AN ALGORITHM FOR AUTOMATIC-ANALYSIS OF PORTAL IMAGES - CLINICAL-EVALUATION FOR PROSTATE TREATMENTS

The aim of this study is to assess the clinical value of an algorithm for automatic analysis of portal images by measuring the method's perf ormance in a clinical study of treatment of prostate cancer. The algor ithm is based on chamfer matching and measures displacements of patien ts relative to prescribed radiation beam positions. In this paper we p ropose a method to quantify the mean standard deviation (MSD) of the p erformance of automatic analysis relative to the MSD of the performanc e of trained radiographers using the clinical data set only, i.e. with out using additional phantoms or simulations. The clinical data set in this study consists of 99 regional AP prostate images of 15 different patients. To assess the performance the automatic analysis in relatio n to that of the human observers, we studied the results of the unsupe rvised automatic analysis, as well as the results of a less-trained hu man observed and a well-trained human observer assisted by the automat ic analysis (in this combination, automatic analysis is done first and the result is modified by the well-trained observer if the observer d oes not agree). First, the intra-observer variations of the well-train ed observer are measured by repetitive analysis of a small subset of t he clinical data. The distribution of differences in analysis between two arbitrary observers is described by the chi(2) distribution and is tabulated in literature. We define the agreement histogram of an obse rver O as an estimator for the chi(2) distribution between O and the w ell-trained human observer, parametrized by the ratio of the intra-obs erver variations of O and the well-trained observer. The ratio paramet er indicates the MSD of O relative to the MSD of the well-trained obse rver and is quantified by fitting the agreement histogram with the chi (2) distribution for three degrees of freedom (two translations and on e rotation). The mean relative standard deviation of the well-trained human observer is 1.0 (about 0.6 mm translation and 0.4 degrees rotati on). The mean relative standard deviation of the less-trained human ob server was found to be 2.1, 1.4 for the automatic analysis (discarding 4% failures), and 1.5 for the well-trained human observer assisted by the automatic analysis. The average computation time of the automatic analysis is around 2 s on a 66-MHz 80486 PC, compared with 30 s to 1 min for the human observers. From these numbers it is concluded that t he well-trained human observer mainly corrects the failures in the aut omatic analysis. The small decrease in precision of the well-trained h uman observer when assisted by the automatic analysis is paid for by a large reduction in workload. Thirdly, the performance of the less-tra ined human observer can be improved when this observer is assisted by the automatic analysis in such way that the observer only corrects the failures of the automatic procedure.