THE DETECTION OF METASTATIC PROSTATE CANCER WITH SIMULTANEOUS DUAL RADIOISOTOPE SPECT IMAGES:

PRELIMINARY RESULTS

ABSTRACT

Whole body, planar, and SPECT imaging with the indium¹¹¹ labeled monoclonal antibody capromab pendetide (indium¹¹¹MoAb CYT-356) has been shown to dramatically increase the detection of early spread of disease in prostate cancer patients. However, at the time of delayed imaging, In¹¹¹MoAb-CYT-356 can be seen in tumor tissue (including perivascular nodes) as well as in the vascular system. This can make metastatic tumor detection difficult especially if involved nodes are located in close proximity to blood vessels. We have developed an alternate imaging approach to the detection of metastatic prostate tumors, based on a single simultaneous In¹¹¹MoAb and Tc^99m RBCs SPECT acquisition. In this article we describe the "first" version of this method with preliminary results from twenty patients. This new method detected 59% more suspicious lesions than the routine evaluation of the In¹¹¹MoAb images alone, and led to the reversal of the negative diagnoses in three cases. Thirty one percent of these new lesions have been confirmed. Future versions of this method will include a statistical analysis of the differences of In¹¹¹ and Tc^99m images and a unique Compton scatter correction program for Tc^99m images (see APPENDIX II).

1. INTRODUCTION

Metastatic prostate cancer in pelvic lymph nodes is a common finding. Spread of prostate cancer occurs by both hematogenous and lymphatic routes especially along the pelvic and abdominal great vessels. It is estimated that lymphatic or extraprostatic extension of disease occurs in 40 to 60% of patients with clinically localized prostate cancer [1,2]. Early metastatic nodal disease from prostate cancer is usually small (< 1 cm) and can frequently be missed with high resolution anatomically based imaging procedures. Nevertheless, it is important to distinguish patients with confined prostate cancer (Stages A & B) from those with lymph node and/or distant spread (Stages C & D) because the corresponding treatment approaches can be dramatically altered. It is also important to distinguish local residual or recurrent disease in the prostatic fossa from nodal or distant metastases in treated patients with rising serum PSA values.

Staging newly discovered prostate cancer or restaging patients with suspected recurrent or residual disease has been limited in the past. The ability to accurately determine the presence of extra-capsular disease in newly discovered disease by serum PSA levels is limited by specificity [3]. Ultrasound imaging by itself is similarly limited because 50% of patients with presumed localized disease are found to have capsular penetration on pathologic review [4]. Digital rectal examination, although specific, is not sensitive [5]. Current high resolution anatomic imaging procedures such as computerized tomography (CT) suffer from poor sensitivity for detecting capsular penetration, seminal vesicle involvement and lymph node extension. Sensitivity for CT has been reported to be in the range of 6 to 30% [6]. Magnetic resonance imaging (MRI), with and without an endorectal coil, has been investigated to improve the accuracy of local staging, but thus far without reported success [7,8]. Therefore, in most patients lymph node status can only be accurately assessed by performing bilateral pelvic lymphadenectomy which does not accurately assess the presence of distant disease [9]. Many of the same staging limitations noted in the patient with newly discovered disease also exist for patients with suspected recurrent prostate cancer.

Recently the Food and Drug Administration (FDA) approved the indium¹¹¹ labeled monoclonal antibody (MoAb) capromab pendetide (In¹¹¹ MoAb 7E11-C5.3, CYT-356, or ProstaScint^TM) for staging men with prostate cancer. Initial clinical trials have demonstrated that this new staging procedure is particularly sensitive for detecting soft tissue metastases often found in pelvic and extra pelvic nodes [10-12]. One Phase III clinical trial reported that in 152 newly discovered prostate cancer patients who underwent surgery, capromab pendetide radioimmunoscintigraphy showed a sensitivity of 62%, specificity of 72% and an overall accuracy of 68% [12]. Another study reported that in 181 patients with suspected residual or recurrent prostate cancer following radical prostatectomy, the antibody localized in the prostatic fossa in 32 patients, in the fossa and extrafossa sites in 30 patients and in the extrafossa sites in 42 patients. Thus, 76 of 181 patients had MoAb scan evidence of extrafossa disease, including 42 patients with localization to abdominal lymph nodes when all other tests were negative or equivocal [11]. In another study, ProstaScint^TM sensitivity, specificity, positive and negative prognostic values for the detection of pelvic lymph node metastases were reported to be 59%, 85%, 72% and 76%, respectively. In contrast, the overall sensitivity of computed tomography (CT) in this population was 6%. The specificity of CT was high (100%; 108/108) because of the strong tendency toward negative scan interpretations [13].

The performance and interpretation of ProstaScint^TM radioimmunoscintigraphy can be difficult and labor intense. The manufacturer's recommended and FDA approved imaging procedure involves obtaining single photon emission computed tomography (SPECT) images of the pelvis thirty minutes post-infusion of In¹¹¹ MoAb. Day 0 images display the vascular anatomy and bladder activity which can be used when interpreting delayed images on days 4, 5, or 6. Activity on delayed images can be seen in the bone marrow, vascular structures and tumor tissue (if present). Nonspecific accumulation of activity can also be seen in the liver, small and large bowel, and the kidneys. Reconstructed 1 cm thick SPECT slices of the pelvis obtained on day 0 and day 5 are compared for the presence of prostate, prostatic fossa and nodal activity (disease). Accurate alignment of slices on these two separate days can be problematic. This can make metastatic tumor detection difficult especially if involved nodes are located in close proximity to blood vessels.

