US 4,856,528 AGrant
Tumor Volume Determination
Issue Date:1989-08-15
•10 Claims
•10 Drawing Sheets
Abstract
A computer-implemented arrangement (apparatus and method) for semi-automatically determining from CT image data the volume of a tumor. A primary application is to determining a tumor volume for patients with hepatoma and primary hepatic cholangiocarcinoma. CT image data of a scanned organ, stored on a magnetic tape, is read into a computer. This data includes for each of plural slices of predetermined thickness a plurality of pixels defining the slice and a CT number representing tissue density corresponding to each pixel. An image of the first slice is displayed on a monitor. A region of interest (ROI) approximating an outline of the organ is automatically generated, which outline is visually modified, if necessary, by an operator to closely correspond to the organ's outline. The particular pixels within the organ's outline and the CT numbers associated therewith are stored as a local histogram of the slice, the local histogram including data indicative of the number of pixels within the organ's outline having each unique CT number. Local histograms, produced in a similar manner for each slice, are summed to generate a global histogram indicative of the number of pixels within the organ outline's of respective slices having each unique CT number. From the global histogram, a demarcation CT number distinguishing tumor tissue from normal organ tissue is determined. From the global histogram and demarcation CT number, the organ volume and tumor volume are determined.
Metadata
Assignee
- John Hopkins University
Inventors
- Nai-Chuen Yang
- Peter K. Leichner
Application Information
Application Number:US 0665541
Filing Date:1987-06-26
Priority Date:1987-06-26
Art Unit:335
Classifications
IPC:
A61B 603
Field of Search:
364378382128358414;415;5649016;18;22653;659111
Patent Drawings (10 sheets)
Description
BACKGROUND OF THE INVENTION
This invention relates in general to the radiological arts. More specifically, it provides a semi-automated process for determining the volume of a tumor within a body organ. The invention has particular application to cancer treatment programs and research.
With the now widespread use of CT scanners, it has become increasingly recognized that rapid and accurate organ and tumor volume determinations from Computed Tomography (CT) image data can be of great value in radiotherapy treatment planning. See for example, the following publications - Hobday P, Hodson NJ, Husband J, Parker RP, MacDonald JS, "Computed tomography applied to radiotherapy treatment planning: techniques and results" Radiology 1979; 133:477-82; and Van Dyk J, Battista JJ, Cunningham JR, Rider WD, Sontag MR, "On the impact of CT scanning on radiotherapy planning" Comput Tomogr 1980; 4:55-65. Tumor volume determinations are also important for radiation dose estimates for normal and tumor issues in radiolabeled antibody cancer therapy. See for example the following publications - Leichner PK, Klein JL, Garrison JB, et al "Dosimetry of .sup.131 I-labeled antiferritin in hepatoma: a model for radioimmunoglobulin dosimetry" Int J Radiat Oncol Biol Phys 1981;7:323-33; Leichner PK, Klein JL, Siegelman SS, Ettinger DS, Order SE "Dosimetry of 131I-labeled antiferritin in hepatoma: specific activities in the tumor and liver" Cancer Treat Rep 1983;67:647-58; and Leichner PK, Klein JL, Fishman EK, Siegelman SS, Ettinger DS, Order SE "Comparative tumor dose from .sup.131 I-labeled polyclonal anti-ferritin, anti-AFP, and anti-CEA in primary liver cancers" Cancer Drug Delivery 1984; 1:321-8. Such volume determinations are also important for the assessment of tumor response to new treatment modalities. See for example the following publication--Order SE, Klein JL, Leichner PK, et al, "Radiolabeled antibodies in the treatment of primary liver malignancies" In:Levin B, Riddell R, eds. Gastrointestinal cancer, New York: Elsevier-North Holland, 1984;222-32.
Volume computations from CT have been investigated by several authors using a variety of methods. For example, see the following publications--Heymsfield SB, Fulenwider T, Nordinger B, Barlow R, Sones P, Kutner M. "Accurate measurement of liver kidney, and spleen volume and mass by computerized axial tomography" Ann Intern Med 1979;90:185-7; Henderson JM, Heysfield SB, Horowitz J, Kutner MH, "Measurement of liver and spleen volume by computed tomography" Radiology 1981; 141:525-7; Moss AA, Cann CE, Friedman MA, Marcus FS, Resser KJ, Berninger W., "Volumetric CT analysis of hepatic tumors" J Comput Assist Tomogr 1981;5:714-8; Moss AA, Friedman MA, Brito AC, "Determination of liver, kidney, and spleen volumes by computed tomography: an experimental study in dogs" J Comput Assist Tomogr 1981; 5:12-4; Breiman RS, Beck JW, Korobkin M, et al, "Volume determinations using computed tomography. AJR 1982; 138:329-33; Oppenheimer DA, Young SW, Marmor JB, "Work in progress, serial evaluation of tumor volume using computed tomography and contrast kinetics" Radiology 1983; 147:495-7; Reid MH, "Organ and lesion volume measurements with computed tomography" J Comput Assist Tomogr 1983;7:268-73.
