A recurrent neural network for rapid detection of delivery errors during real-time portal dosimetry. (April 2022)

• Record Type: Journal Article
• Authors: Bedford, James L.; Hanson, Ian M.
• Is Part Of: Physics and imaging in radiation oncology. Volume 22(2022)
• Page Start: 36
• Page End: 43
• Publication Date: 2022-04
• Subjects: In vivo dosimetry -- Electronic portal imaging device -- Artificial neural network -- Volumetric modulated arc therapy; Radiotherapy -- Periodicals; Radiation dosimetry -- Periodicals; Cancer -- Imaging -- Periodicals; Oncology -- Periodicals; 615.842
• DOI: 10.1016/j.phro.2022.03.004
• Languages: English
• ISSNs: 2405-6316

Abstract

Highlights: Real-time portal dosimetry can detect errors in volumetric modulated arc therapy. Neural networks avoid false positive errors during intrafraction portal dosimetry. Error detection is 30% earlier with an artificial neural network than with thresholds. Abstract: Background and purpose: Real-time portal dosimetry compares measured images with predicted images to detect delivery errors as the radiotherapy treatment proceeds. This work aimed to investigate the performance of a recurrent neural network for processing image metrics so as to detect delivery errors as early as possible in the treatment. Materials and methods: Volumetric modulated arc therapy (VMAT) plans of six prostate patients were used to generate sequences of predicted portal images. Errors were introduced into the treatment plans and the modified plans were delivered to a water-equivalent phantom. Four different metrics were used to detect errors. These metrics were applied to a threshold-based method to detect the errors as soon as possible during the delivery, and also to a recurrent neural network consisting of four layers. A leave-two-out approach was used to set thresholds and train the neural network then test the resulting systems. Results: When using a combination of metrics in conjunction with optimal thresholds, the median segment index at which the errors were detected was 107 out of 180. When using the neural network, the median segment index for error detection was 66 out of 180, with no false positives. The neural network reduced the rate of false negative results from 0.36 to 0.24. Conclusions: The recurrent neural network allowed the detection of errors around 30% earlier than when using conventional threshold techniques. By appropriate training of the network, false positive alerts could be prevented, thereby avoiding unnecessary disruption to the patient workflow.