In this article we describe the imaging results of 20 patients who underwent the manufacturer's recommended imaging procedure plus an additional SPECT acquisition of the pelvis using two isotopes on day 5 (delayed imaging). We also describe the implementation and results of the first version of our computer blood pool subtraction method for detecting malignant lesions including the mathematical equations on which the method is based. Furthermore, we illustrate the application of the method with images produced by our computer subtraction program on several cases from our database.

2. MATERIALS AND METHODS

2.1 PATIENTS

All twenty patients who underwent ProstaScint^TM radioimmunoscintigraphy examination were part of a multicenter, phase III study sponsored by Cytogen Corporation (Princeton, NJ). Patients suspected of having primary (6) or recurrent (14) prostate cancer were studied (age range = 57-81 yrs; mean = 67.8). All patients had histologically confirmed adenocarcinoma of the prostate, elevated serum PSA values, and negative CTs/MRIs and bone scan studies. Patients with a second active primary malignancy, or serious illnesses involving the cardiac, respiratory, CNS, renal or hepatic organ systems were excluded from the study. Previous administration of a murine antibody (other than capromab pendetide) was also an exclusion criterion. Written informed consent was obtained from each patient in accordance with the guidelines established by the Institutional Research Board of the University of Illinois. Patients were asked to undergo the standard recommended imaging procedure (manufacturer's protocol) as well as an additional SPECT acquisition on day 5 using a second isotope. During the second SPECT acquisition on day five we used Tc^99m labeled red blood cell (RBC) activity (as well as In¹¹¹ MoAb activity) to define the vascular system, and to suppress the vascular component in In¹¹¹ MoAb based images. Accurate removal of the vascular structures requires flawless spatial co registration of the In¹¹¹ and Tc^99m images which is achieved during simultaneous SPECT acquisition of two isotopes (see a related analysis [14] of the propagation of the registration errors in brain SPECT imaging).

2.2 IMAGING PROTOCOL

SPECT images of the pelvis were obtained at 30 minutes and 5 days after injection of 5-6 mCi of In¹¹¹ CYT-356 MoAb. Five days post infusion was selected as the ideal time for delayed imaging based on a visual evaluation of image target-to-background ratios observed on days 4, 5, and 6 in five patient studies performed before this investigation (data not included). Patients were required to take a cathartic (4 liters of Colyte) on the afternoon and evening of day 4 post infusion. If significant colonic activity was seen on planar images on day 5, the imaging was discontinued and the patient was asked to take another dose of cathartic and return for imaging on day 6 post infusion. On day five (or day of delayed imaging) a 3 ml sample of whole blood was initially collected from the patient for RBC labeling using the Ultra-Tag^® Kit procedure (Mallinckrodt Medical, St. Louis, MO). During the RBCs labeling procedure whole body images and planar images of the chest, abdomen, and pelvis were obtained. SPECT images of the pelvis (and abdomen when indicated) were then obtained with care to estimate the exact position that was used on day 0. After all recommended planar and SPECT images were obtained, the patient was infused with 15 to 20 mCi of Tc^99m RBC to assure high counts in the Tc^99m window for our CT-SPECT registration program (not described in this paper). Thirty minutes after the infusion of Tc^99m RBCs the patient underwent a simultaneous dual isotope SPECT imaging of the pelvis (and abdomen if indicated).

Whole body and planar images were performed on a large-field-of-view single rectangular headed gamma camera (SMV America SOPHY DSX, Twinsburg, OH) interfaced with a dedicated computer. A medium energy collimator was employed and both photopeaks of indium¹¹¹ were used with 15% energy windows. Whole body images were acquired in a 1024 x 2084 matrix with a table imaging speed of 8 cm/minute. SPECT pelvic images were obtained using a dedicated triple-headed camera system (Picker PRISM 3000XP) with an attached UNIX based computer workstation (Picker Odyssey VP). Each camera head was equipped with a medium energy collimator. Fifteen percent windows were used for the two In¹¹¹ photons and a ten percent window for the Tc^99m photon. A narrow window of 10% was chosen for Tc^99m to decrease the contribution of amount of Compton scatter from the In¹¹¹ photons. Visual inspection suggested that the narrower window increased the signal-to-background ratio for the Tc^99m blood pool images. (It was estimated from a Tc^99m and Tc^99m-In¹¹¹ source study that a loss of 5.5% in the total counts occured when the window is narrowed from 15% to 10%). On the average, less than 8% total counts in Tc^99m window originated from In¹¹¹ downscatter. The raw projection images were acquired on a 128 x 128 matrix with 40 steps per head and 20 seconds per stop on day 0 and 50 seconds per stop on day 5. First order Chang attenuation correction was then applied to the filtered data set. Orthogonal images were displayed for visualization purposes.

Two sets of images were obtained for purposes of this study. Image Set 1 consisted of SPECT images of the pelvis obtained on days 0 and 5. Whole body and planar images obtained on day 5 were also included in this data set. Only In¹¹¹ based images were collected in this imaging set. Image Set 2 consisted of dual isotope SPECT images of the pelvis (and abdomen when indicated) obtained on day 5 as well as whole body and planar images obtained on day 5 before the injection of Tc^99m labeled RBCs. In¹¹¹ and In¹¹¹/Tc^99m based images were collected in Image Set 2.

2.3 IMAGE PROCESSING METHODS

The basic idea of the computer processing of reconstructed SPECT In¹¹¹ and Tc^99m images seems to be quite simple: just subtract the activity in Tc^99m image from the activity in In¹¹¹ image. However, a number of problems emerge when one analyzes the actual data:

	(1)	can hidden metastatic tumors be visualized by a simple subtraction technique?
	(2)	if yes, is a single subtraction enough to show all suspected lesions ?, and
	(3)	how much of the Tc99m image should be subtracted to obtain a meaningful difference image?

While we deal with these problems only briefly here, we address them and the mathematical aspects of the proposed solution in detail in APPENDIX I. For simplicity and for a moment, let us assume that the vascular activity components in both images are mutually proportional, i.e. that the vascular In¹¹¹ component is a constant fraction of the corresponding Tc^99m component. If one could find a region that exhibits only vascular uptake, this fraction would be known and a simple corresponding subtraction could be performed over all image couples of the study. However, based on the experience from twenty cases, not only is it difficult to find a vascular region consistently, the value of the fraction for optimal separation of different tumor sites and the vascular component in the same study is not constant. To avoid these problems, the question of elusive background and the problem of optimal normalization of intensities, we have developed a more general approach to the detection of the domains with disproportionally higher uptake of In¹¹¹ in comparison to that of Tc^99m. Our approach has resulted in several mutually complementary methods, that are in a large degree insensitive to errors/inconsistencies in intensity normalization. Consequently, while we normalized both In¹¹¹ and Tc^99m voxel intensities by scaling their 99.9 or 99.5 percentiles of the cumulative histogram of intensities to the same level (1000), any reasonable linear intensity scaling that maintains both images viable would yield acceptable results.

The first method, the method of dynamic subtractions, presents the results in the cine mode: a sequence of difference images is calculated by continuously increasing the amount of the subtracted Tc^99m image. This image sequence is then displayed either as forward-backward image loop or the operator interactively displays individual difference images. Examples of such a cine display together with the original In¹¹¹ and Tc^99m images are presented in FIG.1. One can observe the disappearance of the vascular component and reappearance of the previously masked In¹¹¹ structures. By visual analysis of this cine difference image sequence one can search for lesions that are not apparent in the original In¹¹¹ image. The operator is free to set the number and size of the subtraction steps (see the parameter A(t) in APPENDIX I).

The histogram blob method is based on the realization that in the 2 D histogram scatter plot of In¹¹¹ and Tc^99m voxel intensities, a protuberance (a blob) to the right (i.e., toward high In¹¹¹ intensities) of the main body represents the disproportionally high uptakes of In¹¹¹ and may consequently be a tumor signature (FIG.2). The voxels generating this tumor signature can be mapped back on the original In¹¹¹ image to point out the location of the tumor (FIG.3).

The 2-D histogram projection method is an attempt to compromise with the original intention to find the best single difference image. The constant fraction is determined by the direction of the straight line that separates the histogram's protuberance tumor signature from the rest of the histogram. As there may be more than one suspicious histogram blobs, each of them may require a separation line of a different direction and consequently more than one difference images may be needed.

It must be stressed that a disproportionally high uptake of In¹¹¹ is not necessarily associated with a metastatic tumor: it may be related, for example, to bone marrow activity and it is up to the nuclear medicine physician to interpret it correctly. The anatomical interpretation can be further aided by registration/fusion of CT/MRI image data with the SPECT images. . Details of this procedure are being prepared for publication.

3. CLINICAL RESULTS

Image Sets 1 and 2 were obtained as described in Section 2.2. Single and dual isotope SPECT images from the two sets of imaging data were separated in time by 30 to 40 minutes. At the end of the study, single SPECT images obtained on day 0 (from Image Set 1) and Tc^99m window images from the dual isotope SPECT images obtained on day 5 (from Image Set 2) were visually inspected and found to be essentially identical.

Data from Image Set 1 were presented as the official results of the clinical trial site to the sponsor. At the end of the study all 20 dual isotope SPECT imaging studies (Image Set 2) were processed on a Microsoft Windows based personal computer by the image subtraction software (APPENDIX I). Image Set 2 was read by the same nuclear medicine physician (MJB) 8 months after the evaluation of Image Set 1. Image Set 2 studies were read without referring to the films obtained from the manufacturer's recommended protocol. Readings based on Image Set 2 were then compared with the official results of the clinical trial (Image Set 1).

The results of the routine planar and SPECT film readings (Image Set 1) showed abnormal uptake in 14 patients and normal uptake in 6 patients (no evidence of disease). Two patients demonstrated evidence for prostatic fossa disease only, 8 for fossa and/or pelvic nodal disease, and 4 for fossa, pelvic and extra pelvic nodal disease. The results of the dual isotope SPECT film readings (Image Set 2) showed abnormal uptake in 17 patients and normal uptake in 3 patients (no evidence of disease). One patient demonstrated evidence for prostatic fossa disease only, 11 with fossa and/or pelvic nodal disease, and 5 for fossa, pelvic and extra pelvic disease. Lesions outside the pelvis and abdomen were detected on whole body images which were included in Image Set 2. These results are presented in Table 1 below.

Image Set 1 readings were compared with the results obtained from the planar and dual isotope SPECT dynamic subtraction images (Image Set 2) and from the histogram blob analysis. The results of the comparison are as follows: 7 pelvic node lesions were detected in computer generated images in 3 patients who were visually read as normal or having no evidence of disease. In 14 patients, 32 lesions were found visually and 44 by computer subtracted program. Both imaging protocols did not detect any lesions in three patients. Table 2 lists the 51 lesions detected in 17 patients using computer based images (Image Set 2).

The data from Image Set 2 after processing by the computer subtraction program increased the lesion detection rate by 62% of which 31% are presently confirmed. Clinical confirmation was accomplished with other imaging modalities, physical exam, and/or biopsy results. Two head and neck lesions and one lung lesion was confirmed by biopsy. Multiple abdominal and pelvic lesions were confirmed with CT/MRI findings during follow up examinations. Abnormal uptake was detected in 17 patients and no evidence of disease was noted in 3 patients.

The overall effect of the new dual isotope SPECT imaging protocol in the 20 study patients was the clinical upstaging of 3 patients (previously thought to be disease free). Three patients were found to be disease free by both imaging protocols. The same number of lesions were found in two patients by both imaging protocols. There was an increase in the number of pelvic and abdominal lesions detected by the dual isotope procedure in 12 patients. Although the increase in the number of lesions did not change the clinical staging in 12 patients, it gave a better indication of the amount of metastatic disease present.

All patients are being followed for verification of assumed positive lesions. Most unsuspected computer detected lesions were found in the pelvic lymph nodes. Bone marrow as well as small and large bowel activity continued to present scan interpretation problems with both protocols. Whole body planar imaging was found to be essential as a left supraclavicular node in one patient and single lung lesion in two patients would not have been found with SPECT imaging of the abdomen and pelvis alone. The supraclavicular node and one lung lesion have been confirmed.

4. DISCUSSION

Based on our limited experience of comparing 20 patient studies using the manufacturer's recommended imaging protocol with the dual isotope SPECT imaging and blood pool subtraction protocol, we found the latter method to be easier to read and interpret, and probably more sensitive. The overall clinical outcome of the new dual isotope SPECT imaging protocol was the upstaging of 3 patients (previously thought to be disease free). One patient was found to have extra pelvic disease (abdominal SPECT) by the dual isotope method that was not detected by the original protocol and this influenced the treatment approach. Three patients were found to be disease free by both imaging methods. The same number of suspicious pelvic nodes were found in two patients by both protocols and there was an increase in the number of suspicious extra prostatic fossa nodes detected by the dual isotope procedure in 12 patients. Although the increase in the number of lesions did not change the clinical staging in these 12 patients, it presented better estimate of the amount of metastatic disease present. The improved sensitivity of this new alternate procedure is presumed as only 31% of the "new" lesions have been confirmed. Consequently further follow up and a larger patient database are needed in order to confirm this increased sensitivity and to describe the interaction between sensitivity and specificity with this new technique.

The visual comparison of the day 0 SPECT images with the Tc^99m window images from the dual isotope SPECT images on day 5 assured us that both sets of data are essentially the same in terms of their respective ability to represent the blood pool. We found in this, and in a previous study that it is practically impossible to reposition patients for a subsequent scan on day 5 within an accuracy of 1/4 or 1/8 of a voxel [14]. It is also very difficult to achieve accurately aligned images with registration processing software after data acquisition [14]. Perfect registration of blood pool and delayed MoAb images can be obtained with a single simultaneous dual isotope SPECT acquisition where the following advantages are immediately realized:

	(1)	registration problems are eliminated,
	(2)	time required for a patient to lie still in a SPECT camera can be decreased (initial blood pool SPECT on day 0 can be eliminated),
	(3)	system throughput is doubled if day 0 imaging is eliminated and,
	(4)	patient scheduling is simplified.

There are a number of problems that still exist with this this new dual isotope procedure. For example, receiver operating characteristic (ROC) analysis was not performed during this pilot study because of lack of sufficient numbers of patients and comparably skilled readers/operators (ROC analysis will be applied in the later stages of this investigation). Consequently, we do not have enough information to evaluate specificity reliably. However, based on the limited data set available, we believe that the lost specificity is more than offset by the gained sensitivity.

Another problem is the verification of presumed lesions detected by our computer subtracted images which are not seen by the classic imaging protocol or by any other method. We found that some of the lesions detected by the dual isotope SPECT imaging of the pelvis with blood pool subtraction can be invisible to the naked eye in the original In¹¹¹MoAb images. Complete subtraction of normal and well defined structures such as blood vessels and urinary bladder can be dynamically observed on the computer screen. Both In¹¹¹ and Tc^99m activity can be seen in the urinary bladder and this structure can be localized and removed from our final display using our software. Only nodes and bone marrow which have excess In¹¹¹ counts (compared with Tc^99m) cannot be effectively subtracted. Long term followup of all patients is needed and is in progress to validate presumed positive lesions.

In summary, we report the establishment of a protocol for the simultaneous dual isotope SPECT imaging of the pelvis in men with prostate cancer utilizing In ¹¹¹-MoAb CYT-356 and Tc^99m-RBC on day 5 post infusion. We have also developed three methods for visualizing disproportional counts of In¹¹¹ MoAb CYT-356 (compared to Tc^99m RBCs) in order to aid in the detection of metastatic prostate tumor sites. Each of these methods alone provided superior results in comparison to the routine film reading of the pelvis. While the dynamic subtraction method may be more demanding of the operator's attention, we prefer it against the single subtraction approach as it provides more information about the consistency of the In¹¹¹ uptake and is more helpful in the detection of lesions associated with different subtraction coefficient . We have found that the histogram blob method is very useful for the confirmation of the results established by the subtraction methods and for investigation of the significance of the disproportional In¹¹¹ MoAb uptake. Attempts to utilize ROIs, drawn to define vascular domains in order to define a suitable subtraction coefficient, yielded inconsistent results, inferior to both histogram blob method and the histogram projection method.

As previously mentioned, the generated images should be interpreted by an experienced nuclear medicine physician. ProstaScint^TM image interpretation can be very difficult. Our data suggests that correctly identified presence of focal In¹¹¹ MoAb accumulation in lymph nodes by computer analysis of SPECT data can be very helpful in the diagnosis of metastatic tumors. As we have demonstrated [15], the extraction of the tumor information from the simultaneously obtained images is plagued by the Compton scatter spillover from one energy window into the other (i.e., from 245 keV and 171 keV In¹¹¹ window into 140 keV Tc^99m window). This can be corrected by calculating the Compton image in the Tc^99m energy window and subtracting it from the acquired Tc^99m image . While we have not implemented this correction in the first software version, the planned Compton scatter correction method [16,17] is briefly described in the APPENDIX II and the corresponding results will be presented in future communications. Further, practically every tumor's signature is contained in a cluster of voxels. Some tumors may have single voxel intensities too low to be detected as tumor voxels individually by a single-voxel statistical analysis (see (3) below).

Based on the experience obtained with the first, preliminary version of the techniques described above, we are working on a further improvement of the method and our current developmental work is concerned mainly with the following problems:

	(1)	a deeper analysis of the tumor signature (detection of latent blobs in 2-D histogram),
	(2)	the Compton scatter correction (see APPENDIX II),
	(3)	an analysis of clusters in dynamic difference images in order to estimate the statistical significance of the corresponding In111 uptake,
	(4)	automatic and optimized search for tumor signatures, and
	(5)	fusion of CT/MRI pelvic image data with SPECT images.

ACKNOWLEDGMENT

We thank Mrs. H. Levi and Mr. B. Patel for their conscientious and efficient management of patients and of the image data acquisition. This authors have contributed to the presented work in the following way: M. Blend initialized the project by recognizing the need for suppression of the vascular component, proposed to search for solution by the dual energy acquisition and subtraction, and while supervizing the project has been responsible for the clinical part of the project, including the clinical evaluation of results. K. Lin processed the raw SPECT data, extracted tomo reconstructed images for PC analysis and transformed the text of the paper into HTML format. J. Sychra developed the presented solutions based on the simultaneous acquisition approach, the corresponding algorithms and software, proposed the Compton scatter correction method, analyzed the SPECT data of twenty cases, and obtained the resulting images.

REFERENCES

[1] Murphy, G.P., Natarajan, N., Pontes, J.E., Schmitz, R.L., Smart, C.R., Schmidt, J.D., and Mettlin, C.: The National survey of prostate cancer in the United States by the American College of Surgeons. J. Urol, 127: 928-936, 1982.

[2] Scardino, P.T. : Early detection of Prostate Cancer. Urol. Clin. N. Amer. 126: 635-648, 1989.

[3] Kuban, D.A., El-Mahdi, A.M., Schellhammer, P.F.: Prostate specific antigen for pretreatment prediction and posttreatment evaluation of outcome after definitive irradiation for prostate cancer. Int J. Radiation Oncol Biol Phys. 32: 307-316, 1995.

[4] Hamper, U.M., Sheth, S., Walsh, P.C., Holtz, P.M., Epstein, J.I.: Capsular transgression of prostatic carcinoma: evaluation with transrectal US with pathologic correlation. Radiology, 178: 791-795, 1991.

[5] McSherry, S.A., Levy, F., Schiebler, M.L., Keefe, B., Dent, G.A., Mohler, J.L.: Preoperative prediction of pathological tumor volume and stage in clinically localized prostate cancer: comparison of digital recta examination, transrectal ultrasonography and magnetic resonance imaging. J of Urol, 146: 85-89, 1991.

[6] Platt, J. F., Bree, R.L., Schwab, R.E.: The accuracy of CT in the staging of carcinoma of the prostate. AJR, 149: 315-318, 1987.

[7] Rifkin, M.D., Zerhouni, E. A., Gatsonis, C.A., Quint, L.E., Paushter, H.M., Epstein, J.I., Hamper, U., Walsh, P.C., McNeil, B.J.: Comparison of magnetic resonance imaging and ultrasonography in staging early prostate cancer. N Eng J Med, 323: 621-626, 1990.

[8] Chelsky, M.J., Schnall, M.D., Seidmon, E.J. Pollack, H.M.: Use of endorectal surface coil magnetic resonance imaging for local staging of prostate cancer. J Urol, 150: 391-395, 1993.

[9] Sands, M.E., Zagars, G.K., Pollack, A., van Eschenbach, A. C.: Serum prostate-specific antigen, clinical stage, pathologic grade and the incidence of nodal metastases in prostate caner. Urology, 44: 215-220, 1994.

