Author: Louis Archambault, Nicolas Drouin, Alexis Horik, Simon Thibault 👨🔬
Affiliation: Département de Physique, de Génie Physique et D'optique, et Centre de Recherche sur le Cancer, Université Laval, Département de Physique, de Génie Physique et D'optique, et Centre d'optique, photonique et laser, Université Laval 🌍
Purpose: To develop a novel type of real-time 3D dosimeter for the quality assurance of linear accelerators used in external beam radiotherapy.
Methods: An experimental setup was constructed using a 10 cm3 scintillating cube, two plane mirrors and a CCD camera. The two plane mirrors were placed such that 3 orthogonal views of the volume were simultaneously imaged by the camera. Because the scintillation is proportional to the deposited dose, the light emission pattern is proportional to the dose distribution. A deep learning model was built to reconstruct the 3D light emission pattern from a single image, thus giving the relative dose distribution. The model, using advantages from Transformers and Convolutional Neural Networks, was trained to reconstruct the volume with a resolution of 1 mm3 in as little as 34.23±0.11 ms using a NVIDIA GeForce RTX 3090 GPU. Pre-training involved 30 000 synthetic data points generated with Python, and fine-tuning utilized 250 experimental data points. Each data point featured an image containing 3 orthogonal views along with the planned dose distribution calculated by the treatment planning software used as ground truth.
Results: Evaluation metrics were calculated on test data values exceeding 10% of the maximum deposited dose. On a test dataset of 25 samples, the model achieved a mean gamma success rate of 89.2±0.7 with a 3%/3mm criterion. The mean values for structural similarity index measure (SSIM) and mean squared error (MSE) were 0.875±0.006 and 0.0050±0.0004, respectively. Gamma maps indicated good results and only minimal discrepancies at high dose gradients.
Conclusion: The presented work shows great potential for a new type of dosimeter that leverages the advantages of plastic scintillators and deep learning to measure 3D dose maps in real-time. Future investigation will improve the model's architecture, feature an expanded dataset to adress overfitting and add a temporal resolution.