Totally automatic definition of renal regions of interest from 99mTc-MAG3 renograms: validation in patients with normal kidneys and in patients with suspected renal obstruction

Abstract

Introduction: An image-processing algorithm (AUTOROI) has been developed to totally automatically (or with manual assistance) detect whole-kidney contours and generate renal regions of interest (ROI) for the extraction of the quantitative measurements used in the interpretation of Tc-mercaptoacetyltriglycine (Tc-MAG3) renograms.

Methods: The 18-20th min dynamic frames post-MAG3 injection were used to automatically define boxes surrounding each kidney, which were then transposed to an early composite image for interpolative and directional background subtraction. Sobel operator and unsharp masking were applied for edge enhancement, and the resulting image histograms were equalized to better define poorly functioning kidneys. AUTOROI searched radially from the center of mass to define each kidney's ROI coordinates. AUTOROI was validated using MAG3 studies from 79 patients referred for suspected obstruction (79 left, 77 right kidneys) and 19 kidney donors with normal kidney function and no obstruction. Renal ROIs were manually defined by a nuclear medicine technologist with 20+ years of experience (reference standard) and an American Board of Nuclear Medicine certified physician. AUTOROI and physician ROIs were automatically compared with the reference standard to determine the border definition error.

Results: AUTOROI totally automatically detected the renal borders in 89% (172 of 194) of the kidneys from the entire group of 98 patients. The 22 kidneys missed automatically were subsequently detected with the assistance of a single manually placed fiducial point demarcating the liver/kidney boundary. These 22 kidneys were shown to be associated with markedly reduced MAG3 clearance. The mean error of AUTOROI for all 194 kidneys was 6.66+/-3.77 and 7.31+/-4.52 mm for the left and right kidney, respectively. The physician's error was 6.78+/-2.42 and 6.65+/-2.05 mm for the left and right kidney, respectively. This error difference between AUTOROI and the physician was not statistically significant.

Conclusion: AUTOROI provides an objective and promising approach to automated renal ROI detection.

Figure 1

Automatic kidney ROI definition from a MAG3 study in a normal patient. Steps 1 through 5 refer to the image processing steps described in the methods section. See kidney borders compared to reference standard and to physician border definition in Figure 4.

Figure 2

Automatic kidney ROI definition from a MAG3 study in a patient with an obstructed left kidney. Steps 1 through 5 refer to the image processing steps described in the methods section. See kidney borders compared to reference standard and to physician border definition in Figure 5.

Figure 3

Error measurement between the borders automatically determined by AUTOROI and those manually assigned by the expert that served as the reference standard for each of 40 angular samples. The border defined by the reference standard is depicted in this figure as a dotted ellipse just for ease of display and explanation purposes.

Figure 4

Automatic kidney ROI definition from a MAG3 study in a normal patient (from Figure 1) compared to the human expert used as a reference standard and to physician's border definition. For this study, the mean radial error for AUTOROI compared to the reference standard in this patient was 4.3 mm and 4.7 mm for the left and right kidney respectively. The mean radial error for the physician compared to the reference standard in this patient was 5.7 mm and 5.1 mm for the left and right kidney respectively.

Figure 5

Automatic kidney ROI definition from a MAG3 study in a patient with obstructed left kidney compared to the human expert used as a reference standard and to physician's border definition. For this study, the mean radial error for AUTOROI compared to the reference standard in this patient was 3.8 mm and 6.7 mm for the left and right kidney respectively. The mean radial error for the physician compared to the reference standard in this patient was 6.2 mm and 6.5 mm for the left and right kidney respectively.

Figure 6

Comparison of Totally Automatic to Fiducial Point (FP) manually assisted kidney ROI definition from a MAG3 study in a patient with an obstructed right kidney in close proximity and uptake to the liver. Steps 2 through 5 refer to the image processing steps described in the methods section. The top row is an example of a study where the totally automatic algorithm missed detecting part of the right, upper cortex due to the loss of border definition as indicated by the dotted arrow on Step 4. Note also on the bottom row how the border definition of the right kidney is improved by the use of a single manually placed FP. This patient's right kidney exhibited a markedly low MAG3 clearance of 58 mL/min/1.73m².