Encrypted login | home

Program Information

Automatic Detection of Patient Identification and Patient Positioning Errors Using 3D Setup Images

no image available
S Jani

S Jani*, D O'Connell , P Chow , N Agazaryan , D Low , J Lamb , University of California, Los Angeles, Los Angeles, CA


SU-C-BRD-4 Sunday 1:00PM - 1:55PM Room: Ballroom D

Purpose: To develop an automated system to detect patient identification and patient positioning errors using algorithmic comparison between megavoltage CT (MVCT) and kilovoltage CT (kVCT) planning images.

Methods: MVCT images from 35 head and neck (H&N) patients and 19 pelvis patients were collected from a Tomotherapy machine, along with the corresponding planning kVCTs. MVCTs and kVCTs were manually aligned according to clinical protocols at our institution. Patient identification errors were simulated by aligning MVCTs and kVCTs from different patients. Positioning errors were simulated by misaligning MVCTs and kVCTs by 1cm to 5cm in the each of the six anatomical directions. For each image pair, a pixel-by-pixel cross-correlation metric was computed within the MVCT image body contour. To eliminate the effect of daily variations in bowel gas, the metric was limited to voxels with HU>-700. The kVCT voxel intensities were remapped to the MVCT scale using a third-order polynomial from a publicly available software package.

Results: A threshold pixel-by-pixel cross-correlation value was found that distinguished between correct and incorrect patient setup with a high degree of accuracy. A stratified 10-fold cross-validation analysis yielded average misclassification probabilities of 0.0030 for H&N and 0.00 for pelvis. For misaligned image pairs, cross-validation analysis yielded average misclassification probabilities of 0.00 and 0.0013 for H&N shifts ≥20mm and ≥10mm across all six anatomical directions, respectively. Misclassification probabilities were 0.00, 0.011, and 0.10 for pelvic shifts ≥30mm, ≥20mm, and ≥10mm, respectively. Receiver operator characteristic analysis for misaligned patients yielded areas under the curve ranging from 0.99 to 1.0 for H&N and 0.86 to 1.0 for pelvis.

Conclusion: This proof-of-concept study shows that pixel-by-pixel cross-correlation of MVCT setup images with their corresponding planning CT images can be used to detect wrong-patient errors as well as incorrect patient shifts in the pelvis and H&N regions.

Contact Email: