Click here to

Session: Therapy General ePoster Viewing [Return to Session]

A Deep Neural Network Approach For Enhancing Signal-To-Noise Ratio of MV EPID Images Using Optimized KV-To-MV Image-Regression Models

M Ahmed1*, H Nourzadeh2, (1) Vanderbilt University Medical Center, Nashville, TN, (2) Thomas Jefferson University, Philadelphia, PA


PO-GePV-T-219 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To generate an optimized KV energy spectrum and use it in KV-MV image regression models to improve the signal-to-noise ratio (SNR) of MV measured signals.

Methods: Two patient cohorts, i.e., 22 prostate patients and 25 brain patients, were compiled to generate two independent training datasets for UNet-based KV-MV image-regression deep neural networks (dnns). Using an in-house developed EPID image prediction algorithm, thousands of EPID MV-images were automatically generated, by permuting specific beam parameters; such as gantry angles, source-to-imager distances, beams energy, and isocenter positions. Generated EPID images are compiled into a local database with their beam configuration at time of simulation. The MV energy spectrum of the source model, used during simulation, is then replaced with an optimized KV energy spectrum. A genetic algorithm technique was implemented to exhaustively search for an optimal KV-energy spectrum that maximizes signal contrast for each patient cohort. KV-like images are then generated at beam projections, similar to those previously used to predict MV-images, such that for each beam configuration a KV and MV signals exist and will be used during the image-regression training. Using the generated KV-MV image pairs, two dnns were trained and the resultant models were used to enhance the contrast of real EPID measurements.

Results: Optimization of KV energy spectrum was technically doable and found to be feasible in KV-MV image-regression to improve the SNR of MV images. Dual NVIDIA RTX3090 GPU devices require 5-7 minutes to optimize the energy spectrum, while they require <6 hours to train the image-regression models.

Conclusion: In this work, we demonstrated the feasibility of optimizing and utilizing KV-based portal images in improving the SNR of MV-based images using dnn. Future investigation is necessary to understand the limitations in various KV-MV regressions and to evaluate the specificity of the trained regression models for anatomical sites.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email