Moss et al [Moss AA, Cann CE, Friedman MA, Marcus FS, Resser KJ, Berninger W., "Volumetric CT analysis of hepatic tumors" J Comput Assist Tomogr 1981;5:714-8]have described a computer program for calculating the mean CT number of normal liver tissue in each CT "slice" and obtaining total liver volume by summing over all CT slices containing liver. Tumor volume in each slice was obtained by subtracting a Gaussian distribution of CT numbers for normal liver from the bimodal CT number distribution for the whole liver. The results from all slices were summed to obtain partial tumor and liver volumes for each patient.
It is known to determine tumor and liver volumes from sets of manually contoured CT slices. However, as practiced in the prior art, such determinations are time-consuming and labor-intensive. A radiologist must outline with a grease pencil regions of interest (ROI) corresponding to tumor and normal liver on patients' CT films. These contours are then digitized, the areas computed by numerical integration, and multiplied by slice thickness to obtain tumor and normal liver volumes for each slice. Total volume is obtained by summing over all slices. In spite of the fact that such methodology is extremely cumbersome, it was carried out for several years (1979-1984), and clinically relevant and important results were obtained and published in scholarly journals. The method of Moss et al also required manual contouring, directly on a video monitor, and slice-by-slice analysis of patients' CT scans.
To handle the increased number of patients due to expansion of the radiolabeled antibody treatment programs, it became evident that further automation was required to provide clinicians with timely information about tumor response to therapy and for radiolabeled antibody treatment planning.
SUMMARY OF THE INVENTION
The technique for determining the volume of a tumor within a body organ presented herein is more automated than prior art methods. The claimed method has several advantages over known methodologies. Regions of interest corresponding to tumor and normal liver are generated in a computer-assisted manner which does not require the presence of a trained radiologist. Secondly, the decision as to tumor and normal liver tissues within the regions of interest is based on a global histogram method which includes all CT slices in a patient's scan. This is both computationally faster and statistically more reliable than previous methods.
The present invention provides a more automated technique for determining from CT image data the volume of a tumor within a body organ. The method can be summarized in "outline" form as follows:
CT image data previously collected for a plurality of contiguous organ slices is read.
An image of the first slice is displayed on a monitor.
An operator inputs upper and lower limit CT numbers to define a boundary condition of the organ.
The operator visually identifies a "seed" pixel that is clearly within the organ.
The computer generates and displays on the monitor, based on defining data input by the operator, a region of interest (ROI) corresponding to an outline of the organ.
If the computer-generated ROI is unsatisfactory, the operator visually modifies the computer generated approximate organ outline to produce a modified ROI that more accurately identifies the organ outline.
Once an accurate organ boundary has been visually established, the computer determines which pixels are within the boundary and stores as a local histogram of the slice, information identifying the number of pixels and CT number associated therewith.
This process is repeated for each slice to produce a local histogram of each slice.
The local histograms are summed to produce a global histogram indicative of the number of pixels within the respective organ boundaries of all slices having each unique CT number.
A demarcation CT number is determined that distinguishes between normal organ tissue and tumor. This may be done by the operator looking at the global histogram.
Based on the demarcation CT number and the global histogram, organ volume and tumor volume are computed.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a general flow chart of the tumor volume determination method according to the present invention.
FIG. 2(i a) is a representative global histogram of a CT number distribution with three distinct peaks. Peaks 1 and 2 were associated with normal liver (NL) and tumor (T), respectively. The third peak was associated with necrotic tissue within the core of a hepatoma. The dashed lines were threshold CT numbers for normal liver and tumor.
FIG. 2(b) is the same histogram as in FIG. 2(a) with the three gaussian fitting functions superimposed on the histogram.
FIG. 3(a) is a representative global histogram with two distinct peaks.; Peak No. 1 was associated with normal liver (NL) and Peak No. 2 with tumor (T). The dashed line represents the threshold CT number for tumor and normal liver.
FIG. 3(b) is a global histogram (outside curve) for the same patient as in FIG. 3(a). The three gaussian functions were summed to get the global histogram. The gaussian function with the highest mean CT number corresponded to normal liver (NL), and the gaussian function with the largest amplitude and lower mean CT number to tumor (T).
FIG. 4 is a representative histogram with a single peak for a patient with a relatively small tumor (581 cm.sup.2).
FIG. 5 is a global histogram (outside curve) of the CT number distribution for normal liver and tumor for the same patient as in FIG. 4.
FIG. 6 is a representative histogram with single peak for a patient with a relatively large tumor (1687 cm.sup.3).
FIG. 7 is a global histogram (outside curve) for the same patient as in FIG. 6.
FIG. 8(a) is a representative CT slice of a patient with a solid hepatoma. The tumor (dark area) was highlighted according to the global histogram method.
