Comprehensive Summary
The article presents a new method to reduce the radiation dose during 4D CT imaging, which is crucial for radiotherapy planning of tumors in the chest and abdomen. Current imaging protocols acquire excess projection data, leading to unnecessary radiation exposure. The research introduced a deep learning (DL)-driven approach utilizing a Long Short-Term Memory (LSTM) network. This network was trained on patient breathing signals to predict and trigger only the necessary "beam-on" events for data acquisition in real-time. Testing on independent clinical 4D CT scans demonstrated a median dose reduction of 29%. The reconstructed dose-reduced images showed only minimal differences compared to the full-dose reference images, exhibiting high agreement in segmentations for organs like the lung and liver, negligible impact on image artifacts, and high consistency in tumor segmentation and motion range measurements. Overall, the study concludes that this DL-driven method effectively reduces radiation exposure in 4D CT while preserving diagnostic image quality, thus offering a clinically viable solution consistent with the ALARA (As Low As Reasonably Achievable) principle.
Outcomes and Implications
This research is highly important because it addresses the critical need to minimize radiation exposure during essential diagnostic procedures like 4D CT imaging for cancer treatment planning. By successfully demonstrating that a deep learning model can reduce the radiation dose by a median of 29% while preserving diagnostic quality, the work offers a concrete and clinically viable pathway to better adhere to the ALARA principle. The authors' use of a prototype reconstruction software closely mimicking a potential scanner implementation suggests that this breathing signal-guided, DL-driven data acquisition is highly relevant to clinical practice and could be implemented in CT scanners in the near future, offering significant benefits to patients undergoing radiotherapy.