Program Information

An Automatic 3D Ultrasound Image Registration Algorithm for Daily Prostate Localization in Radiotherapy

A Grimwood

H Zhou¹ , H Rivaz¹ , A Grimwood²*, H McNair³ , A Tree³ , E Harris² , (1) University of Concordia, Montreal, Quebec, (2) Institute of Cancer Research, London, (4) Royal Marsden NHS Foundation Trust, Sutton, Surrey.

Presentations

SU-F-708-3 (Sunday, July 30, 2017) 2:05 PM - 3:00 PM Room: 708

Purpose: To develop and evaluate the performance of an automatic ultrasound registration algorithm for ultrasound-guided radiotherapy of the prostate.

Methods: Three-dimensional transperineal ultrasound images of the prostate were acquired from 13 patients at simulation (US-SIM) and daily treatment (US-Tx) using the Elekta Clarity Autoscan system. US-Tx images were randomly selected for each patient and divided into training (20 US-Tx) and test (21 US-Tx) datasets. Shifts in prostate position between US-SIM and US-Tx were determined by 3 methods: (1) manual identification of 3 or more landmark (LM) positions (3 observers); (2) Clarity Guide Review software (3 observers) and (3) prostate registration framework (PRF) which used 3D normalised cross-correlation of volume-of-interests (VOIs) split into N equally-sized patches and a hierarchical framework with L levels. Training data was used to determine N, VOI and L. Mean LM-shifts of 3 observers were used as ground truth. PRF-errors and Clarity-errors were LM-shifts minus PRF-shifts and Clarity-Shifts, respectively. PRF-errors greater than a visually detectable threshold (10mm) were considered registration failures and replaced by Clarity-errors, giving 3 sets of PRF_C-errors for a clinically-supervised PRF. Pair-wise comparisons of errors were performed using the Wilcoxon signed-rank test.

Results: Median absolute error[interquartile range] (mm) were PRF: 2.1[4.9], 1.5[8.6], 1.7[1.6] and Clarity: 0.8[1.2], 1.7[ 2.1], 1.5[2.0] for the test dataset, in the left-right, superior-inferior, anterior-posterior directions, respectively. PRF failed for 5 US-Tx, giving PRF_C: 1.2[1.6], 1.1[2.0], 1.6[1.5] for the test dataset. Median PRF_C-errors were significantly smaller than Clarity-Error in the anterior-posterior direction and no different from Clarity-errors in other directions.

Conclusion: PRF combined with Clarity, will provide modest improvements in accuracy if supervised by the operator (i.e., PRF_C). Implementation of PRF may, on-average, result in reduced registration times compared to Clarity manual registration.


Contact Email: