Computer-assisted decision making in portal verification—optimization of the neural network approach ☆

Abstract Purpose: Conventional portal verification requires that a qualified radiation oncologist make decisions as to the set-up acceptability. This scheme is no longer sustainable with the large numbers of images available on-line and stringent time constraints. Therefore the objective of this study was to develop, optimize, and evaluate on clinical data an artificial intelligence decision-making tool for portal verification. The tool, based on the artificial neural network (ANN) approach, should approximate, as closely as possible, portal verification assessments made by a radiation oncologist expert. Methods and Materials: A total of 328 electronic portal images of tangential breast irradiations were included in the study. A radiation oncologist expert evaluated these images and rated the treatment set-up acceptability on a scale from 0 to 10. Translational and rotational errors in the placement of the radiation field boundaries formed seven-dimensional feature vectors that represented each of the 328 portal images/treatments. The feature vectors were used as inputs to a three-layer, feedforward ANN. The neural network was trained on the oncologist’s ratings. Results: The rms discrepancy between the ANN and the expert’s ratings was 1.05 rating points. Using the decision threshold equal to 5 for both sets of ratings, the ANN classifier was capable of detecting 100% of the portals classified as “unacceptable” by the expert. Only 6.5% of portals acceptable to the oncologist were misclassified as “unacceptable” by the ANN. Conclusion: The results of this study indicate the feasibility of using the ANN portal image classifier as an automated assistant to the radiation oncologist. Its role would be to recommend an appropriate decision as to the acceptability or otherwise of a given treatment set-up depicted in a portal image.