Author: Hao Chen, Kai Ding, Xiaoyu Hu, Xun Jia, Heng Li, Devin Miles 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Johns Hopkins University 🌍
Purpose: Accurately delivering radiation dose is critical in intensity-modulated proton therapy (IMPT), where intrafraction motion management plays a pivotal role. Our proton therapy system equipped x-ray system for real-time acquisition of orthogonal x-ray projections, potentially enabling tracking the motion. While fiducial marker-based tracking is feasible, this approach increases patient risks. This study developed a real-time volumetric image reconstruction approach from two orthogonal projections and patient-specific prior information for motion management.
Methods: A two-part deep neural network was developed. The first part, a deep autoencoder (DAE), consisted of an encoder that learned key features of patient-generic prior volumetric CT images and a decoder that generated corresponding CT images. The second part, the real-time imaging (RTI) model, learned to extract 2D information from two orthogonal projections, transformed it into a 3D shape, and integrated it with the patient-specific prior CT image features. Finally, the decoder converted the new feature into the volumetric image. 4DCT images of 32 patients were collected. The DAE was trained with 30 patients where 250 images were used for training and 30 for validation. The remaining 20 images with 20 images from the rest 2 patients were used for testing. 3600 projections were simulated where 2880 projection pairs were employed for training and 720 for validation. The 25%-phase CT was generated via interpolation as prior CT image, and the 65%-phase CT and its 360 projection pairs were utilized for testing.
Results: For the DAE, independent tests achieved RMSE 2.6±0.5 HU, SSIM 0.9998±0.0001 and PSNR 61.3±1.5. For the RTI, the RMSE was13.06±0.15 HU, the SSIM was 0.9960±0.0000 and the PSNR was 49.9±0.1. The average time of reconstructing images from two orthogonal projections was 5 ms excluding the projection loading time.
Conclusion: Numerical results demonstrate that the proposed neural network successfully achieved real-time volumetric x-ray imaging with high accuracy.