The Impact of Segmentation Accuracy on Treatment Planning for Regional Hyperthermia

Abstract

In a previous publication a statistical shape model for semi-automatic segmentation of the pelvic bone in MR image data was presented. The model was created from a training set of 23 male patients’ CT scans. The completeness of the model was checked in a leave-it-out test, and the segmentation procedure on its own was evaluated in a leave-it-in test. The mean surface distances between model-based and manual segmentation were 0.6 mm for the leave-it-in test and 1.8 mm for the leave-it-out test, respectively; the maximal distances were 4.7 mm and 15.6 mm. It was concluded that the shape model was not yet complete, and the training set should be extended.

For estimating the required training set size, the impact of segmentation accuracy on the outcome of a treatment planning was investigated. For this task, a regional hyperthermia using BSD’s Sigma-Eye applicator was simulated for three patients from the training set. The simulations included an E-field calculation, the solution of the steady-state bio-heat transfer equation, and a phase optimization for the 12 channels of the Sigma-Eye.

Simulations were done for three different geometric models of each patient. Only the shape of the pelvic bone differed between them. It was taken from manual segmentation, which was assumed to be exact, from the leave-it-in and leave-it-out segmentation.

Comparison of the leave-it-in with the exact simulations showed that

• the optimized phases differed by maximally 25 degrees,
• when changing from the exact optimal phases to the leave-it-in phases, the T90 for the target region decreased by less than 0.05 centigrade, and the mean temperature change in the whole patient model was about 0.1 centigrade.

Correspondingly, for the leave-it-out simulations the maximal phase difference was 60 degrees, the T90 decreased by about 0.1 centigrade, and the mean temperature change was about 0.2 centigrade.

Hans Lamecker

Director, Software Development

Advancing 3D analysis, planning, design and manufacturing using innovative computational methods and tools