Gross patient error detection via cine transmission dosimetry

TL;DR

This study evaluates the effectiveness of EPID-based cine transmission dosimetry in detecting gross patient anatomical errors during radiotherapy, demonstrating reliable error detection methods for patient misalignments and rotations.

Contribution

The paper introduces and validates specific image analysis techniques for detecting gross patient errors using cine transmission dosimetry in radiotherapy.

Findings

Error detection methods reliably distinguish gross patient misalignments and rotations.

Image distance algorithm outperforms other methods based on AUC.

Methods can differentiate between patient variations and different patients in prostate cases.

Abstract

$\textbf{Purpose:}$ To quantify the effectiveness of EPID-based cine transmission dosimetry to detect gross patient anatomic errors. $\textbf{Method and Materials:}$ EPID image frames resulting from fluence transmitted through multiple patients anatomies are simulated for 100 msec delivery intervals for hypothetical 6 MV VMAT deliveries. Frames simulated through 10 head-and-neck CTs and 19 prostate CTs with and without 1-3 mm shift and 1-3 degree rotations were used to quantify expected in-tolerance clinical setup variations. Per-frame analysis methods to determine if simulated gross errors of (a) 10-20 mm patient miss alignment offsets and (b) 15-20 degree patient rotations could be reliably distinguished from the above baseline variations. For the prostate image sets, frames simulated through the reference CT are intercompared with (c) frames through 8-13 different CT's for the same patient to quantify expected inter-treatment frame variation. ROC analysis of per-frame error discrimination based upon (i) frame image differences, (ii) frame histogram comparisons, (iii) image feature matching, and (iv) image distance were used to quantify error detectability. $\textbf{Results:}$ Each error detection method was able to distinguish gross patient miss-alignment and gross rotations from in-tolerance levels for both H&N and prostate datasets. The image distance algorithm is the best method based on AUC. $\textbf{Conclusion:}$ In-field gross error detection was possible for gross patient miss-alignments and incorrect patients. For prostate cases, the methods used were able to distinguish different patients from daily patient variations.