Implanted fiducial markers, which are used to correct for day-to-day variations, may potentially also be used to correct for intrafraction motion measurements. However, before any treatment can make use of, and react to, the position of the inserted markers they have to be segmented, either manually through expert user intervention or automatically from an imaging system. In the current study, we aimed to establish a robust and autonomous segmentation method for implanted cylindrical gold markers in a single set of projections from a cone-beam computed tomography(CBCT).Methods:
Multiple cylindrical gold markers were segmented in the projection images of CBCT scans by five sequential steps. Initially, marker candidates were identified in all projections with a blob detection routine, and then traced in subsequent projections. Traces inconsistent with a 3D marker position were rejected, and the best remaining traces were identified and used for the construction of a 3D marker constellation model, consisting of the size, position and orientation of the markers. Finally, projections of the model were used to generate templates for the final template-based marker segmentation. Hereby, challenging situations such as overlap of markers and low contrast regions were taken into account. The segmentation method was tested in 63 CBCT scans from 11 patients with 2–4 cylindrical gold markers implanted in the prostate and for 62 CBCT scans from six patients each with 2–3 cylindrical gold markers implanted in the liver and up to two cylindrical markers placed externally on the abdomen. After segmentation all projections of the 125 scans were manually inspected, and a successful segmentation was registered if the segmented position was within the projection of the marker.Results:
For prostate markers, the segmentation was successful in 99.8% of the projections. For the liver patients, liver markers and external markers were segmented successfully in 99.9 and 99.8% of the projections, respectively. All markers were identified in the 3D marker constellation model. The most common source of segmentation error was low contrast and motion of markers relative to each other, which resulted in a discrepancy between the template and actual projection appearance during marker overlap. Markers were overlapping in 20, 2.7, and 0.1% of the projections for prostate, liver, and external, respectively.Conclusions:
We have successfully implemented a new method that, without prior knowledge on marker size, position, orientation, and number, autonomously segments cylindrical gold markers from CBCT projections with a high success rate, despite overlap, motion, and low contrast.
This work was supported by research grants from Varian Medical Systems, Palo Alto, CA, The Danish Cancer Society, and CIRRO—The Lundbeck Foundation Center for Interventional Research in Radiation and the Danish Council for Strategic Research.
II. MATERIALS AND METHODS
II.A. Automatic marker segmentation method
II.A.1. Step 1: Marker candidate identification
II.A.2. Step 2: Trace candidate identification
II.A.3. Step 3: Inconsistent trace candidate rejection
II.A.4. Step 4: Construction of 3D marker constellation model
II.A.5. Step 5: Final template-based segmentation
II.B. Test of segmentation algorithm
Data & Media loading...
Article metrics loading...
Full text loading...