MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Reconstruction of Continuous Volume Averaging Effect-Free Beam Profiles From ICProfiler Measurements Using a Machine-Learning Technique

K Mund*, G Yan

Presentations

(Sunday, 7/14/2019) 1:00 PM - 2:00 PM

Room: 221AB

Purpose: Ion chamber (IC) arrays have great potential for linac QA, but cannot replace water tank scan due to poor spatial resolution and volume averaging effect (VAE). We aimed to restore high spatial resolution and VAE-free beam profiles from IC array measurements using artificial neural network (ANN).

Methods: Beam profiles of a 6 MV photon beam (from 2x2 to 10x10 cm² at 1.5 and 5 cm depth) were measured twice with a commercial IC array (ICProfiler), first aligned to beam central axis, then shifted 2 mm with the couch. The exact shift was determined as the distance between the 50% intensity of the two measurements, with which the two measurements were combined. Then, spline fit was used to restore continuous profiles. A three-layer ANN was used to perform deconvolution (the removal of VAE). A sliding window (SW) was used to extract data points from the spline-fit profile as input to the ANN; the ANN output the deconvolved value at the center of the SW. Diode-measured beam profiles were used as desired output to train the ANN. We optimized the SW length and number of hidden neurons by evaluating the NN’s mean squared error and the penumbra width difference (PWD) between the deconvolved and diode-measured profiles.

Results: Optimum SW length and number of hidden neurons for the ANN depended on the measuring depth. Good agreement between the deconvolved and reference profiles was achieved for all tested field sizes. The average PWD decreased from 2.28±0.88 mm to 0.16±0.22 mm and from 2.09±0.59 mm to 0.26±0.56 mm for 1.5 cm and 5 cm depth profiles, respectively.

Conclusion: Using spline fit and ANN-based deconvolution, we can restore high spatial resolution and VAE-free beam profiles from two ICProfiler measurements. Measuring depth-specific ANN models should be trained to achieve optimal results.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email