4D algoithm six motion coected volumes, each of which using a patially distinct set of pojection images, wee econstucted such that all 1146 acquied and (accoding to thei phase distance to t = 80) spatially waped pojection images wee consideed duing MV-FDK econstuction. Diffeent standad deviations σ fo the Gaussian weighting between the multiple econstuctions whee investigated as shown in Figue 1. The MV-FDK econstuctions ae compaed against the RG-FDK econstuction. The between the aows measued SNR (defined as the atio of the mean intensity value to the standad deviation) is (σ, SNR) = (15, 66.07), (30, 73.58), (60, 74.45), (90, 74.14), (120, 73.94), (150, 73.73), (200, 73.73), (RG F DK, 57.7) (see Fig. 1). An example of epesentative image quality of a RG-FDK econstuction in compaison with an MV-FDK econstuction is shown in Figue 2.
225 Fig. 1. Intensity pofile plot along a line as shown in the top left image. The image shows a slice of an MV-FDK econstuction (6 4s multi-sweep scan, using 1146 pojection images duing econstuction). The intensity pofile of the complete line is shown in the top ight panel. A magnification of the intensity pofile measued between both aows is shown in the bottom panel. Each MV-FDK econstuction is a combination of six econstuctions, that ae econstucted using seveal patially distinct subsets of the acquied pojection data. Inside the left venticle the intensity pofile of the MV-FDK econstuction is moe homogeneous compaed to the RG-FDK econstuction 4 Conclusion and discussion We could show that using the MV-FDK algoithm the SNR can be impoved by appoximately 30 pecent compaed to a standad RG-FDK econstuction. The measued intensity pofile inside the venticle shows a highe SNR. Intensity vaiations inside high density egions like the contast filled venticle ae deceased and the location of edges emain while only a vey slight bluing is noticed such that impoved segmentation esults fo clinical applications can be expected. In conclusion we can say that impovement of SNR by including additional coected and esampled pojections fom cadiac phases outside the tempoal window of the tageted econstuction phase duing econstuction can be achieved. 5 Acknowledgments This wok was suppoted by Siemens AG, Medical Solutions, Fochheim, Gemany, NIH gant R01 EB 003524 and by the Lucas Foundation, HipGaphics, Towson, Mayland, USA, BaCaTec and by Deutsche Foschungsgemeinschaft (DFG), SFB 603, TP C10.
226 Fig. 2. Column (a) shows thee othogonal multi-plana econstuctions (MPRs) of the heat fom an animal model using RG-FDK whee 191 etospectively selected pojection images wee used (t = 80). Column (b) shows a multiple-volume FDK econstuction consideing 1146 acquied pojection images fom a 6 4s scan. The standad deviation of the Gaussian kenel used fo the MV-FDK econstuction (b) was σ = 60. The MV-FDK econstuction (b) is less noisy compaed to the RG-FDK econstuction (a) while edges in pinciple emain as povided by the RG-FDK appoach Refeences 1. Feldkamp LA, Davies LC, Kess JW. Pactical cone-beam algoithm. J Opt Soc Am A1 1984:612 619. 2. Lauitsch G, Boese J, Wigstöm L, Kemeth H, Fahig R. Towads Cadiac C-am Computed Tomogaphy. IEEE Tans Med Imaging 2006;25:922 934. 3. Li T, Scheibmann E, Thondyke B, Tillman G, Boye A, Koong A, et al. Radiation dose eduction in fou-dimensional computed tomogaphy. Med Phys 2005;32(12):3650 3660. 4. Pümme M, Wigstöm L, Honegge J, Boese J, Lauitsch G, Stobel N, et al. Cadiac C-Am CT: Efficient motion coection fo 4D-FBP. Pocs IEEE Medical Imaging Confeence, San Diego, 2006, accepted fo publication. 5. Modesitzki J. Numeical Methods fo Image Registation. Oxfod Univesity Pess, 2004.