[10] Babasian, R.J., Lamki, L.M.: Radioimmunoscintigraphy of prostate cancer. Sem. Nucl. Med. 19: 309-326, 1989.

[11] Kahn, D., Williams, R.D., Seldin, D.W., Libertino, J.A., Hirschhorn, M., Dreicer, R., Weiner, G.J., Bushnell, D., Gulfo, J.: Radioimmunoscintigraphy with In¹¹¹ labeled CYT-356 for the Detection of Occult Prostate Cancer Recurrence. J of Urol 152: 1490-1495, 1994.

[12] Babaian, R. J., Sayer, J., Podoloff, D.A., Steelhammer, L. C., Bhadkamkar, V.A., Gulfo, J.V.: Radioimmunoscintigraphy of Pelvic Lymph Nodes with In¹¹¹-labeled Monoclonal Antibody CYT-356. J of Urol. 152: 1952-1955, 1994.

[13] Wright GL.Jr., Feng Q, Beckett ML, et al.: Characterization of a new prostate carcinoma-associated marker: 7E11-C5.3. Antibody, Immunoconjugates, and Radiopharmaceuticals 3:39, 1990.

[14] Sychra JJ, Pavel DG, Trepashko DW, Chen Y. The accuracy of brain SPECT activation images. Part 1: Propagation of registration errors. J. Med. Physics: 21(10) 1585-1590, 1994

[15] Sychra J, Blend M. A method for simultaneous dual energy SPECT imaging of prostate cancer. SNM Annual meeting, Denver, June 1996

[16] Sychra JJ, Blend MJ, Jobe T. A strategy for Compton scatter correction in brain SPECT by orthogonal image expansions assisted by neural networks. Neurological Research, vol. 17, 435-439, Dec. 1995.

[17] Sychra JJ, Blend MJ, Jobe T. Compton scatter correction in case of multiple crosstalks in SPECT imaging. Neurological Research, vol. 18, 31-32, Feb. 1996.

[18] Sychra JJ, Bandettini PA, N Bhattacharya, Q Lin. Synthetic images by subspace transform. Part 1: Principal component images and related filters, Med. Phys. 21(2):193-201, February 1994.

[19] Buvat I, Benali H, Todd-Pokropek A, Di Paola R. Scatter correction in scintigraphy: the state of the art. Eur J Nucl Med, 21:675-694, 1994

[20] Hannequin MJ, Younes R, Bellaton B, Bidet R. Scatter correction in planar imaging and SPECT by constrained factor analysis of dynamic structures (FADS). Phys Med Biol 35: 1451-1465, 1990

[21] Buvat I, Benali H, Frouin F, Bazin JP, Di Paola R. Target apex-seeking in factor analysis of medical image sequences. Phys Med Biol, 38:123-138, 1993

APPENDIX I

A1.1 INTRODUCTION

Let P and T be 3-D image blocks resulting from the tomographic reconstruction of SPECT image data obtained by simultaneous dual energy (In¹¹¹ and Tc^99m) pelvic scan, respectively, and let P(x) and T(x) be the corresponding voxel intensities (counts) at the location x = [x₁,x₂,x₃].

One may propose a simple model by assuming that T(x) is proportional approximately to the vascular component of P(x), i.e. the sought-after non-vascular component P_o(x) of P(x) is given by

Po(x) P(x) - A T(x) .

(A1.1)

However, the coefficient A is not easy to estimate. First, it is difficult to reliably find a location x where P_o(x)=0 and the remaining components of (A1.1) are large enough to calculate A directly from (A1.1). Not even a use of ROIs drawn over "purely vascular" domains of P_o(x)=0 yields consistent estimates of A. This can be explained by A being not a constant but a function of the location x, which is also confirmed by 2-D histograms H(P,T) - see below. To overcome these difficulties, we have avoided the model (A1.1) with constant A, and developed mutually complementing methods for the detection of the nonvascular component (more precisely: methods for detection of domains with disproportionally higher uptake of In¹¹¹ in comparison to that of Tc^99m).

A1.2 THE METHOD OF DYNAMIC SUBTRACTION

The first approach is based on a visualization of a progressive dynamic subtraction of the Tc^99m image from the In¹¹¹ image. In this case the parameter A in (A1.1) becomes a function of time and the resulting difference image D is displayed in a cine mode:

D(x,t) P(x) - A (t) T(x) .

(A1.2)

The method has been implemented on an IBM PC compatible computer, under MS Windows, using 4th generation language IDL, and its execution consists of the following steps:

A1.2.1 The image input and preprocessing.

At start, the "raw" reconstructed image blocks P and T are read into the computer memory. Currently, two sizes of the image cubes are supported: 64³ or 128³ voxel cubes. Next, sagittal, coronal and transverse integral views of the In¹¹¹ image data are displayed to enable the operator to select a tomographic slab for further analysis. For example, the integral view R in (x,y) plane is calculated by summing voxel intensities along the z-axis,

R(x,y) = P(x,y,z)

(A1.3)

A1.2.2 Definition of the analytical parameters.