FIG. 8(b) is the same CT slice as in FIG. 8(a), but manually contoured by an experienced observer.
FIG. 9 is a comparison of liver volumes obtained by the global histogram method (computer assisted) and from manually contoured CT slices for a sample of 10 patients.
FIG. 10 is a comparison of tumor volumes computed by the global histogram method and from manually contoured CT slices for the same patients as in FIG. 9.
FIG. 11(a) is a CT slice of a patient with diffuse hepatoma.
FIG. 11(b) is the same slice as in FIG. 11(a) but with tumor bearing regions highlighted according to the global histogram method.
FIG. 12 shows the relationship between mean CT numbers of normal liver tissue and the threshold values of CT numbers used in tumor volume computations for global histograms characterized by a single peak.
FIG. 13 shows the relationship between mean CT numbers for normal liver and threshold CT numbers used in tumor computations for histograms which were characterized by two and three distinct peaks.
FIG. 14 is a block diagram of a computer arrangement for carrying out the method set forth in the FIG. 1 flow chart.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT
FIG. 1 is a general flow chart of the tumor volume determination technique according to the invention.
A block diagram of a computer for carrying out the FIG. 1 method is shown in FIG. 14.
At step 100 CT image data, previously gathered by CT scanning the patient, is read and images of the CT "slices" are displayed. Usually CT image data collected by a CAT scan procedure is stored on a magnetic tape. Therefore, this step includes reading the CT image data from that magnetic tape, previously produced. The CT image data includes data for each of a plurality of contiguous slices of predetermined thickness of the tumor bearing organ. Preferably, the slices are 8 mm thick, but other slice thicknesses can be used. The CT image data includes a CT number representing tissue density for each of a plurality of pixels defining the slice.
Steps 100-128 are carried out for each slice (image). For a given slice an image of that slice is displayed. The operator is asked at step 102 whether the boundary for the previous slice is to be applied to the now displayed slice. Of course, if the operator is viewing the first slice of a series of slices, there is no prior boundary to use. If, at step 102, the operator answers "yes", program control proceeds to step 104 and the boundary of the previous slice is superimposed on the now displayed slice. This gives the operator a short cut starting point for identifying the organ's boundary in the slice now being displayed. In steps 106 and 108, the operator can set to zero all pixels for the now displayed slice that are outside the ROI carried over from the previous slice.
If the operator is viewing the first slice, or for some other reason does not want to start with the boundary from the previous slice, he answers "no" at step 102 and program control flows to step 110. At step 110, the operator is asked whether he wants to use the cursor. The cursor appears as a cross mark on the monitor and is under the control of a mouse. If the operator answers "yes", program control then proceeds to step 112. The operator can move the cursor, using the mouse, to any point on the image. The computer will provide the x and y position of the pixel and its CT number. The process at step 112 can help the operator determine lower and upper thresholds of CT numbers that he will have to specify at step 120.
If the operator has decided at step 110 not to use the cursor, he is asked to input lower and upper thresholds of CT numbers. These thresholds are used by the computer to look for boundary pixels. These are pixels at the edge of the tumor bearing organ. Using a "seed" pixel visually selected by the operator and the thresholds input at step 120, the computer automatically generates what it thinks is the boundary of the organ. To input the seed pixel, the operator moves the cursor to an arbitrary pixel that is clearly within the boundary of the organ, as it appears on the monitor. Beginning at the seed pixel, the computer searches in accordance with a predetermined routine from one pixel to the next until it locates a pixel meeting the threshold criteria input at step 120. The first pixel so located is considered to be the first boundary pixel. The nearest neighboring pixels to the first boundary pixel are examined to determine if any of them meet the threshold condition. Each such pixel located is also considered to be a boundary pixel. A vector is "drawn" from the first boundary pixel to the second boundary pixel, etc. This process continues from one pixel to the next until the entire boundary line of the organ has be drawn by the computer.
From step 122, program control proceeds to step 124 where the operator is asked whether he wants to calculate the area inside the now displayed boundary. If the operator is confident that the boundary now shown on the monitor is accurate, no adjustment is necessary and the operator can answer "yes" and proceed to step 126. However, if the boundary is not accurate, the operator must answer "no" and make a boundary correction. If the operator answers "no", program control then proceeds to step 110. The operator can use the cursor to adjust the boundary to be more accurate. This allows the operator to correct the boundary line for errors caused by tissues abutting the organ that have close CT numbers to the organ tissue. For example, in the case of a liver scan, the boundary line automatically generated by the computer often includes soft tissue between ribs. This boundary correction is carried out at steps 114-118. At step 114 where the operator is asked if he wants to blank a special area. If he answers "yes" program control then proceeds to step 116 where the operator can adjust the boundary generated by the computer. He does this by moving the cursor so as to identify two separate and distinct points on the true boundary (not the one generated by the computer). The computer than adjusts the boundary to include those points, thereby eliminating from within the boundary points that do not belong.
