Volumetry of hepatic metastases in computed tomography using the watershed and active contour algorithms

Abstract

The liver is a common site of metastatic disease. Colorectal liver metastases (CLM) alone is diagnosed in approximately 50,000 patients each year in the United States. Treatment for liver metastases is monitored according to the size and number of the hepatic metastases visualized with contrast-enhanced computed tomography (CT). In routine clinical practice, lesion size is assessed according to uni-or bi-dimensional criteria. However, measurements made using these criteria are subject to inter-observer variability and to the inherent limitations of dimensional measurements to reflect volume. Thus, an investigation was conducted of computational methods for measurement of the area and volume of colorectal metastases to the liver. The study included CT images of the liver from five patients with a history of colorectal cancer. The image data included a total of ten hypo-attenuating lesions ranging in size from 0.4 to 79 cc. The lesions were segmented by manual contouring, by the 2D Watershed algorithm, by an active contour initialized by manual contouring and by a dual-scale active contour initialized with an ellipse. Each method is interactive and was applied two times to each image of each lesion (36 image slices in total) to determine the reproducibility of area and volume measurements. The Watershed algorithm was ineffective since it produced gross segmentation errors on a majority of the lesion slices while the active contour with initialization by manual contouring produced no significant improvement in reproducibility. The average difference between the first and second measurements of lesion areas by the multiresolution active contour was slightly less than for manual measurements (0.53 0.83 and 0.58 0.60 cm 2 ) but the difference was not statistically significant. However, area measurements by the dual-scale active contour had less variability in a greater number of slices than did manual contouring (p= 0.00003). Greater variability in area measurements occurred on the end slices than interior slices for each lesion (r=0.948 between first and second measurements for end slices, r=0.989 for interior slices). We conclude that a dual-scale active contour may be an effective alternative to manual contouring for measurement of volumes of liver metastases. Reduced slice thickness may reduce variability in volume measurement.

Liver tumour segmentation using contrast-enhanced multi-detector CT data: performance benchmarking of three semiautomated methods

“…on an organic human liver phantom [8]. CT-based volumetry studies used watershed, active contour and fuzzy c-means algorithms [9,10] and a shape-constraint region growing algorithm to automatically delineate liver metastases on CT images [12]. Seo KS [11] proposed a multi-stage, automatic hepatic tumour segmentation scheme that included liver structure segmentation, removal of hepatic vessels and hepatic tumour extraction by using composite hypotheses and minimal total probability error.…”

“…Hence automatic, accurate and robust methods for liver tumour segmentation and volumetry are increasingly receiving attention and research efforts from medical imaging, computer vision and pattern recognition communities. Some semiautomated methods, including region growing, isocontour, active contour, watershed, fuzzy cmeans, etc., have been developed and evaluated for liver tumour segmentation and volumetry [8][9][10][11][12][13][14]. Most of these studies suggest that computerised results are comparable to or highly correlated with manual measurement.…”

3D general lesion segmentation in CT

“…There has been very little interest in liver lesion segmentation [4,5]. The work has focused on relatively simple hypodense lesions which are nicely contrasted against the parenchyma.…”

Liver segmentation for hepatic lesions detection and characterisation

“…Recently, it has been demonstrated that a bias exists between the clinical practice of inferred measurements of hepatic lesions based on onedimensional or two-dimensional criteria and volume measurement analysis [3]. The results of semi-automatic techniques, using watershed and active contours, guarantee the robustness and reliability of these methods compared to manual segmentation ( [4], [5]). …”