Comparison and assessment of semi-automatic image segmentation in computed tomography scans for image-guided kidney surgery
Example images with hand segmentations from two image sets. The left image shows multiple pathologies on an image without contrast agent. Note some pathology may be of roughly equal intensity when compared to kidney parenchyma, while some may be darker. The right image shows an image taken in an early contrast phase (most contrast is still in the kidney cortex). Some parenchyma takes up this contrast more quickly than others, resulting in variable intensity. This tumor does not take up contrast well compared to normal parenchyma.
Semi-automatic algorithm corrected by parameter change. Left: Threshold level set method overestimates the segmentation due to easily followed vasculature on CT. Right: The same method’s results using lower propagation scaling, automatically implemented upon detection of the preceding failure.
Atlas slices from both a polar and central region of the kidney. Each atlas slice represents multiple aligned, overlapping hand segmentations. The value of a slice was normalized, yielding a way to estimate how likely a pixel was to represent kidney tissue for a given axial depth.
Semi-automatic algorithm corrected by method change. Left: Threshold level set method using low propagation scaling cannot contain an oversize segmentation caused by vessel following. Middle: Curvature detection method’s optimal results, showing only slight improvement in avoiding the vessel bleed. Right: Geodesic active contour method results, demonstrating restriction of the vessel overestimation.
Algorithm Overview. Once a slice is classified as middle or end (axially extreme slices), the order of methods is set. Each method has its own particular parameters to employ based on axial classification, contrast state, whether it has previously failed to segment the current slice, and whether that failure was too large or too small.
Preoperative CT and intraoperative laser range scans for three cases, (a)–(c). Registered laser range scan shown with autosegmented (left) and hand segmented (right) CT surfaces. Laser surface intensity indicates closest point distances to the CT surface, with the highest distances highlighted in white. Tumor is visible in case (a) at top right, with the white highlight located over the tumor itself. Tumors are visible for case (b) at both top left and bottom right. Case (c) contains a very endophytic tumor.
Visualization of distance grouping categories used in evaluation of the transformation variability introduced by using semi-automatic segmentation versus hand segmentation. Points were classified into five categories by increasing distance from target tumor centroid.
Training kidney results. Averages are shown for each contrast phase as well as for all kidneys together.
Testing kidney results. Averages are shown for each contrast phase as well as for all kidneys together.
Closest point differences between registered laser range scan data and segmented CT surfaces (all distances in mm).
Average distances between resultant physical space points when both hand and semi-automatic segmentation inverse transformations are applied to image space coordinates, divided by category. Each category represents a distance grouping from the tumor’s centroid, with category (1) being the closest.
Article metrics loading...
Full text loading...