Only after the operator is satisfied with the boundary appearing on the monitor does he answer "yes" at step 124. Program control then moves to step 126. At step 126, the computer displays the number of pixels inside the boundary. At step 128, data is stored in the form of a local histogram for the slice then being displayed. The local histogram includes data indicative of the number of pixels having each particular CT number.
Steps 100-128 are repeated for each slice so that a local histogram of each slice is produced. The local histograms are summed to produce a global histogram indicative of the number of pixels within the organ boundaries of respective slices having each unique CT number. Then, there is determined from the global histogram a demarcation CT number that distinguishes between normal organ tissue and tumor. Depending upon the histogram, this is done in various ways. From the demarcation CT number and the global histogram, organ volume and tumor volume are computed.
The manner of determining the demarcation CT number and the determination of tumor and organ volumes will be further described in the following material related to a study. In part, the manner for determining automatically tumor volume from the global histogram depended upon determining some constants to be used in calculation empirically.
In a study of 56 patients with primary liver cancers it became evident that the distribution of CT numbers was not necessarily bimodal, and that a slice-by-slice determination of mean CT numbers did not provide sufficient statistical information to distinguish between normal and tumor bearing liver tissues. An algorithm was, therefore, developed to generate histograms of CT number distributions for all of the CT slices in patients' liver scans (global histograms) without having to determine the mean CT numbers corresponding to normal liver in individual slices. This provided more reliable statistical information and had the advantage of being computationally faster.
All patients in this study had histologically confirmed primary liver cancers, and CT examinations were performed using a Siemens Somatom DR3 body scanner. Livers were scanned at contiguous 8-mm intervals with an 8-mm slice thickness while patients suspended respiration at resting lung volume. Reconstructed transaxial slices were stored on magnetic tape in 256.times.256 matrices and analyzed on a minicomputer. Semiautomatic computer software, as described above, was used to define a region of interest (ROI) corresponding to the boundary of the whole liver in each slice, and a local histogram of the CT numbers within the ROI was generated. A global histogram was then obtained by summing over the local histograms for each slice. Total liver volume was computed from the number of pixels in the global histogram and the known pixel size and slice thickness of 8 mm.
Tumor volume computations were based on an analysis of global histograms. The global histograms fell into three categories. Some of the global histograms had three distinct peaks, some had two distinct peaks, and some had only a single peak. Quantitative information about liver and tumor volumes was extracted from these CT number distributions in a consistent manner by fitting them to a sum of three gaussian functions given by ##EQU1##
In the above equation, n represented the CT numbers and ni their mean value for each of the gaussian functions; A.sub.i and .sigma..sub.i were the corresponding amplitude and variance, respectively.
A representative CT number distribution with three distinct peaks is shown in FIG. 2(a), and the three gaussian fitting functions superimposed on the histogram are displayed in FIG. 2(b). The gaussian function with the largest amplitude and the highest mean CT number (peak no. 1) corresponded to normal liver (NL) and the gaussian function with the second-largest amplitude and lower mean CT number (peak no. 2) to tumor (T). By highlighting pixels in CRT displays of CT slices, it was determined that the third peak in FIGS. 2(a) and 2(b) was representative of tissue within the core of solid hepatomas. The mean value of CT numbers in the third peak was even lower than that of the tumor itself and indicated the presence of necrotic tissue.
The gaussian fitting functions also made it possible to define threshold CT numbers for tumor and liver tissues and to determine the probability for normal pixels to be representative of these tissues. For example, the dashed lines in FIG. 2(a) were located at CT numbers corresponding to the minima between gaussian functions 1 and 2 and gaussian functions 2 and 3 in FIG. 2(b), respectively. All pixels to the right of the dashed line between peaks 1 and 2 in FIG. 2(a) were interpreted as normal liver, and pixels falling between the two dashed lines as tumor. This interpretation was verified by highlighting these pixels in CT slices on CRT displays. From the gaussian fitting functions it followed that the probability for pixels within these two ranges of CT numbers to be representative of normal liver and tumor tissues was close to unity (>0.999). All pixels corresponding to tumor were summed in the global histograms, and tumor volumes were computed in the same manner as whole liver volumes.
