In order to estimate organ dose, we need first be able to accurately quantify organ activity at one time point. In vivo organ activity estimations are usually done by conventional planar (C-Planar) processing methods which are based on conjugate-view whole body scans and geometric mean (GM) method. Figure 1 below shows typical steps in a C-Planar method. This method is based on acquisition of anterior and posterior images that are scatter corrected using the triple energy window method. Attenuation compensation is the performed using the geometric mean method using thickness correction with a measured transmission image. ROIs are then drawn on the planar image. Since organs overlap in the 2D projections, and since it is hard to see the true organ outlines in the 2D projections, ROIs are often drawn smaller than the true projected organ size and, in some cases, additional corrections for background and organ overlap is performed. The last step is the use of a calibration factor to convert counts in the ROI to total organ activity.

Figure 1. Typical steps of C-Planar method

There are some obvious shortcomings of the C-Planar method. First, the GM method only approximately compensates for attenuation. Also, the triple energy window method is only approximate, results in an increase in image noise, and may require more energy windows than are available on commercial gamma cameras. Also, there are no compensations for collimator-detector-response or partial volume effects, which are very important for tumor dosimetry. Perhaps the major problem is organ overlap and background activity, as illustrated in Figure 2.

Figure 2. Overlapping problems in the C-Planar method

The ROIs on the leftmost image are the 2D projection of the 3D organ VOIs defined in co-registered SPECT and CT images. We can see that the organs overlap in the 2D projection view. In C-Planar methods, smaller ROIs are often used to reduce the overlap and the impact of background activity. However, this will result in leaving out activity in some portions of the organ. In some methods, the missing activity is extrapolated based on the activity in the non-overlapping portion and background activity is estimated using background ROIs. However, the entire process is very subjective and it is difficult to know a priori, for example, which set of ROIs will provide more accurate activity estimates. For example, the ROIs in the rightmost image will drastically reduce the effects of background and overlap, but result in leaving out activity larger portions of each organ.

One potential solution to the overlap and background problem is the use of SPECT images, as demonstrated in Figure 3. Since we can define organ volumes of interest in 3D, the overlap problem is largely eliminated. However, SPECT images without compensation for physical effects are not quantitative. Even with compensation, the limited resolution will result in partial volume effects that will degrade quantitative accuracy. We have developed and are continuing to develop quantitative SPECT (Q-SPECT) methods in our lab that provide comprehensive compensation for these effects. We have applied, refined, and evaluated these methods for organ activity estimation.

Figure 3. The three leftmost columns show transaxial views of SPECT, CT and fused SPECT/CT images, respectively, overlaid with the manually-defined organ ROIs for the corresponding slice. The rightmost 3 columns show coronal (top row) and sagittal (bottom row) slices from the (left to right) SPECT, CT, and fused SPECT/CT images.

In the Q-SPECT methods, we used the OS-EM algorithm with the compensation for attenuation, scatter, and the full collimator-detector-response, which includes modeling penetration and scatter in the collimator. A perturbation-based geometric transfer matrix (pGTM) method developed in our lab was used for partial volume compensation. Organ activities were then calculated from the total reconstructed image intensity divided by the system sensitivity.

Although, Q-SPECT can provide very accurate activity estimates, there are also some limitations on using it clinically. First, it requires more complex imaging protocols. For example, it requires 2 scans to cover all the important organs without the whole-body SPECT capacity. However, this may become a less important factor in the future if spiral scan whole-body SPECT protocols are developed. The acquisition time is normally longer than WB scans, though in other imaging applications there is some data to suggest that good image quality can be obtained with the same acquisition time as for the planar scans. Also, the computation time is much longer than for QSPECT methods and the definition of the 3D VOIs is tedious since robust and general automatic segmentation tools are not available. However, the availability of co-registered SPECT/CT does ease and improve the accuracy of image segmentation.

To address some of the limitations of Q-SPECT, we have developed a quantitative planar method we will refer to as Q-Planar. As demonstrated in Figure 4, we still use the conjugate whole body scans, but using an iterative algorithm to estimate the organ activities from these two projections in Q-Planar method. As shown here, we still make use of 3D information from CT or SPECT. First, we segment out each organ in 3D, and uniformly fill each VOI so sum of all the voxels inside the VOI is 1. Then we use our model-based projector to project the VOI image. The true projection will then be a linear combination of the separate organ VOI projections. Since the projections are Poisson distributed, we can use the ML-EM algorithm to estimate the scale factors, Ar, for the organ VOI projections and these scale factors will simply be the total activity in the organ. The major difference between this and SPECT reconstruction is that we need to estimate only a few parameters and thus can estimate them quite effectively, as we will show, from only 2 projection views.

Figure 4. Flow chart of activity quantification steps in Q-Planar method. First, each organ VOI is segmented from the registered 3D SPECT/CT data, and uniformly filled so that the sum of all the voxels inside the VOI is 1. Then the projection of each organ VOI is calculated using the model-based projector. The true projection will then be a linear combination of the separate organ VOI projection. Since the projections are Poisson distributed, we can use the ML-EM algorithm to estimate the scale factors, Ar, for the organ VOI projections and these scale factors will simply be the total activity in each organ.