During the following step, the operator selects the active view (sagittal, coronal or transverse, respectively). Subsequently, the range <a,b> of the tomographic slices (the main tomographic slab) to be analyzed is determined interactively by moving ranging bars on any of the other integral views. (Optionally, this initial slab may be automatically split into four slabs of approximately same thickness to be analyzed independently later). In the next step the range <A₁,A₂> of the parameter A is requested, as well as the its incremental step s, i.e. at the instance t the used value of A is

where t = 0,1, ..., (A₂ - A₁)/s. Further, to optimize the image displays, one may select a display intensity clipping (ranges) and a mixture of histogram modifying (including histogram equalization) and image filtering transforms (high/low pass filters). Any of these parameters may be selectively modified any time later, and the corresponding analyses can be executed without reloading the input image data.

A1.2.3 Display of the results.

One may either step manually through the display of the 2-D version of the differential image D,

D(x,y,A) P(x,y,z) - A T(x,y,z) .

(A1.5)

by interactively changing A, or the dynamic differential image is displayed in "infinite" cine loop with A changing forward and backward on the interval <A₁,A₂> in agreement with (A1.4). Examples of such a cine display together with the original In¹¹¹ and Tc^99m images are presented in FIG.1. As each displayed image of the series is independently intensity-wise normalized, one can observe, during the transition from pure In¹¹¹ image to a difference image with the progressively more subtracted Tc^99m component, disappearance of the vascular component and reappearance of the previously masked In¹¹¹ structures. By visual analysis of this cine image sequence one can search for lesions that are not apparent in the original In¹¹¹ image.

FIG. 1. The dynamic difference method.
The first column: Tc^99m images, the second column: In¹¹¹ images, the third column: dynamic (cine) difference images (In A Tc) with ten different values of the subtraction parameter A, the forth column: the location of the slab. Each row represents a different case: the first two cases are easier to diagnose from In¹¹¹ images than the last two).

A1.3 THE 2-D HISTOGRAM BLOB METHOD

Let us assume as an instance that (A1.1) is valid and that both In¹¹¹ and Tc^99m images contain only vascular components. In absence of noise, the graph of the 2-D histogram H(P,T) of In¹¹¹ (horizontal axis) and Tc^99m (vertical axis) intensities would be composed of segments of a straight line passing through the origin [0,0]. When a tumor or other tissues that uptake only In¹¹¹ are present they will be represented as horizontal histogram extensions to the right of the line

However, because of noise, because of uneven uptake of In¹¹¹ and Tc^99m in other tissues, and because A is not a constant, the actual histogram graph has a "smeared-like" form. Further, the actual histograms suggest that the idealization (A1.1) may be replaced by a piece-wise "broken-upward" straight line (or by two lines)

P - Ai T = Bi, I=1,2

(A1.7)

where the constants B₁ = 0 for P <P₁, and B₂ = P₁ ( A₁ - A₂ ) / A₁. P₁ is the idealized uptake level of In¹¹¹ above which the true and greater vascular Tc^99m vs In¹¹¹ uptake ratio

A2-1 = T / P

(A1.8)

is superimposed over the uptake ratio A₁^-1 of other tissues.

The tumor signature then often takes the form of a "protuberance" (a blob) to the right of the main body of the histogram scatter plot (FIG. 2). Once the voxels generating this signature are identified, they can be mapped back on the original In¹¹¹ image to suggest the location of the tumor (FIG. 3). The distance of the signature from the main body of the histogram is obviously associated with the magnitude of the disproportional uptake of In¹¹¹ in comparison to that of Tc^99m. (It is possible that a tumor may be represented by a "latent blob" that is "hidden" in the main body of the H(P,T) histogram. The development of a method for detection of these blobs is a subject of our current investigation).

FIG. 2. Two dimensional histogram of In¹¹¹(blue axis) and Tc^99m (red axis) intensities (counts) of two cases. Each dot represents a voxel in the intensity space. The flashing red "protuberances" to the right of the histogram obviously represent disproportionately high uptakes of In¹¹¹ vs. Tc^99m. These potential tumor signatures may be mapped back into the original images (see FIG. 3).

FIG. 3. Tumor detection by histogram blob method
The first column: the original In¹¹¹ image, the second column: a difference image, the third column: In¹¹¹ images with voxels corresponding to histogram protuberances marked in red, and the forth column: histograms, with protuberances displayed in red.

One has to stress that a disproportionally high uptake of In¹¹¹ is not necessarily associated with a metastatic tumor site. It may be related, for example, to bone marrow and it is up to the nuclear medicine physician to interpret it correctly. The anatomical interpretation can be further aided by registration/fusion of CT/MRI image data with the SPECT images.

A1.4 THE 2-D HISTOGRAM PROJECTION METHOD

The dynamic subtraction and the 2-D histogram blob analysis complement and strengthen each other. However, one may insist on a use of equation (A1.2) with a restriction to a single value of A (to a single difference image) only. As we have mentioned above, the use of ROIs drawn over "purely vascular" domains has yielded inconsistent estimates of A. A more consistent estimate can be obtained by fitting the straight line (A1.7, I=2) to the upper left edge of the graph H(P,T). Obviously,

A2 = - sin r / cos r , -90o < r < 0o ,

(A1.9)

The calculation (A1.2) of the difference image D is then equivalent to the orthogonal projection of the histogram H(P,T) on the vector e and to the subsequent reconstruction of the difference image D from these new synthetic voxel intensities. The factor e = [cos r, sin r] is the unit vector perpendicular to the line (A1.7) and aiming toward high P and low T. In other words, if a voxel x has intensity value h(x) = [P,T], the difference image D then has voxel value

(to avoid negative pixel values, a constant shift and/or clipping of the negative values is usually employed before the actual image is calculated and displayed).

