A Deep-Learning Error Detection System in Radiation Therapy

Delivering radiation therapy based on erroneous or corrupted treatment plan data has previously and unfortunately resulted in severe, sometimes grave patient harm. Aiming to prevent such harm and improve safety in radiation therapy treatment, this work introduces a novel, yet intuitive algorithm for strategically structuring the complex and unstructured data typical of modern treatment plans so their treatment sites may automatically be verified with deep-learning architectures. The proposed algorithm utilizes geometric and dose plan parameters to represent each plan's data as a heat map to feed a deep-learning classifier that will predict the plan's treatment site. Once it is returned by the classifier, a plan's predicted site can be compared to its documented intended site, and a warning raised should the two differ. Using real head-neck, breast, and prostate treatment plan data retrieved at two hospitals in the United States, the algorithm is evaluated by observing the accuracy of convolutional neural networks (ConvNets) in correctly classifying the structured heat map data. Many well-known ConvNet architectures are tested, and ResNet-18 performs the best with a testing accuracy of 97.8% and 0.979 F-1 score. Clearly, the heat maps generated by the proposed algorithm, despite using only a few of the many available plan parameters, retain enough information for correct treatment site classification. The simple construction and ease of interpretation make the heat maps an attractive choice for classification and error detection.