Deep learning-based fluorescence image correction for high spatial resolution precise dosimetry.

Area of Science:

• Medical Physics
• Radiotherapy
• Image Processing

Background:

• Radiation-excited fluorescence imaging offers high spatial resolution for 2D dose distribution measurement.
• Image quality is often compromised by Cherenkov light, scattered light, and background noise.

Purpose of the Study:

• To develop a novel deep learning model for correcting fluorescence images.
• To enhance the accuracy of dosimetric applications using corrected fluorescence images.

Main Methods:

• Acquired fluorescence images using a complementary metal-oxide semiconductor camera from quinine hemisulfate solution irradiated with photon beams.
• Utilized a convolutional neural network (CNN) trained with projected dose distributions and beam angles.
• Employed an empirical Cherenkov emission calibration method for angular dependency.

Main Results:

• Empirical Cherenkov calibration yielded noise-free images compared to uncalibrated distributions.
• The CNN model accurately predicted projected dose distributions, reducing mean absolute error from 2.02 to 0.766 mm·Gy.
• CNN correction resulted in higher gamma index passing rates than conventional methods.

Conclusions:

• The deep learning-based method significantly improves the accuracy of radiation dose distribution measurements.
• This technique shows potential for optical signal denoising and Cherenkov light discrimination in other imaging modalities.
• The developed method provides a high-resolution, accurate dose verification tool for radiotherapy.