Abstract
The experimental results indicate that the authors' new algorithm outperforms the conventional FBP algorithm in effectively reducing the image artifacts due to undersampling and suppressing the image noise due to the low mAs level.
Introduction
Four-dimensional computed tomography (4DCT) is a vital tool in cancer radiotherapy, particularly for tumors in the thorax and upper abdomen that move as a patient breathes. By capturing a sequence of 3D images over time, clinicians can accurately track tumor motion and target radiation more precisely. However, 4DCT requires a much higher radiation dose—often six times that of a conventional CT—because the scanner must acquire data over multiple respiratory cycles.
While clinicians can reduce this dose by lowering the X-ray tube current (mAs) or taking fewer projections (undersampling), standard reconstruction methods like Filtered Backprojection (FBP) result in images riddled with noise and "streaking" artifacts. This paper introduces a novel reconstruction algorithm designed to achieve high-quality 4DCT images from low-dose data by exploiting temporal redundancies.
Methods
The authors propose a simultaneous reconstruction of all breathing phases using an optimization framework, moving away from the standard practice of independent phase reconstruction. The core innovation is the Temporal Non-local Means (TNLM) regularization term.
Unlike standard NLM, which searches for similar patches within a single image, TNLM searches for similar anatomical structures across different temporal phases. Since anatomy changes only slightly between successive breathing phases, the algorithm can "borrow" information from neighboring time points to fill in the gaps caused by low-dose acquisition. To solve this complex optimization problem efficiently, the team employed Forward-Backward Splitting and GPU acceleration, bringing reconstruction times down to a clinically relevant range of 40–90 seconds per slice.
Results
The study utilized a digital NCAT (New Computerized Anthromorphic) thorax phantom to test scenarios involving high noise (low mAs) and high streaking (undersampling).
As shown in Figure 1, the TNLM algorithm significantly outperforms the industry-standard FBP. In undersampled cases where FBP produced heavy streaks that obscured anatomy, the TNLM images remained clear and usable. Quantitatively, the Contrast-to-Noise Ratio (CNR) improved by a factor of 3.9 to 10.2, making low-contrast structures like the vertebral body much easier to see. The Signal-to-Noise Ratio (SNR) also saw improvements by factors of 2.1 to 5.9.
Conclusion
The TNLM algorithm effectively utilizes inter-phase similarity to suppress noise and artifacts in low-dose 4DCT. While challenges remain—such as blurring in fast-moving regions like the heart and the need for validation on real patient data—this approach demonstrates a significant step toward reducing radiation exposure in radiotherapy planning without compromising diagnostic quality.
Summary updated