A representative global histogram with two distinct peaks is shown in FIG. 3(a). FIG. 3(b) shows the three gaussian functions and the histogram fit resulting from the addition of these functions. The third gaussian function (dotted line) with the lowest amplitude was required to obtain a satisfactory fit to the asymmetric portion of the histogram in the range of the lowest CT numbers. Additionally, for CT numbers from 30 to 46 the third gaussian corresponded to low-density structures such as blood vessels, fatty tissue, and bile ducts. This was ascertained by highlighting pixels in this range of CT numbers. In the case of global histograms with only one or two peaks, the third gaussian did not reflect necrotic tissue because it extended over a range of CT numbers that was higher than that of the third gaussian in FIG. 2(b). As before, the gaussian function with the highest mean value of CT numbers was interpreted as representing normal liver (peak no. 1), and the gaussian fitting function for peak no. 2 as representing tumor. Threshold CT numbers for normal liver and tumor in these histograms corresponded to the minimum value of the gaussian fitting functions in the overlap region between peaks 1 and 2 in FIG. 3(b). The dashed line in FIG. 3(a) was obtained in this manner. All pixels with CT numbers above this threshold were counted as normal liver and all others as tumor. Pixels in these two ranges of CT numbers were highlighted in different colors on CRT displays of CT slices and identification of normal liver and tumor determined to be satisfactory by experienced observers. An analysis of the gaussian fitting functions showed that the probability for pixels to represent normal liver and tumor in each of the two CT number ranges was 0.960 and 0.942, respectively.
A representative CT number distribution with a single peak for a patient with a small tumor is shown in FIG. 4. FIG. 5 shows the three gaussian functions and the histogram fit resulting from the addition of these functions. The three gaussian functions were summed to get the global histogram for this patient. The gaussian function with the largest amplitude and the highest mean CT number was representative of normal liver. The position of the arrow corresponded to one-fourth of the maximum of the dominant gaussian function (NL), and all pixels with CT numbers lower than this were empirically determined to be attributable to tumor For single-peak histograms and small tumors, the gaussian with the largest amplitude and highest mean CT number was very reproducible and represented normal liver. This gaussian had a standard deviation ranging from 2-6 Hounsfield Units (HU, 1000 scale). The gaussian with the second-largest amplitude and lower mean CT number had a standard deviation ranging from 3-10 HU. The lowest-amplitude gaussian had a more variable standard deviation and included pixels with low CT numbers and sometimes also those corresponding to tumor and normal liver.
Although in principle tumor volumes could be computed by summing pixels under the gaussian function with the second largest amplitude and lower mean CT number, in practice this was not feasible for the following reasons. The fitting parameters for the two lower-amplitude Gaussians could be varied considerably without significant changes in the goodness of the fit. The fit for these two functions was, therefore, not nearly as reproducible as for the dominant gaussian (NL) in FIG. 5. Secondly, highlighting of pixels in CT slices demonstrated that the gaussian with the second-largest amplitude included a large number of normal liver pixels as evidenced by the overlap of the fitting functions in FIG. 5.
An empirical method was, therefore, developed to compute tumor volumes based on the gaussian fitting function with the largest amplitude. This method was based on a quantitative comparison of tumor volume computations from global histograms and manually contoured CT slices. Volume determinations from manually contoured CT slices were carried out. Additionally, pixels corresponding to tumor were highlighted on CRT displays of CT slices and reviewed by experienced observers. For relatively small tumors, reproducible and satisfactory results were obtained by determining the CT number corresponding to one-fourth of the amplitude to the left of the mean of the gaussian function for normal liver of the (gaussian with the largest with the largest amplitude and the highest mean CT number), as indicated by the arrow in FIG. 5. All pixels to the left of the arrow, in the direction of lower CT numbers, were counted as tumor.
This procedure was justified because pixels with CT numbers above the threshold determined by left hand one-fourth of the amplitude of the dominant (NL) Gaussian had a high probability of representing normal liver tissue. For this threshold, Gaussian error analysis indicated that 95.2% of the normal liver Gaussian was above the threshold and occupied 1,395.63 area units. Using the tumor Gaussian in FIG. 5, 25.4% was above threshold and occupied 174.75 area units. Therefore, the probability of a normal liver pixel being above the threshold was 0.889 whereas the probability of a tumor pixel being above the threshold was 0.111.
A global histogram for a patient with a large tumor is shown in FIG. 6. The three gaussians and the resulting histogram fit are displayed separately in FIG. 7. The gaussian with the largest amplitude and lower mean CT number was representative of the tumor, and the gaussian with the second largest amplitude and the highest mean CT number of normal liver. As before, the gaussian function with the lowest amplitude and the lowest mean CT number extended from low CT numbers into the range corresponding to tumor and normal liver. For these large tumors, good results were obtained by determining the CT number corresponding to three-fourths of the amplitude to the right of the mean of the gaussian function for tumor, as indicated by the arrow in FIG. 7. Tumor volumes were computed by summing over all pixels to the left of the arrow and multiplying by slice thickness.
The probability that a pixel below the threshold determined by three-fourths of the amplitude of the tumor Gaussian represented tumor tissue was estimated in the same manner as above. Namely, 77.6% of the tumor Gaussian was below the threshold and occupied 1,371.19 area units; 14.2% of the normal liver Gaussian in FIG. 7 was below the threshold and occupied 194.11 area units. In addition to that, the third Gaussian contributed 108 area units. Therefore, the probability of a tumor pixel, a normal liver pixel, or a non-tumorous low-density pixel being below the threshold was 0.820, 0.116, and 0.064, respectively.
Using a smaller fraction of the amplitude of the tumor Gaussian would have increased the probability of a tumor pixel below the threshold. However, this was equivalent to raising the CT number threshold for tumor pixels and had the undesirable effect of including normal liver in tumor volumes. This was ascertained by highlighting pixels in CT slices over a range of CT numbers of the tumor Gaussian.
Results of Study
Tumor and liver volumes of 51 patients with hepatoma and 5 patients with cholangiocarcinoma were computed from CT scans. The procedures for volume determinations were the same for these two groups of patients. For the first consecutive 10 patients, volumes computed from manually contoured CT slices were compared with volumes obtained by the global histogram method.
FIG. 8(a) shows a CT slice of a hepatoma patient with a solid tumor and tumor pixels highlighted according to the global histogram for this patient. For comparison, the same slice is shown in FIG. 8(b) with the liver and tumor-bearing region defined manually by an experienced observer directly on the CT film. Results obtained by these two methods are shown in FIGS. 9 and 10 for liver and tumor volumes, respectively.
FIG. 9 is a comparison of liver volumes obtained by the global histogram method (computer assisted) and from manually contoured CT slices for a sample of 10 patients. Two of the volumes were nearly identical, as indicated by the number 2. The solid line was generated from a least-squares fit with a correlation coefficient of 0.992. The dashed line is the line of identity.
FIG. 10 is a comparison of tumor volumes computed by the global histogram method and from manually contoured CT slices for the same patients as in FIG. 9. The correlation coefficient was 0.995 (solid line). The dashed line is the line of identity. For both liver and tumor volumes, results were clustered about the line of identity with correlation coefficients of 0.992 (liver) and 0.995 (tumor).
For the remaining 46 patients, normal and tumor pixels in CT slices were highlighted and displayed on a color monitor. A representative CT slice of patient with diffuse hepatoma that would have been difficult to contour manually is shown in FIG. 11(a) The same slice highlighted according to the global histogram method is shown in FIG. 11(b). Normal liver and tumor ROI's in CT slices for this and all other patients were reviewed by experienced observers and determined to be satisfactory.
Global histogram structures of the CT numbers for 51 patients with hepatoma and 5 patients with cholangiocarcinoma are summarized in Table 1.
Thirty-seven (66%) of these patients had histograms that were characterized by a single peak (FIGS. 4 and 6), 17 (30%) had double-peak, and 2 (4%) had triple-peak histograms (FIGS. 2 and 3).
FIG. 12 shows the relationship between mean CT numbers of normal liver tissue and the threshold values of CT numbers used in tumor volume computations for global histograms characterized by a single peak. An example is indicated by the dashed lines. The mean CT number for normal liver is 55 HU and the corresponding threshold CT number for tumor computations is 47 HU. Also shown is the least-squares fit line.
FIG. 12 demonstrates the relationship between mean CT numbers of normal liver tissue and the threshold values of CT numbers used in tumor volume computations, based on global histograms characterized by a single based on global peak. For this group of patients, mean CT numbers for normal liver ranged from 44-73 HU, whereas threshold CT numbers for tumors ranged from 37-65 HU. An illustration of this relationship is indicated by the dashed lines in FIG. 12. In this example, the mean CT number for normal liver is 55 HU, and the corresponding threshold CT number for tumor volume computations is 47 HU. All pixels with CT numbers less than or equal to 47 HU would be counted as tumor.
FIG. 13 shows the relationship between mean CT numbers for normal liver and threshold CT numbers used in tumor computations for histograms which were characterized by two and three distinct peaks. Also shown is the least-squares fit line.
For these CT number distributions, mean CT numbers for normal liver ranged from 52-72 HU, and tumor threshold value from 40-60 HU. These data showed that for all types of global histograms encountered, there was a nearly linear relationship between mean CT numbers for normal liver and threshold CT numbers for tumor volume computations.
The number of CT examinations in FIGS. 12 and 13 included patients who were scanned prior to and following therapy. There were, however, no systematic changes in either the mean CT number for normal liver or the threshold CT number for tumor volume computations following therapy.
An example of the clinical application of liver and hepatoma volume calculations is provided by the data in Table 2.
Liver and tumor volume computations were made prior to and following three administrations of I-131 labeled antiferritin IgG. After the third treatment, the tumor volume had been reduced from 2480 cm.sup.3 to 298 cm.sup.3, and the patient underwent surgery for removal of the residual tumor.
To assess tumor response to a given therapy, accurate and timely tumor volume computations are required. Such computations are essential in planning treatment strategies and making appropriate changes in therapy, if required. Computer assisted volume computations have proven to be clinically important for these reasons and because of their reproducibility. Repeat volume computations by different observers were consistent to within 2-4%. Additionally, the semi-automatic ROI software was faster than manual outlining using a track ball, light pen, cursor, etc., to define boundaries. This is of considerable practical importance when analyzing large numbers of CT slices in ongoing clinical trials.