Since we can estimate the organ activity from both planar and SPECT scans, there are at least three different acquisition strategies for estimating cumulated activity or residence time. As listed here:

1.         A time series of conjugate view whole body scans;

2.         A time series of SPECT scans;

3.         A time series of conjugate view whole body scans plus a SPECT scan at one time point, which referred as the hybrid method.

The residence times can be estimated using a class of methods for each protocol. Note that the planar acquisition protocol will be, in general, easier to implement than the SPECT acquisition protocol, and the hybrid protocol is a practical compromise between the two.

Figure 5. Flow chart of residence time quantification steps in planar, SPECT, and hybrid planar/SPECT methods.

Representative Result

We performed a Monte Carlo (MC) simulation (MCS) study using the 3D NCAT phantom. The organ activity concentrations were based on the averages from 8 clinical studies using Indium-111 Zevalin. We used a non-uniform activity distribution in the heart and lungs.

The simulation parameters were appropriate for a GE VH/Hawkeye camera with a 1 inch crystal and a medium energy general purpose collimator. We used a modified version of SimSET and PHG code that includes modeling of collimator interactions to simulate SPECT projections and planar images. The low-noise projections were generated using this code. The resulting organ and background projection images were scaled to represent organ activities at different time points (1, 5, 24, 72 and 144 hours) based on the average organ time activity curves obtained from 6 patient studies and summed to form a low-noise set of projections for each time point. Fifty different Poisson noise realizations were generated to study the precision of the methods. The SPECT projections were simulated at 120 views over 360 degree, the count level at 24 hour is equivalent to 30 seconds scan per view. The count level of planar scans at 24 hour is equivalent to 20 minutes whole body scans.

Figure 6. From left to right the images are: coronal slice through activity distribution, same coronal slice through attenuation map, low-noise anterior projection, the noisy anterior SPECT projection, and the noisy anterior planar projection.

As mentioned, for the CPlanar method, organ overlap is an important issue. For the MC study we investigated 3 different cases of overlap correction, as demonstrated in Figure 7.

1.         For the ideal correction we took advantage of the fact that all the organs were simulated separately in the MC simulation. We applied the CPlanar method to the projection of each organ to estimate the organ activity. Thus there was no overlap with other organs and no background activity. This is an ideal case and no real method could do better than this.

2.         At the other extreme is no correction. In this case we defined the planar ROIs based on projections of the true 3D organ VOI images, as shown in the topmost image. This method has the maximum overlap.

3.         In between these two extreme cases, we investigated what a realistic overlap and background correction based on the use of manually drawn ROIs. In this case, organs ROIs were intentionally drawn smaller to avoid overlapping. No additional overlap or background corrections were performed. As I mentioned earlier, this method is somewhat subjective, and thus we used the previous two methods to provide brackets for the kind of variability one might have.

Figure 7. Three different cases of overlap correction in C-Planar Method.

Figure 8 shows the percent errors of activity estimates for C-Planar method with the 3 different overlap corrections, Q-Planar and Q-SPECT. The vertical axis is the percent error in organ activity estimate, calculated as the estimated value minus true value divide by true value. A positive error indicates overestimation. The error bars were calculated from the fifty different noise realizations. Please note this is precision due to noise, which is measured using a single phantom over many noise realizations.

Figure 8. Percent errors and standard deviations of errors in organ activity estimates for C-Planar, Q-Planar, and Q-SPECT methods.

The results of this study show that C-Planar w/o overlap correction performed worst, as expected, with errors up to 380%. Even with ideal overlap correction, the C-Planar method still produced errors in the range of -8% to 12%. Most of these errors resulted from the approximate scatter and attenuation compensations. When using manually defined ROIs, the errors were generally between the other two extreme cases, and is perhaps indicative of what would be realized clinically. The precisions for all the methods were similar and much smaller than the errors, indicating that noise is not as important a factor as bias from imaging and processing methods. The results also showed that Q-Planar was significantly better than realistic C-Planar, with the accuracy approaching that for Q-SPECT and with slightly better precision.

Figure 9 shows the percent errors of residence time estimates for C-Planar, Q-Planar, Q-SPECT, and hybrid planar/SPECT methods. Similar to the errors in activity estimates, the results show that C-Planar w/o overlap correction performed worst, with errors up to 302%. Even with ideal overlap correction, the C-Planar method still produced errors in the range of -8% to 7%. When using manually defined ROIs, the results was somewhat between the other two extreme cases. The hybrid C-Planar (Realistic) with QSPECT method performed better than C-Planar methods alone. Q-Planar or hybrid Q-Planar/Q-SPECT was significantly better than C-Planar methods, with the accuracy approaching that for Q-SPECT.

Figure 9. Percent errors and standard deviations of errors in residence time estimates for C-Planar, Q-Planar, Q-SPECT, and hybrid planar/SPECT methods.