Deep Learning-Based Fluorescence Light Discrimination for High Spatial Resolution Radiotherapy Dose Verification

Abstract

<h3>Purpose/Objective(s)</h3> High spatial resolution quality assurance techniques are crucial to assess complex dose distributions of modern radiotherapy. Radiation-excited fluorescence imaging has great potential for real-time absolute dosimetry with high spatial resolution. However, the imaging data are generally contaminated by noise due to Cherenkov and scattered lights. This project aims to establish a novel deep learning-based model to correct the fluorescence images for accurate dosimetric applications. <h3>Materials/Methods</h3> An acrylic box phantom containing 1 g/L quinine hemisulfate water solution was constructed. 200 MU single-aperture static beams were delivered to the phantom with varying aperture shapes, gantry angle and collimator angle, and emitted fluorescent signals were detected by a digital camera in 2.0 frames per second cine mode. Temporal filtering and down-sampling were applied to obtain fluorescence images with 1 × 1 mm spatial resolution at the isocenter plane. 2D projected dose distributions were obtained by performing forward projection calculation of the 3D dose distributions calculated by a clinical treatment planning system. The acquired noisy fluorescence images and projected dose distributions were set to inputs and labels of the deep learning modeling, respectively. A 22-layer 2D residual convolutional neural network (CNN) was generated, and 200-epoch supervised learning was performed using 73 training data. For evaluation, accuracy of the calculated projected dose distributions was compared with that of a conventional scatter-kernel deconvolution method. <h3>Results</h3> The deep learning model yielded accurate projected 3D dose distributions with high spatial resolution. In contrast to the conventional method based purely on the analysis of florescence imaging signal, our deep learning-based approach was able to mitigate the adverse effects of Cherenkov and other optical noises. Mean squared error of 7 normalized projected dose distributions in test data were improved from 0.053 for uncorrected to 0.029 for CNN-corrected images. Computation time requiring to calculate a projected dose distribution was 0.2 second with a GPU workstation. <h3>Conclusion</h3> The proposed deep learning-based method improves the fluorescence image quality and the accuracy of dose distribution measurements in nearly real time. This technique will also be applied to Cherenkov radiation removal or signal denoising in other fluorescence imaging modalities. This method will provide an accurate absolute dose verification tool with high spatial resolution.