The technique according to the present invention is independent of the mean CT number for normal liver in individual slices, it is readily applicable to tumors which have invaded normal liver diffusely and for which single-slice statistics are less reliable. The global histogram approach also makes it possible to define a relationship for the mean CT number associated with normal liver and threshold CT numbers for volume determinations of hepatic tumors.
The use of computer-assisted volume determination has its limitations. With a slice thickness of 8 mm, partial volume effects introduce a small error in volume computations. We estimated this effect by summing over the first and last slices and dividing by two. The variation due this procedure ranged from 0.36-1.9% of total liver volume. As detailed in Moss AA, Cann CE, Friedman MA, Marcus FS, Resser KJ, Berninger W. Volumetric CT analysis of hepatic tumors. J Comput Assist Tomogr 1981;5:714-8 and Ettinger DS, Leichner PK, Siegelman SS, et al. Computed tomography assisted volumetric analysis of primary liver tumor as a measure of response to therapy. Am J Clin Oncol 1985;8:413-8, non-tumorous structures of low density such as bile ducts, portal and hepatic vessels, and fatty tissues are included in tumor volumes. An estimate of this error was obtained by summing over the third gaussian distribution separately. It was determined that for most patients, this gaussian contributed 0.50-5.89% to the liver volumes. In 2.5% of the patients, the CT number distributions were irregular and could not be fit to three gaussians. In spite of these limitations, a comparison of computed liver and tumor volumes and autopsy data of 4 patients with primary hepatic liver cancers showed a maximum difference of +/-6.4% {See Ettinger DS, Leichner PK, Siegelman SS, et al. Computed tomography assisted volumetric analysis of primary liver tumor as a measure of response to therapy. Am J Clin Oncol 1985;8:413-8}. This was consistent with the estimated errors in the present work.
Other embodiments of the invention should occur to those of ordinary skill in the art having the benefit of the teachings presented herein. Therefore such alternative embodiments are intended to be within the scope of the claims herein.
Claims
I claim:
1. A computer-aided method for determining the volume of a tumor within a body organ from CT image data of that organ, comprising the steps of: (a) obtaining CT image data for each of a plurality of slices of predetermined thickness of the body organ, the image data including a plurality of pixels and a CT number corresponding to each pixel; (b) transferring the CT image data into a computer; (c) displaying from the data in the computer an image of one of the slices; (d) computer generating an approximate boundary of the organ and displaying the boundary superimposed on the image of the slice; (e) operator interacting with the computer to modify the boundary generated by the computer to more accurately describe the boundary of the organ; (f) identifying, by the computer, the particular pixels within the organ boundary and the CT number associated with each such pixel and generating therefrom a local histogram of the slice, the histogram including data indicative of the number of pixels within the boundary having a particular CT number; (g) from the predetermined thickness and number of pixels identified within the boundary, determining by the computer, the volume of organ within the slice; (h) repeating steps (c) through (g) for each slice to so as to determine a local histogram corresponding to each slice; (i) summing the local histograms corresponding to the various slices to obtain a global histogram indicative of the number of pixels within the orga boundaries of respective slices having each particular CT number; (j) determining from the global histogram a demarcation CT number that distinguishes between normal organ tissue and tumor; (k) determining from the volume computations for each of the slices, the total volume of the organ; and (l) computing from the global histogram and demarcation CT number, the volume of the tumor.
2. A method according to claim 1 wherein step (d) comprises the steps of: (1) operator specifying a CT number threshold criteria for defining the boundary of the organ; (2) operator specifying a seed pixel by visually ascertaining an arbitrary point within the organ; (3) beginning at the seed pixel, searching from one pixel to the next to locate a first boundary pixel meeting the threshold criteria specified; (4) examining nearest neighboring pixels to the first boundary pixel to determine if they meet the threshold criteria to find a second boundary pixel, (5) drawing a vector from the first boundary pixel to the second boundary pixel, (6) examining pixels neighboring to said second boundary pixel to determine third, fourth, . . . nth boundary pixels and drawing vectors from each last found pixel to a newly found pixel, to generate a series of vectors defining the boundary of the organ.
3. A method according to claim 1 wherein step (d) comprises the step of reading data defining the boundary determined for a previously boundary defined slice and utilizing that previously determined boundary for the current slice.
4. A method according to claim 1 wherein step (j) comprises the step of the operator visually examining the global histogram and selecting the demarcation CT number based on predetermined criteria.
5. A method according to claim 1 wherein step (j) comprises the step of the computer fitting the global histogram to a sum of three gaussian functions given by ##EQU2## wherein n represents the CT numbers and ni their mean value for each of the Gaussian functions, A.sub.i and .sigma..sub.i.sup.2 are the corresponding amplitude, and variance, respectively.