To further increase the separation of the disproportional uptakes of In¹¹¹ in the difference image from the rest, the definition of the projection vector e can be modified by taking into consideration the corresponding protuberances of the 2-D histogram H(P,T). This may be simply achieved by drawing a straight line in the graph of the histogram H(P,T) to separate significant portions of the protuberances from the rest of the histogram. As above, the vector e is then the unit vector perpendicular to this new line (see the red line in FIG. 4).

FIG. 4. Histogram projection method
The green line corresponds to idealized Eq. (7, i=2), the red line is an example of the projection line to maximize the tumor contrast in the difference image

APPENDIX II

COMPTON SCATTER CORRECTION

The existing Compton scatter corrections (CSC) do not address the non-stationary nature of the Compton scattering in a satisfactory way. Many are based on unproven or incorrect assumptions[19] and are yielding results with unknown errors. An excellent review and critique of recent attempts to solve the CSC problem has been written by Buvat et al. [19], with a slight bias toward the factor analysis approach [20,21].

This appendix contains a brief, non-mathematical description of a new method of CSC. It is based on the approaches described in a more rigorous and mathematical manner in [16,17]. The proposed Compton scatter correction is designed to avoid the pitfalls mentioned in [19], especially that of the non-stationarity of the scatter. It will provide an estimate of the Compton image (the image that would be obtained in Tc^99m window when In¹¹¹ is present and Tc^99m is absent from the patient's body) from data obtained in the In¹¹¹ window and optionally from a window between the In¹¹¹ and Tc^99m windows during the simultaneous acquisition. The corrected Tc^99m image will be obtained by subtraction of the Compton image from the image calculated from data obtained in the Tc^99m energy window during the simultaneous acquisition. The proposed method also yields an estimate of the probable error of the calculated CSC. With improved tumor detection and delineation, the diagnosis and grading of tumors will be improved as well.

The amount of the "Compton-spilled-over" photons depends not only on the imaged object but also on the acquisition geometry (collimator, crystal, light guides and photo multiplier tubes, and the distance between the object and the collimator). However, we will not be concerned with acquisition geometry and will assume that it is kept constant. Consequently, the portability of the CSC will not be addressed as well. It is also assumed that, after scaling and normalization by a non-linear spatial co-registration of cases (see below), the errors of CSC caused by differences in the 3-D distribution of the scatter cross-section between individual patients can be neglected (such an assumption is routinely made by current attenuation correction methods).

It can be shown[16,17] that if the primary image (In¹¹¹) is a linear combination of 3-D sub-images then the corresponding Compton image may be approximated by the same linear combination of the corresponding Compton sub-images, i.e. the Compton function (mapping of the In¹¹¹ image on the Compton image) window may be viewed as a metric operator satisfying the additive property. Consequently, if the geometry of all cases in the database is the same and a studied primary image can be approximated by a linear combination of primary images of the database, then the Compton image is the same combination of the corresponding Compton images of the database. However, the pelvic geometries of cases in the real database differ. Nevertheless, after proper modifications, the idea of linear combinations image prototypes and of the corresponding Compton images may be still used to derive CSC:

1)	Let us assume that the available image data base consists of a representative sample of n cases, i.e. of the set of n In111 images, {Ii}i=1...n and of n corresponding Compton images, {Ci}i=1...n. To make studies comparable intensity-wise, normalize I-th In111 image Ii , to the same, preselected mean voxel intensity m by the factor si, I=1...n. Scale the corresponding Compton images by the same factor si, I=1...n.
2)	Select a case and co-register the rest of 3-D In111 images of the data base cases on this model image. The coregistration transform is assumed to be a nonlinear function f(aii,x), I=1...n, where x is the location and aii is the parameter vector defining this function for I-th case. Select the general form of the coregistration function f (for example, a multinomial of the second order).
3)	Perform two separate principal component (PC) analyses[16,18] on the image set {Iii}i=1...n and image set {Cii}i=1...n. in order to obtain the corresponding sets {Pii}i=1...n and {Rii}i=1...n , respectively, of PC images. (each In111 image I of the database can be then approximated by m significant PC images P I = vk Pk (A2.1) and each Compton image by h significant images R C = wk Rk (A2.2)
4)	The search for the Compton function can be obviously replaced by a search for mapping G, W = G(V,a,s) (A2.3) where W = [w1,w2,...,wh] and V = [v1,v2,...,vm]. The function G can be found by a back-propagation neural network trained on the available database (see [16]).
5)	Once the backpropagation neural network is trained it can be used for CSC on image data discussed in the article above, i.e. when In111 and Tc99m image data are acquired simultaneously:

a)	normalize the In111 image I by following steps (1,2) above (obtain s and a),
b)	obtain coefficient vector V of the PC expansion of the normalized image I,
c)	feed the neural network by V,a and s to obtain W.
d)	calculate the Compton image estimate by (A2,2),
e)	re-normalize (reverse steps (1) and (2)) the Compton image,
f)	subtract the Compton image from the uncorrected Tc99m image.

Based on experience with PC analysis of other types of medical images (fMRI, SPECT brain and cardiac planar radionuclide image sets) we expect that n=50 is a sufficient size of the training database and that number of significant PCS, h<m<10. However, a greater size of the data base will promote the accuracy of the derived CSC.