Automated Verification of IGRT-based Patient Positioning

Abstract

A system for automated quality assurance in radiotherapy of a therapist's registration was designed and tested in clinical practice. The approach compliments the clinical software's automated registration in terms of algorithm configuration and performance, and constitutes a practical approach for ensuring safe patient setups. Per our convergence analysis, evolutionary algorithms perform better in finding the global optima of the cost function with discrepancies from a deterministic optimizer seen sporadically.

Figure 1

Search space of various cost functions investigated. The display shows match quality under different OBI‐DRR displacements for various mathematical formulations of an ideal match. Ideally, these displays should have a single distinctive blue spot (cost function minimum). The normalized cross correlation (a) displays such a characteristic. The gradient difference metric (b) has the distinct global minima (blue, center) but also has too many "spots," making it unsuitable to be used with a gradient descent optimizer. The mean reciprocal metric (c) has a central spot that is too broad. The lower row shows the same search spaces in the absence of the preprocessing step. These search spaces in the absence of histogram equalization are unusable.

Figure 2

Metric values at each iteration for ten registrations under identical conditions started from various initial positions. A heuristic and a deterministic optimization algorithm are compared. For the case shown in the left panel, all registrations with both algorithms arrived at a similar solution. However for a different case, solutions found by the two algorithms (right) had similar metrics but represented different translations/rotations. See Fig. 4 for details.

Figure 3

Differences in translations/rotations obtained when comparing the two optimization algorithms. On 7% of the registrations, translational differences were higher than 3 mm, and on 6%, rotational differences were higher than 1°.

Figure 4

The target DRR image (left), and registered OBI images through the regular step optimizer (middle) and one‐plus‐one optimizer (right) for a case where the two optimizers did not match in final translations and rotations.