6. A computer-based arrangement for determining the volume of a tumor within a body organ from CT image data of that organ, comprising: (a) means for obtaining CT image data for each of a plurality of slices of predetermined thickness of the body organ, the image data including a plurality of pixels and a CT number corresponding to each pixel; (b) means for reading the CT image data; (c) means for displaying one at a time, an image of each slice; (d) means for generating, for each slice, an approximate boundary of the organ and displaying the boundary superimposed on the image of the slice; (e) means for operator interacting with said arrangement to modify the boundary generated and displayed for each slice to ore accurately describe the boundary of the organ; (f) means of identifying, for each slice, the particular pixels within the organ boundary and the CT number associated with each such pixel and generating therefrom a local histogram of the slice, the histogram including data indicative of the number of pixels within the boundary having a particular CT number; (g) means for determining, for each slice, from the predetermined thickness and number of pixels identified within the boundary, the volume of organ within the slice; (h) means for summing the local histograms corresponding to the various slices to obtain a global histogram indicative of the number of pixels within the organ boundaries of respective slices having each particular CT number; (i) means for determining from the global histogram a demarcation CT number that distinguishes between normal organ tissue and tumor; (j) means for determining from the volume computations for each of the slices, the total volume of the organ; and (k) means for computing from the global histogram and demarcation CT number, the volume of the tumor.
7. An arrangement according to claim 6 wherein said (d) means for generating comprises: (1) means for operator specifying a CT number threshold criteria for defining the boundary of the organ; (2) means for operator specifying a seed pixel by visually ascertaining an arbitrary point within the organ; (3) means for searching from one pixel to the next, beginning at the seed pixel, to locate a first boundary pixel meeting the threshold criteria specified; (4) means for examining nearest neighboring pixels to the first boundary pixel to determine if they meet the threshold criteria to find a second boundary pixel, (5) means for drawing a vector from the first boundary pixel to the second boundary pixel, (6) means for examining pixels neighboring to said second boundary pixel to determine third, fourth, . . . nth boundary pixels and drawing vectors from each last found pixel to a newly found pixel, to generate a series of vectors defining the boundary of the organ.
8. An arrangement according to claim 6 wherein said generating means comprises means for reading data defining the boundary determined for a previously boundary defined slice and utilizing that previously determined boundary for the current slice.
9. An arrangement according to claim 6 wherein said (j) means for determining comprises means for the operator to visually examine the global histogram and select the demarcation CT number based on predetermined criteria.
10. An arrangement according to claim 6 wherein said (j) means for determining comprises means for computer fitting the local histogram to a sum of three gaussian functions given by ##EQU3## wherein n represents the CT numbers and n.sub.i their mean value for each of the Gaussian functions, A.sub.i and .sigma..sub.i.sup.2 are the corresponding amplitude, and variance, respectively.
Patent Citations (6)
Non-Patent Literature (11)
- Cassell et al., "Automatic Outlining Technique for EMI Scanner Pictures", Med. & Biol. Eng. & Comput., 1979, vol. 17, No. 5, pp. 693-694.
- Cancer Drug Delivery, vol. 1, No. 4, 1984--Comparative Tumor Dose from 131 I-Labeled Polyclonal Anti-Ferritin . . . by Peter K. Leichner et al.
- Radiation Oncology Biology Physics, Mar. 1981, vol. 7, No. 3, Dosimetry of 131-I-Labeled Anti-Ferritin in Hepatoma: A Model for Radioimmunoglobulin--Peter K. Leichner et al.
- Cancer Treatment Reports--vol. 67, Nos. 7-8, Jul./Aug. 1983--Dosimetry of 131 I-Labeled Anti-Ferritin in Hepatoma: Specific Activities in the Tumor and Liver--Peter K. Leichner et al.
- Radiology, May 1983--pp. 495-497--Work in Progress: Serial Evaluation of Tumor Volume Using Computed Tomography and Contrast Kinetics.sup.1, David A. Oppenheimer et al.
- Computerized Tomography, vol. 4, pp. 55 to 65--On the Impact of CT Scanning on Radiotherapy Planning--J. Van Dyk et al.
- Radiology 141:525-527, Nov. 1981--Measurement of Liver and Spleen Volume by Computed Tomography--J. Michael Henderson.
- J. Comput. Assist Tomogr., vol. 5, No. 1, 1981--Determination of Liver, Kidney, and Spleen Volumes by Computed Tomography-Moss et al.
- J. Comput. Assist Tomogr., vol. 5, No. 5, 1981--Volumetric CT Analysis of Hepatic Tumors--Moss et al.
- Feb. 1979--Annals of Internal Medicine, vol. 90, No. 2--Accurate Measurement of Liver, Kidney and Spleen Volume and Mass by Computerized Axial Tomography--Heymsfield et al.
- AJR:138, Feb. 1982--pp. 329-333--Volume Determinations Using Computed Tomography--Richard S. Breiman et al.