Author: Weikang Ai, Xiaoyu Hu, Xun Jia, Kai Yang, Yuncheng Zhong 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Massachusetts General Hospital, Department of Biomedical Engineering, Johns Hopkins University, Johns Hopkins University 🌍
Purpose: Real-time tumor tracking is critically important for respiratory motion management for lung cancer radiotherapy. A previously proposed application of a photon counting detector involves measuring Compton-scattered photon signals from a kilovolt (kV) X-ray source to track tumors during treatment. However, the inferior image quality presents significant challenges for accurate tumor tracking. This study focuses on leveraging a denoising autoencoder to enhance image quality for subsequent template-matching-based real-time tumor tracking in kV scattered X-ray images.
Methods: Experiments were conducted on a Varian TrueBeam LINAC using a CIRS lung phantom inserted with a 20 mm diameter spherical tumor. Scattered photon signals from a slice of the phantom were acquired. Tumor motion was simulated by couch translation. A denoising autoencoder (DAE) network was developed and applied to reduce noise in the acquired experimental images. One image was selected as the template image. A circular region of interest around the tumor was contoured as the template tumor. The template matching technique was performed on the experimental images to identify the best-matching region with respect to the template tumor. The errors between the estimated tumor location from template matching and the ground truth were compared.
Results: The image noise level, measured as the standard deviation of the background region in the scattered X-ray image, was improved by 27.6% after denoising. The mean absolute error (MAE) of the estimated tumor center positions was 0.8 ± 1.04 mm. In contrast, the MAE without denoising was 3.76 ± 5.88 mm. The computational time to process a single experimental image was 0.25 ± 0.15 seconds.
Conclusion: The DAE significantly reduced image noise, enabling accurate detection and tracking of the target tumor and its motion. This work demonstrates the feasibility of accurate real-time tumor tracking during lung cancer radiotherapy using kV scattered X-ray imaging.