Author: Kyle Cuneo, Issam M. El Naqa, Dale W. Litzenberg, Yiming Liu, Xueding Wang, Lise Wei, Wei Zhang, Jiaren Zou 👨🔬
Affiliation: University of Michigan, H. Lee Moffitt Cancer Center 🌍
Purpose: To quantitatively map 3D dose deposition during radiotherapy, empowering real-time adaptive radiation treatment.
Methods: The research features reconstructing dose deposition from acoustic measurements from the ionizing radiation acoustic imaging (iRAI) system. IRAI measurements were performed during Stereotactic Body Radiation Therapy (SBRT) of liver cancer patients and processed with principal component analysis (PCA) to extract signals. To support deep learning model training with paired measurements and initial pressure, we utilized the k-wave toolbox to simulate acoustic propagation in heterogeneous space and collect the corresponding acoustic measurements. We then implemented the traditional delay-and-sum algorithm (DS) to retrieve a preliminary reconstruction of the initial pressure and further improved it with a deep learning (DL) model. The model features a 3-D U-net architecture and was implemented in PyTorch. We conducted experiments on simulation and clinical data and plotted direct visualizations and isodose lines for test data points. The performance was also quantitatively evaluated by mean-squared error and signal-to-noise ratios, and dose-volume histograms were compared between planned and reconstructed doses.
Results: Our experiment results on simulation data proved that the DL model significantly improved the mean-squared error and the signal-to-noise ratio of the reconstruction, outperforming the traditional delay-and-sum method. Experiments on patient data indicated that PCA effectively extracted signal components from noisy measurements, and the DL model enhanced the imaging quality, especially in high-dose regions. With a pre-trained model, the initial reconstruction can be post-processed in around 0.02s, satisfying online imaging system requirements.
Conclusion: The research reveals a promising pathway to more quantitative radiotherapy dose imaging with deep-learning-enhanced iRAI technology. PCA pre-processing enhances the signal quality of iRAI measurements, and the 3D-UNet improves imaging quality for both simulation and patient data. Importantly, the successful reconstruction using patient data demonstrates the approach’s potential application in actual clinical scenarios.