[en] Purpose: The growing interest in conformal therapy, together with the advent of on-line electronic portal imaging, bring the promise and challenge of high-precision radiation therapy. The accuracy of treatment delivery during fractionated radiotherapy is determined by the distributions of the systematic setup displacement Δ and the random displacement δ_i which are present during fractioni. Inter -treatment verification can be applied either with conventional portal films or, more conveniently, with an electronic portal imaging device (EPID). It can, in principle, reduce or eliminate the systematic component Δ by correcting the patient setup at fraction i + 1 according to a specific correction rule. A simple example of a correction rule is to correct the patient setup whenever a field displacement μ_i > β is detected, the applied correction having a magnitude m_i k μ_i where k ≤ 1. Intra-treatment verification, in which each fractionated treatment is stopped after a few monitor units to verify the patient setup, can be used to reduce both systematic and random setup errors. The decision level β, the correction factor k, and the correction rule must all be carefully selected to ensure that the accuracy of treatment delivery is improved, and not compromised. Materials and Methods: We have developed a simulation model by assuming that both systematic and random setup errors have normal distribution functions with standard deviations σ_3D and σ_δ respectively. Numerous studies of localization displacements support this assumption. Ideally, these data would be determined at each center during commissioning of the EPID prior to routine clinical application of interventional verification. A new correction rule has been developed, known as the Adaptive Maximum Likelihood rule (AML), which makes use of the accumulated experience of measurements during previous fractions to determine the required correction from an estimate of the systematic displacement /-tilde = /-bar_i iσ_Δ²/(iσ_Δ² + σ_δ²) where /-bar_i x_jμ_j/i. If a subsequent correction is required, σ_new σ_δ i σ_Δ²/[√ (iσ_Δ² + σ_δ²)] will be used for the evaluation of /-tilde_new. Other rules have been developed for situations where either one or both of the distribution functions is unknown or uncertain. These rules were compared with previously published decision rules by simulating 5000 courses of treatment with 25 fractions each. Values of Δ and δ_i were selected randomly from their respective distributions, and corrections were applied to the data according to the decision rules. Results: Inter-treatment correction is most effective in reducing systematic setup errors for small values of the ratio r =σ_δ/σ_Δ. The AML rule resulted in an improvement in the accuracy of 90% and 55% of the treatment courses for r=0.25 and r=1.5 respectively. A previously published rule¹ led to improved accuracy in 72% and 10% for the same values of r, and its application in situations wherever >0.5 could lead to an overall reduction in treatment accuracy. However, the superior performance of the AML rule is obtained at the price of more frequent corrections to the patient setup, averaging 2.5 and 1.7 corrections per treatment course of 25 fractions for r=0.25 and r=1.5 respectively, compared to 1.1 and 0.3 corrections for the rule by Bel et. al. It was found that the utilization of inappropriate correction rules in treatments where the random setup errors are dominant may not only fail to improve treatment, but may lead to significant reductions in the accuracy of treatment delivery. 1. A. Bel et al. Radiotherapy and Oncology 29; 253-260 (1993)